← Blog

2026-07-23 Postmortem Agents Trading Data

We Gave 11 AI Agents Their Own Wallets and Real SOL. Here Are All 90 Trades.

Two weeks ago, an agent on three.ws closed its first autonomous trade. Four days ago we scaled that single agent into a fleet: eleven strategies, each with its own AI agent, its own Solana wallet, and real SOL in it, and we stopped touching the keyboard. Each agent watches the live pump.fun launch firehose, decides on its own which brand-new tokens to buy, sizes the position, and decides when to sell. No human approves anything. Every decision is written to a tamper-evident ledger next to the on-chain transaction that proves it happened.

This post is the complete record of every closed real-money trade the program has ever made, from the first cautious test on July 3rd through the fleet's four days of full autonomy: 90 trades across 8 strategies and 8 wallets, every entry thesis, every exit, every configuration knob, and a grade for every single trade. And because the trades are only half the story, we also audited everything the fleet did not do: the 82,994 launches it saw and declined, the monsters it missed and whether they were catchable, and a coin-by-coin check of whether any of our 90 sells was too early. The dataset is downloadable at the bottom, every transaction links to a Solana explorer, and every agent's reasoning ledger is public. Nothing here is backtested, simulated, or cherry-picked.

The headline, stated honestly: the fleet is down -0.1034 SOL (about $20 at the time of writing) on 1.93 SOL deployed across those 90 trades. That is a small tuition bill. What the tuition bought is the interesting part: the two profitable strategies are both the ones where a language model, not a rulebook, makes the buy decision, and they own six of the seven biggest wins. The rest of this post is the evidence.

90closed real trades

8strategies that traded

25winners (28%)

-0.103 SOLfleet net PnL

+0.015 SOLLLM-judged arms (33 trades)

-0.118 SOLrules arms (57 trades)

+73.9%best trade

4mmedian hold

82,994launches seen, 72 bought

0sells the market later proved early

Abstract. We armed eleven autonomous trading strategies, each with its own AI agent and custodial Solana wallet, against the live pump.fun launch firehose, and audited everything they did and did not do. Over the study window the fleet observed 82,994 token launches, submitted 12,770 of them to LLM buy-side judges, and closed 90 real-money positions across 8 strategies for a net of -0.1034 SOL on 1.93 SOL deployed. The LLM-judged cohort was the only profitable one (+0.0147 SOL over 33 trades) despite a lower win rate than the rules cohort, an expectancy effect driven by selection rather than prediction: judged coins the models bought went on to pump or graduate 53.3% of the time versus 11.7% for coins they skipped. The platform's conviction scorer proved strongly predictive at scale (its 50-plus band hit 77.5% good outcomes across 71 labeled coins, seven times the base rate) yet the fleet traded none of that band because two gate thresholds bracketed it. A candle-verified paperhands audit found zero premature sells in 90 exits. Five stop-losses overshot by 48 to 85 points on liquidity collapse. Two autonomous learning loops ran silently dead for two days behind healthy status codes. All raw data, on-chain receipts, per-trade grades, and an interactive exit-policy replay are included below.

Contents

The theory: why race rules against a model
The machine: how a buy actually happens
Speed: from token creation to our buy landing
The market we traded in
The results, in five charts
The league table, and every wallet in full
How trades ended, and what each exit reason teaches
Win rate is a vanity metric
What the judges said no to: 12,770 verdicts
The Oracle's report card
The 82,994 coins we did not buy
The ones that got away
The paperhands audit: did we sell too early?
The disasters: when a stop-loss is a request, not a guarantee
The cost of doing business
Which entry signals actually separated winners from losers
The two silent failures that faked autonomy for two days
The fleet's own decisions so far
Interactive: replay every exit with your own rules
All 90 trades, graded
Everything we learned
What happens next
Methods, data, and glossary

1. The theory: why race rules against a model

When our first agent closed its first autonomous trade at +42%, we wrote that most of that exit was luck and that the honest way to find out what actually works is a controlled experiment. This fleet is that experiment. It is a multi-armed A/B test where each arm is a full strategy with its own wallet, its own budget, and one deliberately different idea about how to pick entries:

Rules arms encode a human hypothesis as frozen numbers: market-cap bands, socials requirements, creator-history filters, quality-score thresholds. They are fast, cheap, deterministic, and cannot adapt to anything.
Oracle arms gate entries on our conviction scorer, a model trained on labeled outcomes of past launches (did the coin graduate, pump, or die). One arm requires modest conviction, one requires high conviction.
Intel arms wait out the first minutes of trading and only enter once the intel engine confirms a real two-sided market rather than a painted chart.
LLM arms throw the rulebook away. Every new launch is put in front of a language model with the coin's metadata, the creator's history, and a structured read of the live chart. The model returns {buy, confidence, thesis} and the thesis is stored on the ledger next to the trade. One arm is piloted by a Grok-class model, one by a Claude-class model, one lets the router pick.
A graduation arm ignores new launches entirely and buys tokens at the moment they complete the bonding curve, riding the migration flow with tight exits.

Every arm sits on the identical execution stack, the identical safety rails, and the identical exit ladder machinery, so the only variable being tested is the entry judgment. Position sizes are deliberately small (0.002 to 0.15 SOL per trade) because the point of this phase is information per SOL, not profit.

Three armed strategies (oracle-strict, rules-no-socials, rules-wide-band) never closed a real trade in the window: their filters rejected everything the firehose offered. That is not a malfunction. A strategy that mostly refuses to trade is a strategy expressing an opinion, and the skip log records a reason for every rejection.

2. The machine: how a buy actually happens

A long-lived worker holds a websocket open to the pump.fun firehose and sees every token on Solana the second it exists. From there, a launch has to survive a gauntlet before any wallet key is touched:

The strategy's own entry logic. Rules arms check their filters; Oracle arms ask the conviction scorer; LLM arms ask their model. This is the experimental variable.
The fleet safety bands. Platform-wide clamps individual strategies can tighten but never loosen. A coin whose market cap cannot be read fails closed.
The Mayhem gate. Pump.fun's adversarial launch mode is excluded unconditionally, read straight from the on-chain bonding curve.
The trade firewall. Before spending, the executor simulates the full buy and sell round trip on-chain. If the token cannot be sold back, it is a honeypot and the buy aborts. 65 of the 90 trades in this dataset carry an explicit allow verdict with a score.
Budgets and concurrency. Per-trade size, daily budget, max concurrent positions, SOL headroom. Only then does the agent's encrypted key sign.

Exits run on a strict priority ladder. It is thirty lines of code and it governed every one of the 90 closes below, so here it is in full:

// workers/agent-sniper/exit-logic.js: the single source of truth for
// "should this position exit, and why". Priority order is load-bearing.
export function decideExit(pos, value, peak, now = Date.now(), sentiment = null) {
  const entry = BigInt(pos.entry_quote_lamports || '0');
  if (entry <= 0n) return null;
  const ev = Number(entry);
  const sl = pct(pos.stop_loss_pct);
  const ts = pct(pos.trailing_stop_pct);
  const tp = pct(pos.take_profit_pct);

  if (sl != null && value <= ev * (1 - sl / 100)) return 'stop_loss';
  if (isBearishFlip(sentiment) && value < ev) return 'signal_flip';
  if (ts != null && peak > 0 && value <= peak * (1 - ts / 100)) return 'trailing_stop';
  if (tp != null && value >= ev * (1 + tp / 100)) return 'take_profit';

  const heldS = (now - new Date(pos.opened_at).getTime()) / 1000;
  if (pos.max_hold_seconds != null && heldS >= pos.max_hold_seconds) return 'timeout';
  return null;
}

The order is load-bearing. The stop-loss always wins, a trailing stop outranks a take-profit, and the timeout is the backstop that guarantees no position outlives its policy. Wherever you see timeout in the trade list, it means "no other rule fired and the clock ran out".

For the LLM arms, this is the prompt doing the judging. Note what it is being asked to do: not to predict the future, but to read a chart shape and a creator history the way a discretionary trader would:

// workers/agent-sniper/llm-judge.js: what the model is actually told.
const SYSTEM_PROMPT = [
  'You are the buy-side judge for an autonomous pump.fun sniper on Solana.',
  'You will be shown one brand-new token launch. Decide whether a small,',
  'time-boxed momentum position (held minutes, hard stop-loss) is worth taking.',
  // ...
  'CRUCIAL: when market data is shown, judge the CHART like an experienced trader.',
  'A big opening candle that then stairsteps up on a handful of wallets with no',
  'sellers is a PAINTED chart meant to trap momentum bots, not real demand: skip',
  'it. Genuine momentum has a real two-sided market.',
  // ...
  'Respond with ONLY a JSON object, no prose, no code fences:',
  '{"buy": true|false, "confidence": <0..1>, "thesis": "<one short sentence>"}',
].join(' ');

One honest infrastructure note: the fleet's LLM calls route through a failover chain, and during much of this window the named primary models were falling back to cheaper chain models (you will see fallback:ovh and fallback:groq#instant in the trade receipts). So this data does not yet cleanly separate "Grok vs Claude"; what it does separate cleanly is model judgment vs frozen rules, because whichever model answered, it was a model reading each launch fresh.

Finally, every buy writes a decision to the agent's hash-chained reasoning ledger: rationale, confidence, firewall verdict, prediction. When the position closes, a reconciliation job scores the prediction against realized PnL. That loop is what makes the rest of this article possible, and it is why none of it required reconstruction from memory.

3. Speed: from token creation to our buy landing

Before any question about judgment, a sniper has to answer a question about physics: how long after a token exists can the machine own it? We measured it for every trade with a recoverable creation timestamp, as the gap between the token's on-chain creation and our buy transaction landing:

The raw-feed paths are effectively instant: the median new_mint entry landed 3 seconds after the token was created, and graduation rides caught the migration window in 4 seconds. That covers the full pipeline: firehose event, strategy evaluation, safety bands, the on-chain firewall's simulated buy-and-sell round trip, and transaction confirmation. The intel-mediated paths (intel_confirmed, llm_intel) land around 1.6 minutes by design: they deliberately spend the coin's first ninety seconds watching the market before deciding anything.

Two implications. First, execution speed is not this system's bottleneck and never was; nothing in the loss column is explained by being late. Second, the speed made the LLM arms possible at all: a judge that answers in 2.1s median (section 9) still gets its winning entries filled seconds to minutes into a coin's life, comfortably inside the window where the crossing analysis (section 12) says the upside still exists.

4. The market we traded in

Numbers without their environment mislead, so here is the environment. During the 76 hours the fleet ran, pump.fun produced 82,994 new tokens, roughly 1,092 per hour around the clock:

Of the 42,995 launches our intel engine watched to a labeled conclusion, 84.6% rugged, 3.5% went flat, 9.5% pumped meaningfully, and 2.4% graduated to a full DEX listing. That is the deck every strategy in this experiment was drawing from: a market where roughly seven of eight coins are designed to take your money, where the median coin is dead within the hour, and where the winners, when they happen, can run four or five orders of magnitude. Every result below should be read against that base rate, and it is why this article spends as much time on the coins we skipped as the coins we bought.

One regime note for anyone reproducing this: pump.fun's BOOST mechanic (a five-minute post-migration buyback window) went live mid-July, shortly before our window. It is the reason the boost-ride arm exists, and it means graduation-adjacent outcome statistics from before that change are not directly comparable to ours.

5. The results, in five charts

The white line is the fleet's cumulative realized PnL across all 90 closes (the horizontal axis is trade sequence, labeled by close date); purple is the LLM arms alone; orange is everything rules-driven. The story is visible at a glance: the rules cohort bleeds slowly and steadily, while the LLM cohort chops sideways and then steps up in discrete jumps. Those jumps are take-profit exits in the +45% to +74% range. The fleet line is the sum of an orange slope and purple stairs.

Per arm, the split is stark. The two profitable arms are both LLM-judged (purple labels). Every rules arm is flat to negative, and the two worst performers lost money in completely different ways: intel-quality took a small number of large losses on coins whose charts looked confirmed, while boost-ride took 31 small trades and got wrecked by four catastrophic liquidity collapses (section 14).

This chart is the argument against judging a strategy by its win rate. boost-ride wins 48% of the time and is still net negative; llm-auto wins only 20% of the time and is net positive, because its average winner is +64.5%. Win rate tells you how often you were right. It says nothing about what being right was worth.

47 of 90 trades (52%) ended on timeout: the coin did nothing definitive and the clock expired. Take-profit exits are the only category that is meaningfully net positive (+0.056 SOL across 13 trades), and stop-losses are the biggest single drain (-0.086 SOL across 18).

Each dot is one trade, horizontal axis is hold time on a log scale, vertical is realized PnL. The structure worth noticing: green take-profit dots cluster between ten minutes and half an hour (winners need time to run), red stop-loss dots stack in the first minutes (losers announce themselves fast), and the long right tail is graduation_ride positions that survived hours or days. The five red dots at the very bottom are the liquidity-collapse disasters.

6. The league table

arm	mode	trades	wins	net SOL	avg win	avg loss	best	worst	med hold
llm-grok	llm	9	4 (44%)	+0.0190 SOL	+53.6%	-4.9%	+71.3%	-9.2%	30m
llm-auto	llm	15	3 (20%)	+0.0064 SOL	+64.5%	-10.8%	+73.9%	-37.4%	30m
rules-classic	rules	1	0 (0%)	-0.0006 SOL	+0.0%	-28.1%	-28.1%	-28.1%	4.4d
llm-claude	llm	9	1 (11%)	-0.0107 SOL	+10.6%	-14.7%	+10.6%	-78.3%	30m
oracle-open	rules	14	1 (7%)	-0.0167 SOL	+10.4%	-13.6%	+10.4%	-57.4%	30m
rules-proven	rules	5	1 (20%)	-0.0183 SOL	+42.4%	-19.8%	+42.4%	-26.1%	2m
boost-ride	rules	31	15 (48%)	-0.0413 SOL	+14.9%	-39.8%	+42.9%	-99.9%	3m
intel-quality	rules	6	0 (0%)	-0.0413 SOL	+0.0%	-4.6%	-2.5%	-8.0%	30m

Every wallet, in full

Nothing in this experiment is anonymous. Each arm trades from its own custodial wallet; here is each one's full address, what it holds right now, and its complete lifetime record, with direct links to the on-chain history and the agent's public decision ledger. Balances are live at time of writing and include SOL the auto-funder topped up beyond trading PnL, which is why balance and net PnL do not reconcile one-to-one.

arm	agent	wallet address	balance now	deployed	realized	ROI	links
llm-grok	Swarm 2	`HBgNvnffvbKRzE1vgmvTfKLNq2taETKZJDpZRHcAm8Ca`	0.0369 SOL	0.130 SOL	+0.0117	+9.0%	chain · ledger
llm-auto	Swarm 5	`HXDzfb4creqEuvpkBRoacFJFQKXQDz7YDB7CiFNg3F3s`	0.0592 SOL	0.160 SOL	+0.0056	+3.5%	chain · ledger
rules-wide-band	Swarm 1	`98HyVNAbitNjwGafNZyrd7zuPsCT6cse1ZYGSwMy2cHB`	0.0500 SOL	0.000 SOL	+0.0000		chain · ledger
oracle-strict	My First Agent		? SOL	0.000 SOL	+0.0000		chain · ledger
rules-no-socials	Swarm 3	`8EbtMn6cSZd7tCsFb7y3DBYsdV4jDzqYzqNw7QBRBpww`	0.0500 SOL	0.000 SOL	+0.0000		chain · ledger
rules-classic	Swarm 7	`7KdHYVovYzmrJCE3w6g1iGxeUWJmXoTZLKbBqD7QdWfD`	0.0500 SOL	0.002 SOL	-0.0006	-28.1%	chain · ledger
llm-claude	Swarm 9	`65KauYM8hnLAZVt1KWj3mz6fyhKMyzHXDjz1r5cLw3h8`	0.0230 SOL	0.090 SOL	-0.0107	-11.8%	chain · ledger
oracle-open	Moe Money AI	`54tvt4wmrb9KbxSczoDvbo5mjLJupdUA7ZuePa8hv23M`	0.0371 SOL	0.140 SOL	-0.0167	-11.9%	chain · ledger
rules-proven	three	`3334ZPryymmAuoRx13pBfB2kepL2ULexoirjwqhFKrsc`	0.3515 SOL	0.250 SOL	-0.0183	-7.3%	chain · ledger
boost-ride	Sniper Arm: Boost Ride	`3yjxcqWqjGpESMJ4EuTpDEAyTNwHvNVA38L4ZGyyaQwt`	0.0587 SOL	0.310 SOL	-0.0413	-13.3%	chain · ledger
intel-quality	Nadirah	`D5BGRDsXVNBt6iZZ83P9hFib1w4SR6AzL55HrNnrXypu`	0.2930 SOL	0.900 SOL	-0.0413	-4.6%	chain · ledger
fleet master (auto-funder)		`888BXStWfjmb5CLUL99GGPxqrHQXDhqaj8K4sXChfpvp`	0.9015 SOL				chain

Anyone can audit any of these wallets on any Solana explorer, at any time, without asking us. That is not bravado; it is the constraint that keeps every claim in this article checkable, and it is also live risk transparency: these are hot wallets executing autonomous code, sized so that a total loss of any one of them is an acceptable research cost.

And the seven biggest winners fleet-wide, six of which belong to the LLM arms:

token	arm	realized	exit	hold	receipts
House	llm-auto	+73.9%	take_profit	68s	buy · sell
DEPE	llm-grok	+71.3%	take_profit	2m	buy · sell
DUCK	llm-grok	+65.6%	take_profit	72s	buy · sell
BLACKHOLE	llm-grok	+63.1%	take_profit	2m	buy · sell
WhiteBull	llm-auto	+60.8%	take_profit	3m	buy · sell
掌握	llm-auto	+58.8%	take_profit	84s	buy · sell
HANU.p	boost-ride	+42.9%	take_profit	5m	buy · sell

7. How trades ended, and what each exit reason teaches

take_profit (13 trades, +0.056 SOL). The only systematically profitable exit. Every arm that lacked a configured take-profit either got lucky on a timeout or gave its gains back. After the first trade taught us this, every fleet arm shipped with a TP; the data now confirms it with 13 clean examples.
timeout (47 trades, -0.036 SOL). The modal outcome. Most pump.fun launches simply flatline, and a timeout on a flat coin costs a little slippage each way. 11 timeouts closed green, which is luck, not policy: nothing locked those gains in. The real cost of timeouts is invisible: dead coins held for their full configured hold (4 to 30 minutes depending on the arm) tie up a concurrency slot that a live coin could have used.
stop_loss (18 trades, -0.086 SOL). Working as designed in most cases: the median stop landed within a few points of policy. The tail is the problem; see section 14.
trailing_stop (12 trades, -0.037 SOL). The most misunderstood exit in the ladder. A trailing stop armed while the position is still underwater just converts a small loss into a realized one. Trailing stops earned their keep only on positions that were already green; the fix (arm the trail only above breakeven) shipped from this dataset and is now how the exit ladder behaves (section 22).

8. Win rate is a vanity metric

Put the two cohorts side by side and the asymmetry is the entire story:

cohort	trades	win rate	avg winner	avg loser	net
LLM-judged	33	24%	+52.3%	-10.8%	+0.0147 SOL
rules-driven	57	30%	+16.2%	-23.7%	-0.1182 SOL

The rules cohort wins more often than the LLM cohort (30% vs 24%) and still loses three times as much money as the LLM cohort makes. The LLM arms' average winner (+52.3%) is more than three times the rules arms' average winner (+16.2%), and their average loser is less than half as deep. Same market, same window, same execution stack, same exits available.

Why? Read the 29 stored theses below and a pattern emerges: the model is doing selection, not filtering. A rules arm buys everything that passes its thresholds, which in practice means everything sufficiently mediocre to not trip a filter. The model was asked to behave like a selective trader with no other filters behind it, and it skips launches a rulebook has no vocabulary to reject: tickers that are trying too hard, narratives that are a week stale, charts that stairstep on four wallets. The result is fewer entries with fatter right tails. The obvious caveat is equally important: 33 trades is a directional signal, not a proof, and we treat it accordingly (section 21).

9. What the judges said no to: 12,770 verdicts

The trades only show you the buys. The judgment ledger shows you everything else: every verdict any LLM judge issued, buys and skips, is persisted with its confidence, thesis, latency, and which model actually answered. Over the study window the judges issued 12,770 verdicts and said buy 485 times: a 3.8% buy rate. Asked to behave like a selective trader with no other filters behind them, the models were roughly twenty-five times choosier than a coin flip:

requested model	judged	buy verdicts	buy rate	median latency	answered by fallback
`anthropic/claude-haiku-4.5`	5,101	171	3.4%	1.8s	77%
`openrouter/auto`	4,950	189	3.8%	2.2s	67%
`x-ai/grok-4.3`	2,719	125	4.6%	2.1s	82%

The last column quantifies the routing problem we flagged in section 2: 74% of all verdicts were answered by a fallback model, not the model the arm requested. Until that is fixed, per-model comparisons here measure the failover chain more than the named pilots, which is exactly why the fixed-routing rerun leads the roadmap.

Selectivity is only interesting if the selections were good, and the ledger lets us check that independent of trade execution, because the intel engine later labeled 3,850 of the judged coins with a ground-truth outcome:

Coins the judges wanted went on to pump or graduate 53.3% of the time; coins they skipped, 11.7%. The skip rate matches the market's base rate almost exactly, which is what you want to see: the judges are not skipping winners at an elevated rate, they are pulling winners out of the stream at four and a half times the base rate. This is the cleanest evidence in the whole study that the LLM cohort's trading edge is real selection and not execution luck, because it holds on coins the fleet never traded at all.

Confidence calibration adds a warning label. Verdicts issued at 0.7 confidence went good 73% of the time, the strongest band by far. But the most confident verdicts underperformed: the 0.8 band went good 47% and the 0.9 band went 0 for 4. The samples are small, but the shape is the classic overconfidence signature, and it is actionable: the arms' confidence floor already exists (llm_min_confidence), and this data argues for a confidence ceiling too, treating a 0.9-plus verdict on a seconds-old memecoin as a red flag rather than a green light.

10. The Oracle's report card

The Oracle is our conviction scorer: a trained model that rates each launch 0 to 100 on evidence like buyer diversity, creator track record, and early-chart shape. 47 of the 90 trades had a conviction score at entry (the unscored balance is mostly graduation_ride entries, which the Oracle does not cover).

The gradient goes the right way: sub-30-conviction entries won 19% of the time and lost -0.054 SOL in aggregate, while the 30-to-50 band did meaningfully better on both counts. The damning detail is the empty top band: not one trade in this window entered with conviction above 50. The fleet spent two days paying for exactly the coins its own scorer was lukewarm about, because most arms were not required to listen to it. The conviction gate exists precisely to prevent this; the data says the arms that ignored it funded the lesson. And the top band is not empty because high-conviction coins do not exist: section 11 finds 110 of them in the same window, none bought, for a reason that is almost comically mechanical.

The ledger's own confidence numbers tell the same story from the other side. Snipe decisions recorded with confidence around 0.6 went on to be right 32% of the time (23 of 73); decisions recorded around 0.4 were right 9% of the time (3 of 34). The absolute numbers are humbling, but the ordering is real: the system's self-rated confidence carries signal, which means it can be calibrated rather than discarded.

11. The 82,994 coins we did not buy

Everything above analyzes the trades the fleet took. That is half the record. A sniper's real output is its selectivity, and while the fleet was live the intel engine observed 82,994 launches in three days, scored 45,784 of them with the Oracle, watched 42,995 long enough to label their final outcome, and bought 72. That is a buy rate of less than one in a thousand.

Because those 42,995 labeled coins each carry the conviction score the Oracle gave them, we can finally answer the question the traded sample was too small for: is the Oracle's score actually predictive, at scale?

Yes, and dramatically so. Coins the Oracle scored below 30 pumped or graduated 11.8% of the time, roughly the same as unscored coins (10.5%). The 30-to-50 band improves to 17.4%. And the coins scored 50 or above went on to pump or graduate 77.5% of the time (9 graduations and 46 pumps out of 71 labeled). A seven-times-the-base-rate signal, sitting in production, feeding a live feed.

Now the painful part: the fleet bought 0 of those 110 high-conviction coins. Zero. And the reason is embarrassingly mechanical. Two arms gate on conviction: oracle-open requires 35 or more, and oracle-strict requires 65 or more. The highest score any coin reached in the window was 61. So the strict arm's bar was mathematically unreachable and it sat idle all week, while the open arm's bar of 35 admitted the 30-to-50 band (17.4% good) and it duly went 7%-win negative. The two gates bracketed the profitable band (50 to 61) and covered none of it. Nobody chose this; it fell out of two thresholds set on different days by different reasoning, and no analysis existed to catch it until now.

Honesty requires checking one alternative explanation: conviction scores are recomputed as a coin trades, so maybe the 50+ scores only appeared after the pumps they seem to predict. The score history table refutes it: of the 110 high-conviction coins, 104 have an archived first score, and all 104 were already at 50 or above the first time the Oracle ever scored them, none were rescored up after the fact. The signal was on the board before the outcome. (What history cannot fully rule out is entry slippage: by the time a coin proves enough demand to score 50, its price has moved. A crossing-triggered arm, described in section 22, is how we find out what is actually capturable.)

12. The ones that got away

The counterfactual everyone actually wants: which monsters did we skip? Here are the biggest winners the intel engine observed and passed on during the window, deduplicated by ticker, with the scores the system had at the time:

token	outcome	ATH cap	from-launch multiple	quality	conviction	buyers seen
USOH	graduated	$161M	73,975x	43	unscored	3
DOW	graduated	$134M	61,692x	36	20	3
NTFS	graduated	$76M	35,170x	43	21	3
USWR	graduated	$67M	31,158x	43	21	3
VORF	graduated	$25M	11,674x	43	21	3

Look at the last three columns before feeling bad. Every one of these coins, at the moment our systems watched it, had a handful of visible buyers, mediocre quality, and low or no conviction. The $161M ATH winner showed 3 buyers and quality 43 during its observation window. These were not signals anyone ignored; at minute zero, these coins were indistinguishable from the 36,366 rugs in the same dataset. Their runs started minutes to hours after our observation window closed.

That is the actual lesson, and it is structural: a minute-zero sniper is blind to every winner whose story starts at minute thirty. No threshold tuning fixes it, because the information does not exist yet at decision time. What fixes it is a second look: re-scoring coins at 30 and 60 minutes and entering on demonstrated, sustained demand rather than launch-second vibes. The Oracle's 50+ band (section 11) is exactly that signal, which is why wiring an arm to it is the top item on the roadmap.

Where the fleet did have the information and skipped anyway, the record is defensible: of the 110 coins that crossed conviction 50, the 16 that later rugged would have cost us, but the 55 that ran would have paid for them roughly 3-to-1. We are careful not to claim a specific PnL for a strategy nobody ran (entry slippage, exit behavior, and fill quality are all unknowable in hindsight), but the direction is not subtle.

So we measured how catchable the band actually is, coin by coin, against full candle histories. For each of the 102 measurable crossings, take the close of the first candle at or after the coin's first 50-plus score as a hypothetical entry, and ask what the price did afterward. Three answers matter. First, the crossing lands at a median of 2 minutes after launch: this is a sniper-speed signal, not a lagging one. Second, the upside is real: the median crossing offered 1.23x from the entry candle, 35% of crossings reached at least 1.5x and 19% reached 2x, with the best run at 9.6x, and not a single coin had already topped before crossing. Third, the warning label: the median crossing coin eventually decays to 0.32x its entry, so the profit exists only for a strategy that takes it; buy-and-hope converts this band back into losses. (Method honesty: candle-close entries, no slippage or fill modeling; the live arm's fills will be worse than the candles by some spread we can only learn by trading it.)

13. The paperhands audit: did we sell too early?

Every trader's private fear: the coin you sold kept running. We checked it for all 90 exits, two independent ways.

Way one: where does every coin we sold trade right now? We recomputed each coin's implied market cap at our exact exit fill (our sell price per token times pump.fun's fixed 1B supply, SOL at ~$190) and compared it to its live market cap today:

The answer is a flat no. Not one of the 90 coins we sold trades above our exit today. The best sits at 0.90x our exit price, the median at 0.36x, and 37 of 90 have lost more than 90% of their value since we left. The 0.25-to-0.5x cluster (44 coins) is the pump.fun floor state: dead curves settling at the couple-thousand-dollar residual cap.

Way two: every token's full price history. Current price could hide a coin that ran 10x after our exit and then died. So we pulled the complete candle history for all 85 traded tokens from pump.fun's swap API (price per token; times the fixed 1B supply gives market cap) and, for each of the 90 trades, measured the highest price the token ever printed after our exit candle, at one-minute granularity where the coin's full life fits in minute candles and hourly otherwise. This is the strict test: it catches the spike-after-we-sold case that both the current-price check and outcome labels can miss.

The verdict, coin by coin: 37 of the 90 tokens never printed a single trade after our exit. We were, literally, the last one out the door before the lights went off. Of the 53 that did keep trading, the median post-exit high is 0.43x our exit price, zero reached 1.5x, and only three edged above our exit at all. The closest call in the whole dataset is trade #79: a trailing stop shook llm-claude out of a position at +10.6% and the token printed 1.38x our exit price afterward, worth about 0.0042 SOL of missed upside. The other two are a 1.24x dead-cat bounce off a -98.7% rug floor and a 1.07x wiggle on an hourly candle, both noise. Total SOL left on the table by selling too early, across the entire experiment, at any threshold that clears measurement noise: zero.

Which produces this experiment's strangest compliment: the fleet has no paperhands problem at all. Every one of the 90 sells, including every take-profit, was vindicated by the token's own price history. The exits are not what needs work. The money was lost at the entry, one mediocre coin at a time, and the "money left on the table" cases below are about capturing spikes during the hold, not about holding longer.

The within-trade version is where real improvement lives: 8 positions peaked more than 15% up and still captured less than three-quarters of that peak. The worst: RACOIN peaked +55.0% and realized -78.3%; Poze peaked +19.3% and realized -38.9%; AGIGUY peaked +15.9% and realized -34.4%. A laddered exit (sell a tranche at the target, trail the remainder) is the standard fix and is now on the roadmap; holding longer is not, because the data just told us longer holds end at 0.36x.

14. The disasters: when a stop-loss is a request, not a guarantee

Five trades lost more than 75% of their stake. All five had stop-losses configured at -15% or -30%. This is the table nobody publishes:

token	arm	configured stop	realized	overshoot	hold
`yMDueA…`	boost-ride	-15%	-99.9%	85 pts	3m
`5V3eh6…`	boost-ride	-15%	-98.7%	84 pts	8s
`ChRFYk…`	boost-ride	-15%	-93.2%	78 pts	9s
`2UDry5…`	boost-ride	-15%	-84.3%	69 pts	6s
RACOIN	llm-claude	-30%	-78.3%	48 pts	2m

What actually happens in these rows: the position poller re-quotes the coin every few seconds, sees the value crash through the stop threshold, and fires a market sell into a bonding curve whose liquidity has already evaporated because everyone else is selling into the same hole. The stop triggered on time. The execution of the stop got the price that was left. On a thin curve, that can be pennies.

Every one of these five was a graduation_ride or low-conviction entry on a coin whose holder base was concentrated in a handful of wallets. That is measurable before entry, and it is exactly what the Oracle's whale-concentration and crowd-size features flag. The defenses against this failure mode, in order of effectiveness: do not enter concentrated coins (conviction gate), size down where liquidity is thin (the optimizer now throttles sizes on underperforming arms), and quote faster so the stop fires earlier in the collapse. "Set a stop-loss" is not on the list, because a stop-loss was set every single time.

15. The cost of doing business

Where does money leak in a system like this, besides bad picks? We can account for the frictions precisely because every fill is recorded.

Entry price impact: negligible. Median impact across all 90 entries was 0.036%, the worst single entry was 1.25%, and zero entries exceeded 2%. At 0.002 to 0.15 SOL position sizes, pump.fun bonding curves barely notice us. This matters for interpreting everything else: the fleet's losses are a selection cost, not an execution cost, and the counterfactual analyses (sections 11 and 12) are not being flattered by ignoring slippage that would have eaten the edge at our sizes. It also flags the real scaling question: these results say nothing yet about what impact looks like at 10x the size.
Exit slippage: the tail risk. Normal exits paid a fraction of a percent; the five liquidity-collapse exits (section 14) paid 48 to 85 points past their stops. Exit friction is not a distribution with a mean; it is fine until it is catastrophic.
Network fees: rounding error. Solana transaction fees across all 180 legs summed to well under 0.001 SOL. On this chain, at this size, fees are not a line item worth optimizing.
Inference: the real operating cost. The judges answered 12,770 verdict calls at 2.1s median latency, deduplicated per (coin, model) so same-model arms share one call. Most calls were answered by the free-first failover chain rather than paid primaries, which kept the bill near zero at the cost of the attribution problem in section 9. A properly-routed rerun makes inference the dominant marginal cost of the LLM arms, and per-verdict cost accounting will ship with it.

16. Which entry signals actually separated winners from losers

Every trade card below carries the intel engine's read of the coin at entry. Pooling the 76 closed trades with full intel coverage (20 winners) and comparing median entry-time signals:

signal at entry	median, winning trades	median, losing trades	separates?
intel quality score (0-100)	43	31	yes, +12 points
organic-demand score (0-1)	0.60	0.47	yes, +0.13
unique buyers observed	3	3	no
unique sellers observed	0	1	no
top-10 wallet concentration	1	1	no (saturated at our entry timing)

Two composite signals separate; three raw counts do not. The raw counts fail for an instructive reason: at the speed we enter (section 3), most coins have shown a handful of buyers regardless of their future, so the count is uninformative at that instant; concentration is near 1.0 for everything seconds after launch. The composites (quality, organic) already fold in timing-adjusted structure, which is exactly the kind of feature engineering the Oracle's conviction score does at a larger scale, and section 11 shows what that buys. The caveat is the usual one: 76 trades with 20 winners is a small sample; treat this table as hypothesis-generating, and note that the same comparison on the full labeled population is exactly what the intel engine's weekly weight training already does with n in the tens of thousands.

17. The two silent failures that faked autonomy for two days

This experiment is supposed to be self-improving: an optimizer cron reads each arm's real record every six hours and tunes that arm's own knobs inside hard bounds, and two more crons feed realized trade outcomes back into the Oracle's training data. When we audited the loop after two days of "running", we found both halves silently dead. If you build autonomous systems, these two bugs are worth your time.

Failure one: a driver quirk that turned every write into a no-op. The optimizer applied its tunings with a dynamically-selected column name. Our Postgres driver happily accepts a dynamic identifier in WHERE position but not in SET position, where it quietly parameterizes the column name and produces invalid SQL:

// The bug: the neon serverless driver accepts a dynamic identifier in
// WHERE and FROM position, but NOT in SET position. This line type-checks,
// reads clean in review, and throws "syntax error at or near $1" at runtime:
await sql`update agent_sniper_strategies set ${sql(p.field)} = ${p.to} ...`;

// The per-arm try/catch swallowed the throw, skipped the audit insert, and
// the cron returned 200. Zero rows ever landed in the audit table. The fix
// enumerates every whitelisted column explicitly:
async function applyProposal(strategyId, field, value) {
  switch (field) {
    case 'take_profit_pct':
      return sql`update agent_sniper_strategies
                 set take_profit_pct = ${value}, updated_at = now()
                 where id = ${strategyId}`;
    case 'trailing_stop_pct': /* ... one explicit case per allowed field */
    default:
      throw new Error(`optimizer refused to write unknown field '${field}'`);
  }
}

The throw was swallowed by the per-arm error handler, the audit-row insert was skipped along with it, and the cron returned HTTP 200 every six hours. From every dashboard, the self-improvement loop was healthy. In reality it had never applied a single change.

Failure two: a cron that existed in config but not in the scheduler. The Oracle feedback jobs were registered in the app's cron table and deployed, but the cloud scheduler jobs that actually fire them were never created; adding a cron to the config file does not conjure the scheduler entry. Result: the "Oracle learns from realized PnL" bridge ran zero times while looking fully shipped in the codebase.

Both are fixed, and both produced the same postmortem rule: an autonomous system's health is measured by rows in its audit tables, not by its HTTP status codes. "The cron returns 200" and "the loop is learning" are unrelated statements. We now verify every closed loop by counting its side effects, and the optimizer's first real applied run is in the dataset below (it throttled oracle-open's position size after 14 trades at a 7% win rate, and tightened llm-grok's take-profit from 60 to 45 because winners kept dying on the clock before reaching 60).

18. The fleet's own decisions so far

Since the two silent failures were fixed and deployed, the self-improvement layers have started producing an audited paper trail of their own. Everything below was decided by the system, not by us, and each row links back to bounded rules described in the engineering doc.

The intra-arm optimizer has applied 4 tunings:

oracle-open, 2026-07-23 03:25 UTC: per_trade_lamports 10000000 → 8000000. Sustained underperformance (7% win rate, net -0.017 SOL over 14 trades): throttle size. Consider disabling this arm.
llm-grok, 2026-07-23 03:25 UTC: take_profit_pct 60 → 45. Winners still timing out with take-profit at 60%: lower it toward the realized average (+21.1%) so it actually triggers.
oracle-open, 2026-07-23 06:00 UTC: per_trade_lamports 8000000 → 6400000. Sustained underperformance (7% win rate, net -0.017 SOL over 14 trades): throttle size. Consider disabling this arm.
llm-grok, 2026-07-23 06:00 UTC: min_oracle_score (unset) → 30, take_profit_pct 45 → 30. Realized wins concentrate at Oracle conviction >= 30 over 12 trades: raise the conviction floor to buy where this arm actually wins.

The portfolio-evolution layer has applied 28 budget reallocations across 10 arms, shifting daily budget toward arms with higher Wilson-bounded fitness on realized returns. Every move is journaled with before/after values and the fitness evidence that drove it; the most recent shifted rules-wide-band's daily budget from 0.053 to 0.048 SOL on a fitness of 0.122.

The point of publishing these is symmetry: if we are going to claim the fleet tunes itself, the tunings must be as auditable as the trades. Future editions of this article will grade these decisions against what happened after them, the same way section 20 grades the trades.

19. Interactive: replay every exit with your own rules

Disagree with the exit policy? Test yours. This replays the actual per-minute price history of 77 of the 90 trades (the ones whose full life is covered by minute candles; 11 long-lived coins only have hourly data and are excluded) through the same exit-priority ladder the live worker runs: stop-loss first, then trailing stop, then take-profit, then the clock. Each coin's price is normalized so the first candle after our entry equals 1.0, so a take-profit of +60% means the coin traded 60% above where we got in. The baseline you are measured against is not a fixed number: it is each trade's real fleet config replayed on the identical price path, so your policy and the fleet's are compared apples to apples, same coins, same candles, only the exit rules differ.

Stop-loss Trailing stop Take-profit Max hold arm trailing stop only above breakeven

loading price histories…

Honesty notes: candle-close prices, minute resolution, no slippage or fee modeling, and coins whose market died mid-hold exit at their last printed price exactly as the real fleet's quotes would collapse. Each trade's real fleet config replayed on the same price paths is the baseline. This is a policy explorer, not investment advice.

Spend a minute here before reading on; it teaches the whole thesis of section 7 by hand. Load the "let winners run" preset and watch the result get worse: on a population where most coins are dead within the hour, a patient +150% target just converts small would-be gains into timeouts and stop-losses. Then load "tight scalper" (a fast -15% stop, a +25% target, a ten-minute clock) and watch the fleet's whole loss flip to a small profit. The single most important exit-policy lever in this market is not how you cap losses; it is refusing to wait for a moonshot that, for the median coin, never comes. That is a finding you can derive yourself, from the coins' real price histories, in about thirty seconds.

20. All 90 trades, graded

Every closed real-money trade the fleet has ever made. First a sortable master index (click any column header), then the full record of every trade grouped by arm: trigger, firewall verdict, Oracle conviction, entry intel, model thesis verbatim from the ledger, what the coin did after we left, on-chain receipts, and an honest grade generated from the recorded numbers by fixed criteria, the same for every arm. All cards are expanded by default; use the toggle to collapse them.

Download the full dataset (JSON) Download the backtest price series (JSON)

#	token	arm	trigger	entry SOL	PnL %	PnL SOL	exit	hold s	conviction	peak %
1	Dronenald	rules-classic	new_mint	0.002	-28.1	-0.0006	trailing_stop	377674		0.0
2	BITS	rules-proven	new_mint	0.05	42.4	0.0212	timeout	1804	15	46.5
3	Enthuist	intel-quality	intel_confirmed	0.15	-8.0	-0.0120	timeout	135136		0.0
4	looong	rules-proven	new_mint	0.05	-26.1	-0.0130	trailing_stop	46		0.0
5	Smoke	intel-quality	intel_confirmed	0.15	-7.1	-0.0107	timeout	1806	22	0.0
6	WET	intel-quality	intel_confirmed	0.15	-2.5	-0.0037	timeout	1804		0.0
7	YES YOU	intel-quality	intel_confirmed	0.15	-2.5	-0.0037	timeout	1805		0.0
8	Meowpin	rules-proven	new_mint	0.05	-14.3	-0.0072	trailing_stop	29	25	28.3
9	KYDEN	intel-quality	intel_confirmed	0.15	-5.0	-0.0075	timeout	1802	17	0.0
10	SADCAT	intel-quality	intel_confirmed	0.15	-2.5	-0.0037	timeout	1806	35	0.0
11	Clouseau	rules-proven	new_mint	0.05	-13.4	-0.0067	trailing_stop	117	32	24.8
12	ANSEM	oracle-open	new_mint	0.01	-4.5	-0.0004	timeout	1806	17	0.0
13	Jimothy	oracle-open	new_mint	0.01	-57.4	-0.0057	stop_loss	15		0.0
14	PJT	oracle-open	new_mint	0.01	-7.4	-0.0007	timeout	1804		0.0
15	$DELU	oracle-open	new_mint	0.01	-2.5	-0.0002	timeout	1805	17	4.2
16	Poze	oracle-open	new_mint	0.01	-38.9	-0.0039	stop_loss	63	31	19.3
17	MELON	oracle-open	new_mint	0.01	-2.5	-0.0003	timeout	1803	19	0.0
18	TrippleBBL	oracle-open	new_mint	0.01	10.4	0.0010	timeout	1805	19	10.4
19	SKE	oracle-open	new_mint	0.01	-32.5	-0.0033	stop_loss	29	20	0.0
20	test	oracle-open	new_mint	0.01	-2.5	-0.0002	timeout	1805	15	0.0
21	JARRETT	oracle-open	new_mint	0.01	-2.5	-0.0002	timeout	1806	17	0.0
22	House	llm-auto	intel_confirmed	0.01	73.9	0.0074	take_profit	68	30	73.9
23	AG	rules-proven	new_mint	0.05	-25.3	-0.0126	timeout	1897	28	0.0
24	ROGERS	llm-claude	intel_confirmed	0.01	-2.7	-0.0003	timeout	1806		0.0
25	ROGERS	llm-grok	intel_confirmed	0.01	-5.6	-0.0006	timeout	1809		0.0
26	GABRIEL	llm-auto	intel_confirmed	0.01	-4.4	-0.0004	timeout	1803		0.0
27	ocFkqj	boost-ride	graduation_ride	0.01	5.2	0.0005	timeout	245	21	5.2
28	HOME	llm-claude	llm_intel	0.01	-3.0	-0.0003	timeout	1802		0.0
29	HOME	llm-auto	llm_intel	0.01	-3.1	-0.0003	timeout	1804		0.0
30	xjcfSQ	boost-ride	graduation_ride	0.01	4.7	0.0005	timeout	243	21	4.7
31	Ux3UY8	boost-ride	graduation_ride	0.01	20.9	0.0021	take_profit	153	21	20.2
32	2UDry5	boost-ride	graduation_ride	0.01	-84.3	-0.0084	stop_loss	6		0.0
33	one	llm-claude	llm_intel	0.01	-21.7	-0.0022	trailing_stop	133		1.2
34	one	llm-auto	llm_intel	0.01	-21.7	-0.0022	trailing_stop	136		1.2
35	掌握	llm-auto	llm_intel	0.01	58.8	0.0059	take_profit	84		64.3
36	Elon	llm-auto	llm_intel	0.01	-37.4	-0.0037	stop_loss	23	44	0.0
37	DBrD6t	boost-ride	graduation_ride	0.01	-15.8	-0.0016	stop_loss	5	16	0.0
38	Pink	llm-claude	llm_intel	0.01	-2.7	-0.0003	timeout	1804	38	0.0
39	Pink	llm-auto	llm_intel	0.01	-2.8	-0.0003	timeout	1806	38	0.0
40	Pink	llm-grok	llm_intel	0.01	-2.8	-0.0003	timeout	1808	38	0.0
41	HMimvR	boost-ride	graduation_ride	0.01	-20.1	-0.0020	stop_loss	8	21	0.0
42	Y4HEU5	boost-ride	graduation_ride	0.01	11.7	0.0012	timeout	245		11.5
43	WhiteBull	llm-auto	llm_intel	0.01	60.8	0.0061	take_profit	172		60.8
44	DEPE	llm-grok	llm_intel	0.01	71.3	0.0071	take_profit	112		81.9
45	CQUACK	llm-claude	llm_intel	0.01	-2.7	-0.0003	timeout	1806		0.0
46	MrSue	oracle-open	new_mint	0.01	-2.5	-0.0003	timeout	1806	19	0.0
47	aQzS8b	boost-ride	graduation_ride	0.01	12.0	0.0012	timeout	244		12.0
48	FLIPPENING	oracle-open	new_mint	0.01	-18.7	-0.0019	timeout	1804		0.0
49	Moon	llm-auto	llm_intel	0.01	-4.1	-0.0004	timeout	1807		0.0
50	JMBush	boost-ride	graduation_ride	0.01	5.0	0.0005	timeout	246		5.0
51	yMDueA	boost-ride	graduation_ride	0.01	-99.9	-0.0100	stop_loss	177	21	4.6
52	Diamond	llm-claude	llm_intel	0.01	-2.5	-0.0002	timeout	1803	17	0.0
53	6dAqWB	boost-ride	graduation_ride	0.01	-22.0	-0.0022	stop_loss	18	18	0.0
54	tutu	llm-auto	llm_intel	0.01	-17.6	-0.0018	trailing_stop	121		18.4
55	fxppSy	boost-ride	graduation_ride	0.01	11.3	0.0011	timeout	247	21	11.3
56	Awyjqw	boost-ride	graduation_ride	0.01	-17.4	-0.0017	stop_loss	7		0.0
57	5V3eh6	boost-ride	graduation_ride	0.01	-98.7	-0.0099	stop_loss	8	21	0.0
58	ChRFYk	boost-ride	graduation_ride	0.01	-93.2	-0.0093	stop_loss	9	22	0.0
59	qFJmT1	boost-ride	graduation_ride	0.01	20.4	0.0020	take_profit	206	21	20.4
60	4Wkuhs	boost-ride	graduation_ride	0.01	-8.7	-0.0009	trailing_stop	19	19	0.0
61	u8oLcb	boost-ride	graduation_ride	0.01	20.2	0.0020	take_profit	224		20.2
62	birry	llm-claude	llm_intel	0.01	-3.7	-0.0004	timeout	1803	28	0.0
63	BLACKHOLE	llm-grok	llm_intel	0.01	63.1	0.0063	take_profit	114		63.1
64	boss	llm-auto	llm_intel	0.01	-4.7	-0.0005	timeout	1803	36	0.0
65	$RUGHUNTER	llm-auto	llm_intel	0.01	-2.5	-0.0002	timeout	1802	17	0.0
66	RCC	llm-auto	llm_intel	0.01	-2.5	-0.0002	timeout	1806	16	0.0
67	COOK	llm-grok	llm_intel	0.01	14.2	0.0014	timeout	1804		14.2
68	doge	llm-grok	llm_intel	0.01	-4.1	-0.0004	timeout	1804	17	0.0
69	REC	llm-grok	llm_intel	0.01	-2.6	-0.0003	timeout	1804	20	0.0
70	DUCK	llm-grok	llm_intel	0.01	65.6	0.0066	take_profit	72	36	65.6
71	Aevon	boost-ride	graduation_ride	0.01	2.0	0.0002	timeout	243		4.8
72	Jrool	oracle-open	new_mint	0.01	-2.5	-0.0002	timeout	1804	21	0.0
73	Walter	boost-ride	graduation_ride	0.01	-2.3	-0.0002	trailing_stop	15		0.0
74	RACOIN	llm-claude	llm_intel	0.01	-78.3	-0.0078	stop_loss	128	38	55.0
75	VFX	boost-ride	graduation_ride	0.01	18.7	0.0019	take_profit	179		20.7
76	FLL	boost-ride	graduation_ride	0.01	20.0	0.0020	take_profit	164		20.0
77	Free	oracle-open	new_mint	0.01	-2.5	-0.0002	timeout	1807	17	0.0
78	SAVIOUR	llm-auto	llm_intel	0.01	-2.5	-0.0002	timeout	1806	17	0.0
79	MCP	llm-claude	llm_intel	0.01	10.6	0.0011	trailing_stop	61		38.6
80	RWA	boost-ride	graduation_ride	0.01	-35.6	-0.0036	stop_loss	116		10.6
81	COREY	llm-auto	llm_intel	0.01	-26.4	-0.0026	trailing_stop	22		0.0
82	APE	llm-grok	llm_intel	0.01	-9.2	-0.0009	timeout	1802	13	0.0
83	CWARDIN	boost-ride	graduation_ride	0.01	-32.8	-0.0033	stop_loss	191		16.8
84	BENNER	boost-ride	graduation_ride	0.01	-34.3	-0.0034	stop_loss	188		15.9
85	AGIGUY	boost-ride	graduation_ride	0.01	-34.4	-0.0034	stop_loss	190		15.9
86	DrqyFF	boost-ride	graduation_ride	0.01	7.7	0.0008	timeout	243		7.7
87	moHcoC	boost-ride	graduation_ride	0.01	-6.4	-0.0006	trailing_stop	114	21	1.9
88	HANU.p	boost-ride	graduation_ride	0.01	42.9	0.0043	take_profit	293		20.3
89	$MRH	boost-ride	graduation_ride	0.01	-30.1	-0.0030	stop_loss	254		10.1
90	BDAG	boost-ride	graduation_ride	0.01	20.4	0.0020	take_profit	135		20.6

llm-grok LLM-judged

Wallet HBgN..m8Ca · decision ledger · 9 trades · 4 wins (44%) · net +0.0190 SOL · avg winner +53.6% · avg loser -4.9% · median hold 30m

Policy at entry: 0.01 SOL per trade · take-profit +45% · stop-loss -30% · trailing stop 20% · 30m max hold

#25 ROGERS -5.6% timeout 2026-07-21 13:17

Opened	2026-07-21 13:17 UTC, committed 0.01 SOL
Trigger	intel_confirmed: the intel engine watched the first minutes of trading and confirmed real two-sided demand
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 45/100 · 0 buyers / 0 sellers observed · organic 45% · top-10 hold 0%
Peak	+0.0%
Exit	`timeout` after 30m, realized -5.6% (-0.0006 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,245 cap, now ~$2,170); never traded again after our exit (no later candles exist); lifetime label: `rugged`, ATH cap ~$2,203
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $ROGERS on a intel_confirmed trigger with 0.04% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: Scratched out near flat (-5.6%) instead of riding a dead coin further down. The token never printed another trade after this exit; the agent was the last one out before the market vanished. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

What should improve: The coin flatlined almost immediately; holding the full 30 minutes tied up budget for nothing. A liquidity-decay early exit would free the slot in a fraction of the time.

#40 Pink -2.8% timeout 2026-07-21 20:58

Opened	2026-07-21 20:58 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Oracle conviction	38/100 at entry
Firewall	warn (score 86), sell simulated on-chain before the buy
Model	`fallback:ovh`, confidence 0.8
Entry intel	quality 34/100 · 89 buyers / 69 sellers observed · organic 52% · top-10 hold 53% · dev sold during observation
Peak	+0.0%
Exit	`timeout` after 30m, realized -2.8% (-0.0003 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,253 cap, now ~$2,172); highest print after our exit: ~$2,183 cap (0.42x our exit, 1m candles)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “Pink grasshopper has a relatively high market realness score indicating genuine two-sided market demand.”

What it did well: Scratched out near flat (-2.8%) instead of riding a dead coin further down.

What should improve: The coin flatlined almost immediately; holding the full 30 minutes tied up budget for nothing. A liquidity-decay early exit would free the slot in a fraction of the time. The model judged this 80% confident and was wrong; theses that cite "memetic quality" alone keep underperforming ones that cite crowd or creator evidence.

#44 DEPE +71.3% take_profit 2026-07-22 00:20

Opened	2026-07-22 00:20 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Model	`fallback:groq#instant`, confidence 0.8
Entry intel	quality 95/100 · 51 buyers / 31 sellers observed · organic 84% · top-10 hold 90%
Peak	+81.9%
Exit	`take_profit` after 2m, realized +71.3% (+0.0071 SOL)
After we left	trades at 0.07x our exit today (below where we sold; we sold at an implied ~$37,396 cap, now ~$2,469); highest print after our exit: ~$19,566 cap (0.52x our exit, 1m candles)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “Strong real-market signal with multiple buyers and substantial wallet holding”

What it did well: Hit the configured +45% take-profit and sold the whole position into strength. This is the exit ladder doing exactly what it exists for. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry. The model's entry thesis was directionally right on this one.

#63 BLACKHOLE +63.1% take_profit 2026-07-22 05:29

Opened	2026-07-22 05:29 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Model	`fallback:ovh`, confidence 0.8
Entry intel	quality 83/100 · 54 buyers / 33 sellers observed · organic 82% · top-10 hold 87%
Peak	+63.1%
Exit	`take_profit` after 2m, realized +63.1% (+0.0063 SOL)
After we left	trades at 0.07x our exit today (below where we sold; we sold at an implied ~$29,585 cap, now ~$2,194); highest print after our exit: ~$14,830 cap (0.50x our exit, 1m candles)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “Genuine two-sided market indicated by high market_realness score.”

#67 COOK +14.2% timeout 2026-07-22 14:16

Opened	2026-07-22 14:16 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Firewall	warn (score 84), sell simulated on-chain before the buy
Model	`fallback:ovh`, confidence 0.8
Entry intel	quality 20/100 · 3 buyers / 1 sellers observed · organic 43% · top-10 hold 100%
Peak	+14.2%
Exit	`timeout` after 30m, realized +14.2% (+0.0014 SOL)
After we left	trades at 0.36x our exit today (below where we sold; we sold at an implied ~$6,867 cap, now ~$2,491); highest print after our exit: ~$2,841 cap (0.41x our exit, 1m candles)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “Strong memetic potential with a catchy name and timely narrative.”

What it did well: Closed +14.2% in the green. The model's entry thesis was directionally right on this one.

What should improve: This exit was luck-shaped: the 30-minute clock expired while the position happened to be up. Nothing locked the gain in. A reachable take-profit or a momentum-fade exit converts this from luck into policy.

#68 doge -4.1% timeout 2026-07-22 15:13

Opened	2026-07-22 15:13 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Oracle conviction	17/100 at entry
Firewall	warn (score 84), sell simulated on-chain before the buy
Model	`fallback:ovh`, confidence 0.8
Entry intel	quality 30/100 · 3 buyers / 1 sellers observed · organic 52% · top-10 hold 100%
Peak	+0.0%
Exit	`timeout` after 30m, realized -4.1% (-0.0004 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,248 cap, now ~$2,170); never traded again after our exit (no later candles exist)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “Strong memetic potential from 'swole' and 'doge' combination”

What it did well: Scratched out near flat (-4.1%) instead of riding a dead coin further down. The token never printed another trade after this exit; the agent was the last one out before the market vanished.

#69 REC -2.6% timeout 2026-07-22 17:29

Opened	2026-07-22 17:29 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Oracle conviction	20/100 at entry
Firewall	warn (score 84), sell simulated on-chain before the buy
Model	`fallback:groq#instant`, confidence 0.7
Entry intel	quality 23/100 · 4 buyers / 2 sellers observed · organic 46% · top-10 hold 100%
Peak	+0.0%
Exit	`timeout` after 30m, realized -2.6% (-0.0003 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,248 cap, now ~$2,170); never traded again after our exit (no later candles exist)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “REC exhibits strong market_realness signals and timely narrative”

What it did well: Scratched out near flat (-2.6%) instead of riding a dead coin further down. The token never printed another trade after this exit; the agent was the last one out before the market vanished.

What should improve: The coin flatlined almost immediately; holding the full 30 minutes tied up budget for nothing. A liquidity-decay early exit would free the slot in a fraction of the time.

#70 DUCK +65.6% take_profit 2026-07-22 22:36

Opened	2026-07-22 22:36 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Oracle conviction	36/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Model	`fallback:ovh`, confidence 0.7
Entry intel	quality 78/100 · 48 buyers / 24 sellers observed · organic 78% · top-10 hold 86%
Peak	+65.6%
Exit	`take_profit` after 72s, realized +65.6% (+0.0066 SOL)
After we left	trades at 0.04x our exit today (below where we sold; we sold at an implied ~$49,102 cap, now ~$2,185); highest print after our exit: ~$40,087 cap (0.82x our exit, 1m candles); completed the bonding curve after our entry
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “DUCK has a decent market_realness score and a relatively balanced buyer-seller ratio.”

#82 APE -9.2% timeout 2026-07-23 01:21

Opened	2026-07-23 01:21 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Oracle conviction	13/100 at entry
Firewall	warn (score 70), sell simulated on-chain before the buy
Model	`fallback:groq#instant`, confidence 0.8
Entry intel	quality 16/100 · 24 buyers / 20 sellers observed · organic 46% · top-10 hold 97% · dev sold during observation
Peak	+0.0%
Exit	`timeout` after 30m, realized -9.2% (-0.0009 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,313 cap, now ~$2,197); never traded again after our exit (no later candles exist)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “A moderate momentum position in APE is worth taking due to a decent two-sided market and a sizable dev buy.”

What it did well: Scratched out near flat (-9.2%) instead of riding a dead coin further down. The token never printed another trade after this exit; the agent was the last one out before the market vanished.

llm-auto LLM-judged

Wallet HXDz..3F3s · decision ledger · 15 trades · 3 wins (20%) · net +0.0064 SOL · avg winner +64.5% · avg loser -10.8% · median hold 30m

Policy at entry: 0.01 SOL per trade · take-profit +60% · stop-loss -30% · trailing stop 20% · 30m max hold

#22 House +73.9% take_profit 2026-07-21 13:18

Opened	2026-07-21 13:18 UTC, committed 0.01 SOL
Trigger	intel_confirmed: the intel engine watched the first minutes of trading and confirmed real two-sided demand
Oracle conviction	30/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 79/100 · 39 buyers / 22 sellers observed · organic 80% · top-10 hold 81%
Peak	+73.9%
Exit	`take_profit` after 68s, realized +73.9% (+0.0074 SOL)
After we left	trades at 0.16x our exit today (below where we sold; we sold at an implied ~$13,784 cap, now ~$2,170); highest print after our exit: ~$5,889 cap (0.43x our exit, 1m candles); lifetime label: `rugged`, ATH cap ~$7,503
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $HOUSE on a intel_confirmed trigger with 0.02% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: Hit the configured +60% take-profit and sold the whole position into strength. This is the exit ladder doing exactly what it exists for. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

#26 GABRIEL -4.4% timeout 2026-07-21 13:19

Opened	2026-07-21 13:19 UTC, committed 0.01 SOL
Trigger	intel_confirmed: the intel engine watched the first minutes of trading and confirmed real two-sided demand
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 13/100 · 5 buyers / 0 sellers observed · organic 39% · top-10 hold 100%
Peak	+0.0%
Exit	`timeout` after 30m, realized -4.4% (-0.0004 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,247 cap, now ~$2,170); never traded again after our exit (no later candles exist); lifetime label: `rugged`, ATH cap ~$3,263
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $GABRIEL on a intel_confirmed trigger with 0.04% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: Scratched out near flat (-4.4%) instead of riding a dead coin further down. The token never printed another trade after this exit; the agent was the last one out before the market vanished. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

What should improve: The coin flatlined almost immediately; holding the full 30 minutes tied up budget for nothing. A liquidity-decay early exit would free the slot in a fraction of the time.

#29 HOME -3.1% timeout 2026-07-21 18:59

Opened	2026-07-21 18:59 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Firewall	warn (score 84), sell simulated on-chain before the buy
Model	`fallback:ovh`, confidence 0.7
Entry intel	quality 24/100 · 3 buyers / 2 sellers observed · organic 47% · top-10 hold 100%
Peak	+0.0%
Exit	`timeout` after 30m, realized -3.1% (-0.0003 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,250 cap, now ~$2,171); never traded again after our exit (no later candles exist)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “HOME token has strong memetic quality and a plausible narrative.”

What it did well: Scratched out near flat (-3.1%) instead of riding a dead coin further down. The token never printed another trade after this exit; the agent was the last one out before the market vanished.

What should improve: The coin flatlined almost immediately; holding the full 30 minutes tied up budget for nothing. A liquidity-decay early exit would free the slot in a fraction of the time.

#34 one -21.7% trailing_stop 2026-07-21 19:49

Opened	2026-07-21 19:49 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Model	`fallback:ovh`, confidence 0.7
Entry intel	quality 76/100 · 52 buyers / 30 sellers observed · organic 77% · top-10 hold 87%
Peak	+1.2%
Exit	`trailing_stop` after 2m, realized -21.7% (-0.0022 SOL)
After we left	trades at 0.12x our exit today (below where we sold; we sold at an implied ~$17,495 cap, now ~$2,171); highest print after our exit: ~$10,006 cap (0.57x our exit, 1m candles)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “Genuine two-sided market and moderate market realness justify a small momentum position.”

What it did well: The trailing stop fired and got out with liquidity still present. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

What should improve: The trail armed below breakeven: the position peaked at +1.2% but the 20% giveback landed at -21.7%. Arming the trail only after the position is green would turn this exit into a scratch instead of a loss.

#35 掌握 +58.8% take_profit 2026-07-21 20:10

Opened	2026-07-21 20:10 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Model	`fallback:ovh`, confidence 0.7
Entry intel	quality 78/100 · 41 buyers / 22 sellers observed · organic 77% · top-10 hold 93%
Peak	+64.3%
Exit	`take_profit` after 84s, realized +58.8% (+0.0059 SOL)
After we left	trades at 0.05x our exit today (below where we sold; we sold at an implied ~$40,784 cap, now ~$1,976); highest print after our exit: ~$33,559 cap (0.82x our exit, 1m candles); completed the bonding curve after our entry
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “Genuine two-sided market and moderate market realness score justify a small momentum position.”

#36 Elon -37.4% stop_loss 2026-07-21 20:24

Opened	2026-07-21 20:24 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Oracle conviction	44/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Model	`fallback:ovh`, confidence 0.7
Entry intel	quality 82/100 · 37 buyers / 22 sellers observed · organic 83% · top-10 hold 84%
Peak	+0.0%
Exit	`stop_loss` after 23s, realized -37.4% (-0.0037 SOL)
After we left	trades at 0.17x our exit today (below where we sold; we sold at an implied ~$12,994 cap, now ~$2,195); highest print after our exit: ~$13,306 cap (1.02x our exit, 1m candles)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “Strong memetic quality and decent market realness justify a small momentum position.”

What it did well: Loss capped close to policy: -30% configured, -37.4% realized. The mandatory stop did its one job. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

What should improve: Entry selection, not exit policy, is what loses here: the coin never traded above the entry quote for long.

#39 Pink -2.8% timeout 2026-07-21 20:58

Opened	2026-07-21 20:58 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Oracle conviction	38/100 at entry
Firewall	warn (score 86), sell simulated on-chain before the buy
Model	`fallback:ovh`, confidence 0.8
Entry intel	quality 34/100 · 89 buyers / 69 sellers observed · organic 52% · top-10 hold 53% · dev sold during observation
Peak	+0.0%
Exit	`timeout` after 30m, realized -2.8% (-0.0003 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,257 cap, now ~$2,172); highest print after our exit: ~$2,183 cap (0.42x our exit, 1m candles)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “Pink grasshopper has a relatively real market with diverse buyers and sellers.”

What it did well: Scratched out near flat (-2.8%) instead of riding a dead coin further down.

#43 WhiteBull +60.8% take_profit 2026-07-22 00:16

Opened	2026-07-22 00:16 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Firewall	warn (score 86), sell simulated on-chain before the buy
Model	`fallback:ovh`, confidence 0.7
Entry intel	quality 30/100 · 37 buyers / 12 sellers observed · organic 50% · top-10 hold 61% · dev sold during observation
Peak	+60.8%
Exit	`take_profit` after 3m, realized +60.8% (+0.0061 SOL)
After we left	trades at 0.05x our exit today (below where we sold; we sold at an implied ~$41,664 cap, now ~$2,170); highest print after our exit: ~$19,796 cap (0.48x our exit, 1m candles)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “Moderately real market with diverse buyers warrants a small momentum position.”

What it did well: Hit the configured +60% take-profit and sold the whole position into strength. This is the exit ladder doing exactly what it exists for. The model's entry thesis was directionally right on this one.

#49 Moon -4.1% timeout 2026-07-22 00:42

Opened	2026-07-22 00:42 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Firewall	warn (score 84), sell simulated on-chain before the buy
Model	`fallback:ovh`, confidence 0.8
Entry intel	quality 4/100 · 1 buyers / 0 sellers observed · organic 28% · top-10 hold 100%
Peak	+0.0%
Exit	`timeout` after 30m, realized -4.1% (-0.0004 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,248 cap, now ~$2,170); never traded again after our exit (no later candles exist)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “Memetic quality of 'TO THE MOON' is high and could attract significant attention”

#54 tutu -17.6% trailing_stop 2026-07-22 02:12

Opened	2026-07-22 02:12 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Model	`fallback:ovh`, confidence 0.7
Entry intel	quality 83/100 · 52 buyers / 33 sellers observed · organic 79% · top-10 hold 85%
Peak	+18.4%
Exit	`trailing_stop` after 2m, realized -17.6% (-0.0018 SOL)
After we left	trades at 0.16x our exit today (below where we sold; we sold at an implied ~$14,169 cap, now ~$2,319); highest print after our exit: ~$11,091 cap (0.78x our exit, 1m candles)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “Moderately genuine market_realness and decent buyer distribution justify a momentum position.”

What should improve: The trail armed below breakeven: the position peaked at +18.4% but the 20% giveback landed at -17.6%. Arming the trail only after the position is green would turn this exit into a scratch instead of a loss.

#64 boss -4.7% timeout 2026-07-22 05:48

Opened	2026-07-22 05:48 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Oracle conviction	36/100 at entry
Firewall	warn (score 86), sell simulated on-chain before the buy
Model	`fallback:ovh`, confidence 0.8
Entry intel	quality 39/100 · 44 buyers / 36 sellers observed · organic 56% · top-10 hold 79% · dev sold during observation
Peak	+0.0%
Exit	`timeout` after 30m, realized -4.7% (-0.0005 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,269 cap, now ~$2,175); highest print after our exit: ~$2,181 cap (0.41x our exit, 1m candles)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “Strong market realness and diverse buyer base justify a momentum position.”

What it did well: Scratched out near flat (-4.7%) instead of riding a dead coin further down.

#65 $RUGHUNTER -2.5% timeout 2026-07-22 08:34

Opened	2026-07-22 08:34 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Oracle conviction	17/100 at entry
Firewall	warn (score 84), sell simulated on-chain before the buy
Model	`fallback:ovh`, confidence 0.8
Entry intel	quality 4/100 · 1 buyers / 0 sellers observed · organic 28% · top-10 hold 100%
Peak	+0.0%
Exit	`timeout` after 30m, realized -2.5% (-0.0002 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,252 cap, now ~$2,172); never traded again after our exit (no later candles exist)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “Strong memetic name and timely narrative suggest potential for real momentum.”

#66 RCC -2.5% timeout 2026-07-22 09:42

Opened	2026-07-22 09:42 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Oracle conviction	16/100 at entry
Firewall	warn (score 84), sell simulated on-chain before the buy
Model	`fallback:ovh`, confidence 0.8
Entry intel	quality 0/100 · 3 buyers / 2 sellers observed · organic 41% · bundle risk 53% · top-10 hold 100%
Peak	+0.0%
Exit	`timeout` after 30m, realized -2.5% (-0.0002 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,248 cap, now ~$2,170); never traded again after our exit (no later candles exist)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “Memetic token name and low market cap may attract speculative buyers.”

#78 SAVIOUR -2.5% timeout 2026-07-23 00:50

Opened	2026-07-23 00:50 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Oracle conviction	17/100 at entry
Firewall	warn (score 84), sell simulated on-chain before the buy
Model	`fallback:groq#instant`, confidence 0.7
Entry intel	quality 4/100 · 1 buyers / 0 sellers observed · organic 28% · top-10 hold 100%
Peak	+0.0%
Exit	`timeout` after 30m, realized -2.5% (-0.0002 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,248 cap, now ~$2,170); never traded again after our exit (no later candles exist)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “SAVIOR may have organic demand due to a compelling narrative and the creator's launch history”

What should improve: The coin flatlined almost immediately; holding the full 30 minutes tied up budget for nothing. A liquidity-decay early exit would free the slot in a fraction of the time.

#81 COREY -26.4% trailing_stop 2026-07-23 01:40

Opened	2026-07-23 01:40 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Firewall	warn (score 86), sell simulated on-chain before the buy
Model	`fallback:ovh`, confidence 0.7
Entry intel	quality 15/100 · 41 buyers / 27 sellers observed · organic 45% · top-10 hold 87% · dev sold during observation
Peak	+0.0%
Exit	`trailing_stop` after 22s, realized -26.4% (-0.0026 SOL)
After we left	trades at 0.32x our exit today (below where we sold; we sold at an implied ~$7,306 cap, now ~$2,369); highest print after our exit: ~$2,522 cap (0.35x our exit, 1m candles)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “Genuine two-sided market and moderate market realness score indicate potential for real momentum.”

What it did well: The trailing stop fired and got out with liquidity still present.

What should improve: The trail armed below breakeven: the position peaked at +0.0% but the 20% giveback landed at -26.4%. Arming the trail only after the position is green would turn this exit into a scratch instead of a loss.

rules-classic rules

Wallet 7KdH..dWfD · decision ledger · 1 trades · 0 wins (0%) · net -0.0006 SOL · avg winner +0.0% · avg loser -28.1% · median hold 4.4d

Policy at entry: 0.002 SOL per trade · take-profit +60% · stop-loss -30% · trailing stop 20% · 30m max hold · Oracle conviction ≥ 55 · socials required

#1 Dronenald -28.1% trailing_stop 2026-07-03 16:19

Opened	2026-07-03 16:19 UTC, committed 0.002 SOL
Trigger	new_mint: sniped seconds after the launch appeared on the live pump.fun firehose
Firewall	allow (score 100), sell simulated on-chain before the buy
Peak	+0.0%
Exit	`trailing_stop` after 4.4d, realized -28.1% (-0.0006 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,253 cap, now ~$2,172); never traded again after our exit (no later candles exist)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $DRONENALD on a new_mint trigger with 0.01% price impact; firewall verdict allow (score 100). Committed 0.0020 SOL expecting a profitable exit.”

What it did well: The trailing stop fired and got out with liquidity still present. The token never printed another trade after this exit; the agent was the last one out before the market vanished. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

What should improve: The trail armed below breakeven: the position peaked at +0.0% but the 20% giveback landed at -28.1%. Arming the trail only after the position is green would turn this exit into a scratch instead of a loss.

llm-claude LLM-judged

Wallet 65Ka..w3h8 · decision ledger · 9 trades · 1 wins (11%) · net -0.0107 SOL · avg winner +10.6% · avg loser -14.7% · median hold 30m

Policy at entry: 0.01 SOL per trade · take-profit +60% · stop-loss -30% · trailing stop 20% · 30m max hold

#24 ROGERS -2.7% timeout 2026-07-21 13:17

Opened	2026-07-21 13:17 UTC, committed 0.01 SOL
Trigger	intel_confirmed: the intel engine watched the first minutes of trading and confirmed real two-sided demand
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 45/100 · 0 buyers / 0 sellers observed · organic 45% · top-10 hold 0%
Peak	+0.0%
Exit	`timeout` after 30m, realized -2.7% (-0.0003 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,251 cap, now ~$2,170); never traded again after our exit (no later candles exist); lifetime label: `rugged`, ATH cap ~$2,203
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $ROGERS on a intel_confirmed trigger with 0.04% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: Scratched out near flat (-2.7%) instead of riding a dead coin further down. The token never printed another trade after this exit; the agent was the last one out before the market vanished. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

What should improve: The coin flatlined almost immediately; holding the full 30 minutes tied up budget for nothing. A liquidity-decay early exit would free the slot in a fraction of the time.

#28 HOME -3.0% timeout 2026-07-21 18:59

Opened	2026-07-21 18:59 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Firewall	warn (score 84), sell simulated on-chain before the buy
Model	`fallback:ovh`, confidence 0.7
Entry intel	quality 24/100 · 3 buyers / 2 sellers observed · organic 47% · top-10 hold 100%
Peak	+0.0%
Exit	`timeout` after 30m, realized -3.0% (-0.0003 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,253 cap, now ~$2,171); never traded again after our exit (no later candles exist)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “Strong memetic quality and timely narrative align with market sentiment.”

What it did well: Scratched out near flat (-3.0%) instead of riding a dead coin further down. The token never printed another trade after this exit; the agent was the last one out before the market vanished.

What should improve: The coin flatlined almost immediately; holding the full 30 minutes tied up budget for nothing. A liquidity-decay early exit would free the slot in a fraction of the time.

#33 one -21.7% trailing_stop 2026-07-21 19:49

Opened	2026-07-21 19:49 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Model	`fallback:ovh`, confidence 0.7
Entry intel	quality 76/100 · 52 buyers / 30 sellers observed · organic 77% · top-10 hold 87%
Peak	+1.2%
Exit	`trailing_stop` after 2m, realized -21.7% (-0.0022 SOL)
After we left	trades at 0.12x our exit today (below where we sold; we sold at an implied ~$17,500 cap, now ~$2,171); highest print after our exit: ~$10,006 cap (0.57x our exit, 1m candles)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “Moderately strong market realness and balanced buyer/seller ratio justify a momentum position.”

#38 Pink -2.7% timeout 2026-07-21 20:58

Opened	2026-07-21 20:58 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Oracle conviction	38/100 at entry
Firewall	warn (score 86), sell simulated on-chain before the buy
Model	`fallback:ovh`, confidence 0.8
Entry intel	quality 34/100 · 89 buyers / 69 sellers observed · organic 52% · top-10 hold 53% · dev sold during observation
Peak	+0.0%
Exit	`timeout` after 30m, realized -2.7% (-0.0003 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,260 cap, now ~$2,172); highest print after our exit: ~$2,183 cap (0.42x our exit, 1m candles)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “Pink grasshopper has a relatively real market with diverse buyers and sellers.”

What it did well: Scratched out near flat (-2.7%) instead of riding a dead coin further down.

#45 CQUACK -2.7% timeout 2026-07-21 23:57

Opened	2026-07-21 23:57 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Firewall	warn (score 84), sell simulated on-chain before the buy
Model	`fallback:ovh`, confidence 0.7
Entry intel	quality 7/100 · 2 buyers / 1 sellers observed · organic 31% · top-10 hold 100%
Peak	+0.0%
Exit	`timeout` after 30m, realized -2.7% (-0.0003 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,248 cap, now ~$2,170); never traded again after our exit (no later candles exist)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “CQUACK's memetic name and symbol may attract initial speculative interest.”

What should improve: The coin flatlined almost immediately; holding the full 30 minutes tied up budget for nothing. A liquidity-decay early exit would free the slot in a fraction of the time.

#52 Diamond -2.5% timeout 2026-07-22 01:38

Opened	2026-07-22 01:38 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Oracle conviction	17/100 at entry
Firewall	warn (score 84), sell simulated on-chain before the buy
Model	`fallback:ovh`, confidence 0.8
Entry intel	quality 7/100 · 2 buyers / 1 sellers observed · organic 31% · top-10 hold 100%
Peak	+0.0%
Exit	`timeout` after 30m, realized -2.5% (-0.0002 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,258 cap, now ~$2,170); highest print after our exit: ~$2,191 cap (0.42x our exit, 1m candles)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “Strong memetic name and ticker align with current market sentiment.”

What it did well: Scratched out near flat (-2.5%) instead of riding a dead coin further down.

#62 birry -3.7% timeout 2026-07-22 03:38

Opened	2026-07-22 03:38 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Oracle conviction	28/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Model	`fallback:groq#instant`, confidence 0.7
Entry intel	quality 84/100 · 51 buyers / 46 sellers observed · organic 84% · top-10 hold 73%
Peak	+0.0%
Exit	`timeout` after 30m, realized -3.7% (-0.0004 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,248 cap, now ~$2,170); never traded again after our exit (no later candles exist)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “Strong momentum with genuine market demand”

What it did well: Scratched out near flat (-3.7%) instead of riding a dead coin further down. The token never printed another trade after this exit; the agent was the last one out before the market vanished. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

What should improve: The coin flatlined almost immediately; holding the full 30 minutes tied up budget for nothing. A liquidity-decay early exit would free the slot in a fraction of the time.

#74 RACOIN -78.3% stop_loss 2026-07-23 00:37

Opened	2026-07-23 00:37 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Oracle conviction	38/100 at entry
Firewall	warn (score 86), sell simulated on-chain before the buy
Model	`fallback:ovh`, confidence 0.8
Entry intel	quality 26/100 · 43 buyers / 8 sellers observed · organic 49% · top-10 hold 59% · dev sold during observation
Peak	+55.0%
Exit	`stop_loss` after 2m, realized -78.3% (-0.0078 SOL)
After we left	trades at 0.39x our exit today (below where we sold; we sold at an implied ~$5,629 cap, now ~$2,171); highest print after our exit: ~$2,192 cap (0.39x our exit, 1m candles)
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “RACOIN has a relatively genuine market with diverse buyers and sellers.”

What it did well: It still sold. Getting any exit through a collapsing bonding curve beats holding a dead token.

What should improve: The stop was set at -30% but the realized loss was -78.3%: the curve's liquidity collapsed faster than the position poller could quote and sell. A stop-loss is a request, not a guarantee. Smaller size in thin curves and a faster quote cadence are the real defenses. The model judged this 80% confident and was wrong; theses that cite "memetic quality" alone keep underperforming ones that cite crowd or creator evidence.

#79 MCP +10.6% trailing_stop 2026-07-23 01:24

Opened	2026-07-23 01:24 UTC, committed 0.01 SOL
Trigger	llm_intel: the intel engine surfaced it, then the strategy’s LLM judge approved the entry
Firewall	warn (score 86), sell simulated on-chain before the buy
Model	`fallback:ovh`, confidence 0.8
Entry intel	quality 37/100 · 108 buyers / 34 sellers observed · organic 55% · top-10 hold 32% · dev sold during observation
Peak	+38.6%
Exit	`trailing_stop` after 61s, realized +10.6% (+0.0011 SOL)
After we left	trades at 0.08x our exit today (below where we sold; we sold at an implied ~$32,977 cap, now ~$2,589); highest print after our exit: ~$45,600 cap (1.38x our exit, 1m candles); completed the bonding curve after our entry
Receipts	buy tx · sell tx · wallet

Model thesis, verbatim: “MCP Genesis has a reasonably high market_realness score indicating genuine two-sided market activity.”

What it did well: Rode the pump to a +38.6% peak and let the 20% trailing stop convert it into a realized +10.6% instead of round-tripping to zero. The model's entry thesis was directionally right on this one.

What should improve: The token printed 1.38x our exit price after we left, the closest thing to a shaken-out-early exit in the dataset.

oracle-open rules

Wallet 54tv..v23M · decision ledger · 14 trades · 1 wins (7%) · net -0.0167 SOL · avg winner +10.4% · avg loser -13.6% · median hold 30m

Policy at entry: 0.008 SOL per trade · no take-profit · stop-loss -30% · no trailing stop · 30m max hold · Oracle conviction ≥ 35

#12 ANSEM -4.5% timeout 2026-07-21 05:56

Opened	2026-07-21 05:56 UTC, committed 0.01 SOL
Trigger	new_mint: sniped seconds after the launch appeared on the live pump.fun firehose
Oracle conviction	17/100 at entry
Firewall	warn (score 70), sell simulated on-chain before the buy
Entry intel	quality 0/100 · 19 buyers / 15 sellers observed · organic 33% · top-10 hold 99% · dev sold during observation
Peak	+0.0%
Exit	`timeout` after 30m, realized -4.5% (-0.0004 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,251 cap, now ~$2,171); never traded again after our exit (no later candles exist); lifetime label: `rugged`, ATH cap ~$5,915
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $ANSEM on a new_mint trigger with 0.03% price impact; firewall verdict warn (score 70). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: Scratched out near flat (-4.5%) instead of riding a dead coin further down. The token never printed another trade after this exit; the agent was the last one out before the market vanished.

What should improve: The coin flatlined almost immediately; holding the full 30 minutes tied up budget for nothing. A liquidity-decay early exit would free the slot in a fraction of the time.

#13 Jimothy -57.4% stop_loss 2026-07-21 06:26

Opened	2026-07-21 06:26 UTC, committed 0.01 SOL
Trigger	new_mint: sniped seconds after the launch appeared on the live pump.fun firehose
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 68/100 · 52 buyers / 57 sellers observed · organic 76% · top-10 hold 100%
Peak	+0.0%
Exit	`stop_loss` after 15s, realized -57.4% (-0.0057 SOL)
After we left	trades at 0.40x our exit today (below where we sold; we sold at an implied ~$5,411 cap, now ~$2,170); never traded again after our exit (no later candles exist); lifetime label: `rugged`, ATH cap ~$5,208
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $JIMOTHY on a new_mint trigger with 0.02% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: It still sold. Getting any exit through a collapsing bonding curve beats holding a dead token. The token never printed another trade after this exit; the agent was the last one out before the market vanished. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

What should improve: The stop was set at -30% but the realized loss was -57.4%: the curve's liquidity collapsed faster than the position poller could quote and sell. A stop-loss is a request, not a guarantee. Smaller size in thin curves and a faster quote cadence are the real defenses.

#14 PJT -7.4% timeout 2026-07-21 06:26

Opened	2026-07-21 06:26 UTC, committed 0.01 SOL
Trigger	new_mint: sniped seconds after the launch appeared on the live pump.fun firehose
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 31/100 · 4 buyers / 1 sellers observed · organic 51% · top-10 hold 100%
Peak	+0.0%
Exit	`timeout` after 30m, realized -7.4% (-0.0007 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,247 cap, now ~$2,170); never traded again after our exit (no later candles exist); lifetime label: `rugged`, ATH cap ~$2,309
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $PJT on a new_mint trigger with 0.03% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: Scratched out near flat (-7.4%) instead of riding a dead coin further down. The token never printed another trade after this exit; the agent was the last one out before the market vanished. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

What should improve: The coin flatlined almost immediately; holding the full 30 minutes tied up budget for nothing. A liquidity-decay early exit would free the slot in a fraction of the time.

#15 $DELU -2.5% timeout 2026-07-21 06:56

Opened	2026-07-21 06:56 UTC, committed 0.01 SOL
Trigger	new_mint: sniped seconds after the launch appeared on the live pump.fun firehose
Oracle conviction	17/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 7/100 · 2 buyers / 1 sellers observed · organic 31% · top-10 hold 100%
Peak	+4.2%
Exit	`timeout` after 30m, realized -2.5% (-0.0002 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,248 cap, now ~$2,170); never traded again after our exit (no later candles exist); lifetime label: `rugged`, ATH cap ~$2,348
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $$DELU on a new_mint trigger with 0.04% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: Scratched out near flat (-2.5%) instead of riding a dead coin further down. The token never printed another trade after this exit; the agent was the last one out before the market vanished. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

What should improve: The coin flatlined almost immediately; holding the full 30 minutes tied up budget for nothing. A liquidity-decay early exit would free the slot in a fraction of the time.

#16 Poze -38.9% stop_loss 2026-07-21 07:27

Opened	2026-07-21 07:27 UTC, committed 0.01 SOL
Trigger	new_mint: sniped seconds after the launch appeared on the live pump.fun firehose
Oracle conviction	31/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 31/100 · 41 buyers / 40 sellers observed · organic 52% · top-10 hold 68% · dev sold during observation
Peak	+19.3%
Exit	`stop_loss` after 63s, realized -38.9% (-0.0039 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,335 cap, now ~$2,170); never traded again after our exit (no later candles exist); lifetime label: `rugged`, ATH cap ~$4,629
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $POZE on a new_mint trigger with 0.02% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: Loss capped close to policy: -30% configured, -38.9% realized. The mandatory stop did its one job. The token never printed another trade after this exit; the agent was the last one out before the market vanished. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

What should improve: Entry selection, not exit policy, is what loses here: the coin never traded above the entry quote for long.

#17 MELON -2.5% timeout 2026-07-21 07:28

Opened	2026-07-21 07:28 UTC, committed 0.01 SOL
Trigger	new_mint: sniped seconds after the launch appeared on the live pump.fun firehose
Oracle conviction	19/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 30/100 · 3 buyers / 1 sellers observed · organic 45% · top-10 hold 100%
Peak	+0.0%
Exit	`timeout` after 30m, realized -2.5% (-0.0003 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,248 cap, now ~$2,170); never traded again after our exit (no later candles exist); lifetime label: `rugged`, ATH cap ~$2,218
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $MELON on a new_mint trigger with 0.04% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What should improve: The coin flatlined almost immediately; holding the full 30 minutes tied up budget for nothing. A liquidity-decay early exit would free the slot in a fraction of the time.

#18 TrippleBBL +10.4% timeout 2026-07-21 07:58

Opened	2026-07-21 07:58 UTC, committed 0.01 SOL
Trigger	new_mint: sniped seconds after the launch appeared on the live pump.fun firehose
Oracle conviction	19/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 27/100 · 3 buyers / 0 sellers observed · organic 49% · top-10 hold 100%
Peak	+10.4%
Exit	`timeout` after 30m, realized +10.4% (+0.0010 SOL)
After we left	trades at 0.36x our exit today (below where we sold; we sold at an implied ~$6,018 cap, now ~$2,170); highest print after our exit: ~$2,313 cap (0.38x our exit, 1h candles); lifetime label: `rugged`, ATH cap ~$2,523
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $TRIPPLEBBL on a new_mint trigger with 0.04% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: Closed +10.4% in the green. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

#19 SKE -32.5% stop_loss 2026-07-21 08:29

Opened	2026-07-21 08:29 UTC, committed 0.01 SOL
Trigger	new_mint: sniped seconds after the launch appeared on the live pump.fun firehose
Oracle conviction	20/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 42/100 · 15 buyers / 13 sellers observed · organic 64% · top-10 hold 100%
Peak	+0.0%
Exit	`stop_loss` after 29s, realized -32.5% (-0.0033 SOL)
After we left	trades at 0.40x our exit today (below where we sold; we sold at an implied ~$5,462 cap, now ~$2,170); highest print after our exit: ~$2,208 cap (0.40x our exit, 1m candles); lifetime label: `rugged`, ATH cap ~$3,971
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $SKE on a new_mint trigger with 0.02% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: Loss capped close to policy: -30% configured, -32.5% realized. The mandatory stop did its one job. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

What should improve: The Oracle had this coin at conviction 20/100 at entry. The conviction gate that now exists would have skipped it entirely.

#20 test -2.5% timeout 2026-07-21 08:29

Opened	2026-07-21 08:29 UTC, committed 0.01 SOL
Trigger	new_mint: sniped seconds after the launch appeared on the live pump.fun firehose
Oracle conviction	15/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 4/100 · 1 buyers / 0 sellers observed · organic 28% · top-10 hold 100%
Peak	+0.0%
Exit	`timeout` after 30m, realized -2.5% (-0.0002 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,248 cap, now ~$2,170); never traded again after our exit (no later candles exist); lifetime label: `rugged`, ATH cap ~$2,196
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $TEST on a new_mint trigger with 0.04% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What should improve: The coin flatlined almost immediately; holding the full 30 minutes tied up budget for nothing. A liquidity-decay early exit would free the slot in a fraction of the time.

#21 JARRETT -2.5% timeout 2026-07-21 09:00

Opened	2026-07-21 09:00 UTC, committed 0.01 SOL
Trigger	new_mint: sniped seconds after the launch appeared on the live pump.fun firehose
Oracle conviction	17/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 4/100 · 1 buyers / 0 sellers observed · organic 28% · top-10 hold 100%
Peak	+0.0%
Exit	`timeout` after 30m, realized -2.5% (-0.0002 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,248 cap, now ~$2,170); never traded again after our exit (no later candles exist); lifetime label: `rugged`, ATH cap ~$2,191
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $JARRETT on a new_mint trigger with 0.04% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What should improve: The coin flatlined almost immediately; holding the full 30 minutes tied up budget for nothing. A liquidity-decay early exit would free the slot in a fraction of the time.

#46 MrSue -2.5% timeout 2026-07-22 00:00

Opened	2026-07-22 00:00 UTC, committed 0.01 SOL
Trigger	new_mint: sniped seconds after the launch appeared on the live pump.fun firehose
Oracle conviction	19/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 25/100 · 2 buyers / 0 sellers observed · organic 36% · top-10 hold 100%
Peak	+0.0%
Exit	`timeout` after 30m, realized -2.5% (-0.0003 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,248 cap, now ~$2,170); never traded again after our exit (no later candles exist)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $MRSUE on a new_mint trigger with 0.04% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What should improve: The coin flatlined almost immediately; holding the full 30 minutes tied up budget for nothing. A liquidity-decay early exit would free the slot in a fraction of the time.

#48 FLIPPENING -18.7% timeout 2026-07-22 00:30

Opened	2026-07-22 00:30 UTC, committed 0.01 SOL
Trigger	new_mint: sniped seconds after the launch appeared on the live pump.fun firehose
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 0/100 · 6 buyers / 3 sellers observed · organic 33% · bundle risk 67% · top-10 hold 100% · dev sold during observation
Peak	+0.0%
Exit	`timeout` after 30m, realized -18.7% (-0.0019 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,247 cap, now ~$2,170); never traded again after our exit (no later candles exist)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $FLIPPENING on a new_mint trigger with 0.03% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: The token never printed another trade after this exit; the agent was the last one out before the market vanished. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

What should improve: Worst kind of exit: the coin bled to -18.7% slowly enough that no stop fired and the clock, not a decision, ended the trade. This is the gap between the stop-loss and the timeout that a mid-hold health check should cover.

#72 Jrool -2.5% timeout 2026-07-23 00:00

Opened	2026-07-23 00:00 UTC, committed 0.01 SOL
Trigger	new_mint: sniped seconds after the launch appeared on the live pump.fun firehose
Oracle conviction	21/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 24/100 · 2 buyers / 0 sellers observed · organic 36% · top-10 hold 100%
Peak	+0.0%
Exit	`timeout` after 30m, realized -2.5% (-0.0002 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,251 cap, now ~$2,171); never traded again after our exit (no later candles exist)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $JROOL on a new_mint trigger with 0.04% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What should improve: The coin flatlined almost immediately; holding the full 30 minutes tied up budget for nothing. A liquidity-decay early exit would free the slot in a fraction of the time.

#77 Free -2.5% timeout 2026-07-23 00:30

Opened	2026-07-23 00:30 UTC, committed 0.01 SOL
Trigger	new_mint: sniped seconds after the launch appeared on the live pump.fun firehose
Oracle conviction	17/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 4/100 · 1 buyers / 0 sellers observed · organic 28% · top-10 hold 100%
Peak	+0.0%
Exit	`timeout` after 30m, realized -2.5% (-0.0002 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,248 cap, now ~$2,170); never traded again after our exit (no later candles exist)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $FREE on a new_mint trigger with 0.04% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What should improve: The coin flatlined almost immediately; holding the full 30 minutes tied up budget for nothing. A liquidity-decay early exit would free the slot in a fraction of the time.

rules-proven rules

Wallet 3334..Krsc · decision ledger · 5 trades · 1 wins (20%) · net -0.0183 SOL · avg winner +42.4% · avg loser -19.8% · median hold 2m

Policy at entry: 0.05 SOL per trade · no take-profit · stop-loss -30% · trailing stop 25% · 30m max hold · socials required

#2 BITS +42.4% timeout 2026-07-19 02:26

Opened	2026-07-19 02:26 UTC, committed 0.05 SOL
Trigger	new_mint: sniped seconds after the launch appeared on the live pump.fun firehose
Oracle conviction	15/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 4/100 · 1 buyers / 0 sellers observed · organic 28% · top-10 hold 100%
Peak	+46.5%
Exit	`timeout` after 30m, realized +42.4% (+0.0212 SOL)
After we left	trades at 0.08x our exit today (below where we sold; we sold at an implied ~$27,376 cap, now ~$2,206); highest print after our exit: ~$15,257 cap (0.56x our exit, 1m candles); lifetime label: `rugged`, ATH cap ~$15,257
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $BITS on a new_mint trigger with 0.05% price impact; firewall verdict allow (score 100). Committed 0.0500 SOL expecting a profitable exit.”

What it did well: Closed +42.4% in the green. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

#4 looong -26.1% trailing_stop 2026-07-20 22:52

Opened	2026-07-20 22:52 UTC, committed 0.05 SOL
Trigger	new_mint: sniped seconds after the launch appeared on the live pump.fun firehose
Firewall	allow (score 100), sell simulated on-chain before the buy
Peak	+0.0%
Exit	`trailing_stop` after 46s, realized -26.1% (-0.0130 SOL)
After we left	trades at 0.09x our exit today (below where we sold; we sold at an implied ~$24,085 cap, now ~$2,247); highest print after our exit: ~$13,713 cap (0.57x our exit, 1m candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $LOOONG on a new_mint trigger with 0.03% price impact; firewall verdict allow (score 100). Committed 0.0500 SOL expecting a profitable exit.”

What should improve: The trail armed below breakeven: the position peaked at +0.0% but the 25% giveback landed at -26.1%. Arming the trail only after the position is green would turn this exit into a scratch instead of a loss.

#8 Meowpin -14.3% trailing_stop 2026-07-21 01:41

Opened	2026-07-21 01:41 UTC, committed 0.05 SOL
Trigger	new_mint: sniped seconds after the launch appeared on the live pump.fun firehose
Oracle conviction	25/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 75/100 · 35 buyers / 26 sellers observed · organic 78% · top-10 hold 97%
Peak	+28.3%
Exit	`trailing_stop` after 29s, realized -14.3% (-0.0072 SOL)
After we left	trades at 0.11x our exit today (below where we sold; we sold at an implied ~$19,386 cap, now ~$2,174); highest print after our exit: ~$6,809 cap (0.35x our exit, 1h candles); lifetime label: `rugged`, ATH cap ~$12,844
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $MEOWPIN on a new_mint trigger with 0.04% price impact; firewall verdict allow (score 100). Committed 0.0500 SOL expecting a profitable exit.”

What should improve: The trail armed below breakeven: the position peaked at +28.3% but the 25% giveback landed at -14.3%. Arming the trail only after the position is green would turn this exit into a scratch instead of a loss.

#11 Clouseau -13.4% trailing_stop 2026-07-21 03:56

Opened	2026-07-21 03:56 UTC, committed 0.05 SOL
Trigger	new_mint: sniped seconds after the launch appeared on the live pump.fun firehose
Oracle conviction	32/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 76/100 · 68 buyers / 36 sellers observed · organic 79% · top-10 hold 88%
Peak	+24.8%
Exit	`trailing_stop` after 2m, realized -13.4% (-0.0067 SOL)
After we left	trades at 0.10x our exit today (below where we sold; we sold at an implied ~$22,143 cap, now ~$2,184); highest print after our exit: ~$9,861 cap (0.45x our exit, 1h candles); lifetime label: `rugged`, ATH cap ~$14,055
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $CLOUSEAU on a new_mint trigger with 0.04% price impact; firewall verdict allow (score 100). Committed 0.0500 SOL expecting a profitable exit.”

What should improve: The trail armed below breakeven: the position peaked at +24.8% but the 25% giveback landed at -13.4%. Arming the trail only after the position is green would turn this exit into a scratch instead of a loss.

#23 AG -25.3% timeout 2026-07-21 12:54

Opened	2026-07-21 12:54 UTC, committed 0.05 SOL
Trigger	new_mint: sniped seconds after the launch appeared on the live pump.fun firehose
Oracle conviction	28/100 at entry
Firewall	warn (score 84), sell simulated on-chain before the buy
Entry intel	quality 63/100 · 26 buyers / 8 sellers observed · organic 74% · top-10 hold 95%
Peak	+0.0%
Exit	`timeout` after 32m, realized -25.3% (-0.0126 SOL)
After we left	trades at 0.46x our exit today (below where we sold; we sold at an implied ~$25,467 cap, now ~$11,594); highest print after our exit: ~$14,830 cap (0.58x our exit, 1m candles); lifetime label: `pumped`, ATH cap ~$14,830
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $AG on a new_mint trigger with 0.03% price impact; firewall verdict warn (score 84). Committed 0.0500 SOL expecting a profitable exit.”

What should improve: Worst kind of exit: the coin bled to -25.3% slowly enough that no stop fired and the clock, not a decision, ended the trade. This is the gap between the stop-loss and the timeout that a mid-hold health check should cover.

boost-ride rules

Wallet 3yjx..aQwt · decision ledger · 31 trades · 15 wins (48%) · net -0.0413 SOL · avg winner +14.9% · avg loser -39.8% · median hold 3m

Policy at entry: 0.01 SOL per trade · take-profit +20% · stop-loss -15% · trailing stop 8% · 4m max hold

#27 ocFkqj… +5.2% timeout 2026-07-21 19:23

Opened	2026-07-21 19:23 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Oracle conviction	21/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 43/100 · 3 buyers / 0 sellers observed · organic 60% · top-10 hold 100%
Peak	+5.2%
Exit	`timeout` after 4m, realized +5.2% (+0.0005 SOL)
After we left	trades at 0.00x our exit today (below where we sold; we sold at an implied ~$34,281,069 cap, now ~$1,620); highest print after our exit: ~$21,212,439 cap (0.62x our exit, 1h candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $OCFK on a graduation_ride trigger with 0.30% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: Closed +5.2% in the green. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

What should improve: This exit was luck-shaped: the 4-minute clock expired while the position happened to be up. Nothing locked the gain in. A reachable take-profit or a momentum-fade exit converts this from luck into policy.

#30 xjcfSQ… +4.7% timeout 2026-07-21 19:28

Opened	2026-07-21 19:28 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Oracle conviction	21/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 43/100 · 3 buyers / 0 sellers observed · organic 60% · top-10 hold 100%
Peak	+4.7%
Exit	`timeout` after 4m, realized +4.7% (+0.0005 SOL)
After we left	trades at 0.00x our exit today (below where we sold; we sold at an implied ~$40,035,795 cap, now ~$1,498); highest print after our exit: ~$1,503 cap (0.00x our exit, 1h candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $XJCF on a graduation_ride trigger with 0.30% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: Closed +4.7% in the green. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

#31 Ux3UY8… +20.9% take_profit 2026-07-21 19:45

Opened	2026-07-21 19:45 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Oracle conviction	21/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 43/100 · 3 buyers / 0 sellers observed · organic 60% · top-10 hold 100%
Peak	+20.2%
Exit	`take_profit` after 3m, realized +20.9% (+0.0021 SOL)
After we left	trades at 0.00x our exit today (below where we sold; we sold at an implied ~$2,120,289 cap, now ~$1,388); highest print after our exit: ~$1,402 cap (0.00x our exit, 1h candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $UX3U on a graduation_ride trigger with 0.99% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: Hit the configured +20% take-profit and sold the whole position into strength. This is the exit ladder doing exactly what it exists for. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

#32 2UDry5… -84.3% stop_loss 2026-07-21 19:49

Opened	2026-07-21 19:49 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 45/100 · 0 buyers / 0 sellers observed · organic 45% · top-10 hold 0%
Peak	+0.0%
Exit	`stop_loss` after 6s, realized -84.3% (-0.0084 SOL)
After we left	trades at 0.18x our exit today (below where we sold; we sold at an implied ~$10,177 cap, now ~$1,814); highest print after our exit: ~$1,914 cap (0.19x our exit, 1h candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $2UDR on a graduation_ride trigger with 1.25% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: It still sold. Getting any exit through a collapsing bonding curve beats holding a dead token. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

What should improve: The stop was set at -15% but the realized loss was -84.3%: the curve's liquidity collapsed faster than the position poller could quote and sell. A stop-loss is a request, not a guarantee. Smaller size in thin curves and a faster quote cadence are the real defenses.

#37 DBrD6t… -15.8% stop_loss 2026-07-21 20:42

Opened	2026-07-21 20:42 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Oracle conviction	16/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 4/100 · 1 buyers / 0 sellers observed · organic 28% · top-10 hold 100%
Peak	+0.0%
Exit	`stop_loss` after 5s, realized -15.8% (-0.0016 SOL)
After we left	trades at 0.01x our exit today (below where we sold; we sold at an implied ~$137,916 cap, now ~$1,489); highest print after our exit: ~$1,577 cap (0.01x our exit, 1h candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $DBRD on a graduation_ride trigger with 1.19% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: Loss capped close to policy: -15% configured, -15.8% realized. The mandatory stop did its one job. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

What should improve: The Oracle had this coin at conviction 16/100 at entry. The conviction gate that now exists would have skipped it entirely.

#41 HMimvR… -20.1% stop_loss 2026-07-22 00:04

Opened	2026-07-22 00:04 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Oracle conviction	21/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 50/100 · 3 buyers / 0 sellers observed · organic 62% · top-10 hold 100%
Peak	+0.0%
Exit	`stop_loss` after 8s, realized -20.1% (-0.0020 SOL)
After we left	trades at 0.01x our exit today (below where we sold; we sold at an implied ~$251,303 cap, now ~$1,423); highest print after our exit: ~$1,437 cap (0.01x our exit, 1h candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $HMIM on a graduation_ride trigger with 1.14% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: Loss capped close to policy: -15% configured, -20.1% realized. The mandatory stop did its one job. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

What should improve: The Oracle had this coin at conviction 21/100 at entry. The conviction gate that now exists would have skipped it entirely.

#42 Y4HEU5… +11.7% timeout 2026-07-22 00:07

Opened	2026-07-22 00:07 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 43/100 · 3 buyers / 0 sellers observed · organic 60% · top-10 hold 100%
Peak	+11.5%
Exit	`timeout` after 4m, realized +11.7% (+0.0012 SOL)
After we left	trades at 0.00x our exit today (below where we sold; we sold at an implied ~$27,595,866 cap, now ~$1,480); highest print after our exit: ~$1,514 cap (0.00x our exit, 1h candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $Y4HE on a graduation_ride trigger with 0.30% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: Closed +11.7% in the green. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

#47 aQzS8b… +12.0% timeout 2026-07-22 00:46

Opened	2026-07-22 00:46 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 43/100 · 3 buyers / 0 sellers observed · organic 60% · top-10 hold 100%
Peak	+12.0%
Exit	`timeout` after 4m, realized +12.0% (+0.0012 SOL)
After we left	trades at 0.00x our exit today (below where we sold; we sold at an implied ~$3,974,392 cap, now ~$1,382); highest print after our exit: ~$1,955,006 cap (0.49x our exit, 1h candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $AQZS on a graduation_ride trigger with 0.89% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: Closed +12.0% in the green. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

#50 JMBush… +5.0% timeout 2026-07-22 01:54

Opened	2026-07-22 01:54 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 43/100 · 3 buyers / 0 sellers observed · organic 60% · top-10 hold 100%
Peak	+5.0%
Exit	`timeout` after 4m, realized +5.0% (+0.0005 SOL)
After we left	trades at 0.00x our exit today (below where we sold; we sold at an implied ~$30,472,559 cap, now ~$1,389); highest print after our exit: ~$30,818,920 cap (1.01x our exit, 1h candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $JMBU on a graduation_ride trigger with 0.30% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: Closed +5.0% in the green. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

#51 yMDueA… -99.9% stop_loss 2026-07-22 02:01

Opened	2026-07-22 02:01 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Oracle conviction	21/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 44/100 · 3 buyers / 0 sellers observed · organic 60% · top-10 hold 100%
Peak	+4.6%
Exit	`stop_loss` after 3m, realized -99.9% (-0.0100 SOL)
After we left	trades at 0.04x our exit today (below where we sold; we sold at an implied ~$34,770 cap, now ~$1,488); highest print after our exit: ~$1,569 cap (0.05x our exit, 1h candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $YMDU on a graduation_ride trigger with 0.30% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What should improve: The stop was set at -15% but the realized loss was -99.9%: the curve's liquidity collapsed faster than the position poller could quote and sell. A stop-loss is a request, not a guarantee. Smaller size in thin curves and a faster quote cadence are the real defenses.

#53 6dAqWB… -22.0% stop_loss 2026-07-22 02:11

Opened	2026-07-22 02:11 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Oracle conviction	18/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 22/100 · 5 buyers / 0 sellers observed · organic 50% · bundle risk 60% · top-10 hold 100%
Peak	+0.0%
Exit	`stop_loss` after 18s, realized -22.0% (-0.0022 SOL)
After we left	trades at 0.02x our exit today (below where we sold; we sold at an implied ~$83,716 cap, now ~$1,516); highest print after our exit: ~$1,537 cap (0.02x our exit, 1h candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $6DAQ on a graduation_ride trigger with 1.20% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: Loss capped close to policy: -15% configured, -22.0% realized. The mandatory stop did its one job. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

What should improve: The Oracle had this coin at conviction 18/100 at entry. The conviction gate that now exists would have skipped it entirely.

#55 fxppSy… +11.3% timeout 2026-07-22 02:26

Opened	2026-07-22 02:26 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Oracle conviction	21/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 43/100 · 3 buyers / 0 sellers observed · organic 60% · top-10 hold 100%
Peak	+11.3%
Exit	`timeout` after 4m, realized +11.3% (+0.0011 SOL)
After we left	trades at 0.00x our exit today (below where we sold; we sold at an implied ~$49,443,662 cap, now ~$1,733); highest print after our exit: ~$2,458 cap (0.00x our exit, 1h candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $FXPP on a graduation_ride trigger with 0.30% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: Closed +11.3% in the green. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

#56 Awyjqw… -17.4% stop_loss 2026-07-22 02:43

Opened	2026-07-22 02:43 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 50/100 · 3 buyers / 0 sellers observed · organic 62% · top-10 hold 100%
Peak	+0.0%
Exit	`stop_loss` after 7s, realized -17.4% (-0.0017 SOL)
After we left	trades at 0.01x our exit today (below where we sold; we sold at an implied ~$257,383 cap, now ~$1,426); highest print after our exit: ~$1,439 cap (0.01x our exit, 1h candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $AWYJ on a graduation_ride trigger with 1.14% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: Loss capped close to policy: -15% configured, -17.4% realized. The mandatory stop did its one job. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

What should improve: Entry selection, not exit policy, is what loses here: the coin never traded above the entry quote for long.

#57 5V3eh6… -98.7% stop_loss 2026-07-22 03:13

Opened	2026-07-22 03:13 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Oracle conviction	21/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 50/100 · 3 buyers / 0 sellers observed · organic 62% · top-10 hold 100%
Peak	+0.0%
Exit	`stop_loss` after 8s, realized -98.7% (-0.0099 SOL)
After we left	trades at 0.90x our exit today (below where we sold; we sold at an implied ~$4,051 cap, now ~$3,644); highest print after our exit: ~$5,005 cap (1.24x our exit, 1m candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $5V3E on a graduation_ride trigger with 1.14% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What should improve: The stop was set at -15% but the realized loss was -98.7%: the curve's liquidity collapsed faster than the position poller could quote and sell. A stop-loss is a request, not a guarantee. Smaller size in thin curves and a faster quote cadence are the real defenses.

#58 ChRFYk… -93.2% stop_loss 2026-07-22 03:37

Opened	2026-07-22 03:37 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Oracle conviction	22/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 50/100 · 3 buyers / 0 sellers observed · organic 62% · top-10 hold 100%
Peak	+0.0%
Exit	`stop_loss` after 9s, realized -93.2% (-0.0093 SOL)
After we left	trades at 0.56x our exit today (below where we sold; we sold at an implied ~$6,737 cap, now ~$3,771); highest print after our exit: ~$3,795 cap (0.56x our exit, 1h candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $CHRF on a graduation_ride trigger with 1.14% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What should improve: The stop was set at -15% but the realized loss was -93.2%: the curve's liquidity collapsed faster than the position poller could quote and sell. A stop-loss is a request, not a guarantee. Smaller size in thin curves and a faster quote cadence are the real defenses.

#59 qFJmT1… +20.4% take_profit 2026-07-22 03:51

Opened	2026-07-22 03:51 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Oracle conviction	21/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 43/100 · 3 buyers / 0 sellers observed · organic 60% · top-10 hold 100%
Peak	+20.4%
Exit	`take_profit` after 3m, realized +20.4% (+0.0020 SOL)
After we left	trades at 0.00x our exit today (below where we sold; we sold at an implied ~$53,484,541 cap, now ~$1,761); highest print after our exit: ~$33,419,046 cap (0.62x our exit, 1h candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $QFJM on a graduation_ride trigger with 0.30% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

#60 4Wkuhs… -8.7% trailing_stop 2026-07-22 03:59

Opened	2026-07-22 03:59 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Oracle conviction	19/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 22/100 · 5 buyers / 0 sellers observed · organic 50% · bundle risk 60% · top-10 hold 100%
Peak	+0.0%
Exit	`trailing_stop` after 19s, realized -8.7% (-0.0009 SOL)
After we left	trades at 0.02x our exit today (below where we sold; we sold at an implied ~$95,125 cap, now ~$1,511); highest print after our exit: ~$1,520 cap (0.02x our exit, 1h candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $4WKU on a graduation_ride trigger with 1.20% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What should improve: The trail armed below breakeven: the position peaked at +0.0% but the 8% giveback landed at -8.7%. Arming the trail only after the position is green would turn this exit into a scratch instead of a loss.

#61 u8oLcb… +20.2% take_profit 2026-07-22 04:03

Opened	2026-07-22 04:03 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 43/100 · 3 buyers / 0 sellers observed · organic 60% · top-10 hold 100%
Peak	+20.2%
Exit	`take_profit` after 4m, realized +20.2% (+0.0020 SOL)
After we left	trades at 0.00x our exit today (below where we sold; we sold at an implied ~$48,359,811 cap, now ~$1,842); highest print after our exit: ~$1,841 cap (0.00x our exit, 1h candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $U8OL on a graduation_ride trigger with 0.30% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

#71 Aevon +2.0% timeout 2026-07-23 00:13

Opened	2026-07-23 00:13 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Firewall	allow (score 100), sell simulated on-chain before the buy
Peak	+4.8%
Exit	`timeout` after 4m, realized +2.0% (+0.0002 SOL)
After we left	trades at 0.03x our exit today (below where we sold; we sold at an implied ~$80,495 cap, now ~$2,634); never traded again after our exit (no later candles exist)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $AEVON on a graduation_ride trigger with 1.25% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: Closed +2.0% in the green. The token never printed another trade after this exit; the agent was the last one out before the market vanished. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

#73 Walter -2.3% trailing_stop 2026-07-23 00:35

Opened	2026-07-23 00:35 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Firewall	allow (score 100), sell simulated on-chain before the buy
Peak	+0.0%
Exit	`trailing_stop` after 15s, realized -2.3% (-0.0002 SOL)
After we left	trades at 0.08x our exit today (below where we sold; we sold at an implied ~$74,229 cap, now ~$6,027); highest print after our exit: ~$31,026 cap (0.42x our exit, 1m candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $WALTER on a graduation_ride trigger with 1.25% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What should improve: The trail armed below breakeven: the position peaked at +0.0% but the 8% giveback landed at -2.3%. Arming the trail only after the position is green would turn this exit into a scratch instead of a loss.

#75 VFX +18.7% take_profit 2026-07-23 00:37

Opened	2026-07-23 00:37 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Firewall	allow (score 100), sell simulated on-chain before the buy
Peak	+20.7%
Exit	`take_profit` after 3m, realized +18.7% (+0.0019 SOL)
After we left	trades at 0.05x our exit today (below where we sold; we sold at an implied ~$99,353 cap, now ~$5,183); highest print after our exit: ~$46,155 cap (0.46x our exit, 1m candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $VFX on a graduation_ride trigger with 1.20% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

#76 FLL +20.0% take_profit 2026-07-23 00:45

Opened	2026-07-23 00:45 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Firewall	allow (score 100), sell simulated on-chain before the buy
Peak	+20.0%
Exit	`take_profit` after 3m, realized +20.0% (+0.0020 SOL)
After we left	trades at 0.03x our exit today (below where we sold; we sold at an implied ~$99,101 cap, now ~$3,467); highest print after our exit: ~$46,459 cap (0.47x our exit, 1m candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $FLL on a graduation_ride trigger with 1.20% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

#80 RWA -35.6% stop_loss 2026-07-23 01:32

Opened	2026-07-23 01:32 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Firewall	allow (score 100), sell simulated on-chain before the buy
Peak	+10.6%
Exit	`stop_loss` after 2m, realized -35.6% (-0.0036 SOL)
After we left	trades at 0.04x our exit today (below where we sold; we sold at an implied ~$78,671 cap, now ~$3,108); never traded again after our exit (no later candles exist)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $RWA on a graduation_ride trigger with 1.20% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What should improve: The stop was set at -15% but the realized loss was -35.6%: the curve's liquidity collapsed faster than the position poller could quote and sell. A stop-loss is a request, not a guarantee. Smaller size in thin curves and a faster quote cadence are the real defenses.

#83 CWARDIN -32.8% stop_loss 2026-07-23 02:04

Opened	2026-07-23 02:04 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Firewall	allow (score 100), sell simulated on-chain before the buy
Peak	+16.8%
Exit	`stop_loss` after 3m, realized -32.8% (-0.0033 SOL)
After we left	trades at 0.04x our exit today (below where we sold; we sold at an implied ~$82,366 cap, now ~$3,038); never traded again after our exit (no later candles exist)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $CWARDIN on a graduation_ride trigger with 1.20% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What should improve: The stop was set at -15% but the realized loss was -32.8%: the curve's liquidity collapsed faster than the position poller could quote and sell. A stop-loss is a request, not a guarantee. Smaller size in thin curves and a faster quote cadence are the real defenses.

#84 BENNER -34.3% stop_loss 2026-07-23 02:10

Opened	2026-07-23 02:10 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Firewall	allow (score 100), sell simulated on-chain before the buy
Peak	+15.9%
Exit	`stop_loss` after 3m, realized -34.3% (-0.0034 SOL)
After we left	trades at 0.04x our exit today (below where we sold; we sold at an implied ~$82,120 cap, now ~$3,038); never traded again after our exit (no later candles exist)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $BENNER on a graduation_ride trigger with 1.20% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What should improve: The stop was set at -15% but the realized loss was -34.3%: the curve's liquidity collapsed faster than the position poller could quote and sell. A stop-loss is a request, not a guarantee. Smaller size in thin curves and a faster quote cadence are the real defenses.

#85 AGIGUY -34.4% stop_loss 2026-07-23 02:17

Opened	2026-07-23 02:17 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Firewall	allow (score 100), sell simulated on-chain before the buy
Peak	+15.9%
Exit	`stop_loss` after 3m, realized -34.4% (-0.0034 SOL)
After we left	trades at 0.04x our exit today (below where we sold; we sold at an implied ~$83,683 cap, now ~$3,029); never traded again after our exit (no later candles exist)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $AGIGUY on a graduation_ride trigger with 1.19% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What should improve: The stop was set at -15% but the realized loss was -34.4%: the curve's liquidity collapsed faster than the position poller could quote and sell. A stop-loss is a request, not a guarantee. Smaller size in thin curves and a faster quote cadence are the real defenses.

#86 DrqyFF… +7.7% timeout 2026-07-23 02:21

Opened	2026-07-23 02:21 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 43/100 · 3 buyers / 0 sellers observed · organic 60% · top-10 hold 100%
Peak	+7.7%
Exit	`timeout` after 4m, realized +7.7% (+0.0008 SOL)
After we left	trades at 0.00x our exit today (below where we sold; we sold at an implied ~$47,830,590 cap, now ~$1,722); highest print after our exit: ~$51,051,658 cap (1.07x our exit, 1h candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $DRQY on a graduation_ride trigger with 0.30% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What it did well: Closed +7.7% in the green. The trade firewall simulated the full buy-and-sell round trip on-chain before spending: the token was provably sellable at entry.

#87 moHcoC… -6.4% trailing_stop 2026-07-23 02:26

Opened	2026-07-23 02:26 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Oracle conviction	21/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 50/100 · 3 buyers / 0 sellers observed · organic 62% · top-10 hold 100%
Peak	+1.9%
Exit	`trailing_stop` after 2m, realized -6.4% (-0.0006 SOL)
After we left	trades at 0.01x our exit today (below where we sold; we sold at an implied ~$230,800 cap, now ~$1,459); never traded again after our exit (no later candles exist)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $MOHC on a graduation_ride trigger with 1.19% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What should improve: The trail armed below breakeven: the position peaked at +1.9% but the 8% giveback landed at -6.4%. Arming the trail only after the position is green would turn this exit into a scratch instead of a loss.

#88 HANU.p +42.9% take_profit 2026-07-23 02:41

Opened	2026-07-23 02:41 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Firewall	allow (score 100), sell simulated on-chain before the buy
Peak	+20.3%
Exit	`take_profit` after 5m, realized +42.9% (+0.0043 SOL)
After we left	trades at 0.04x our exit today (below where we sold; we sold at an implied ~$118,030 cap, now ~$5,081); highest print after our exit: ~$5,242 cap (0.04x our exit, 1m candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $HANU.P on a graduation_ride trigger with 0.95% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

#89 $MRH -30.1% stop_loss 2026-07-23 02:46

Opened	2026-07-23 02:46 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Firewall	allow (score 100), sell simulated on-chain before the buy
Peak	+10.1%
Exit	`stop_loss` after 4m, realized -30.1% (-0.0030 SOL)
After we left	trades at 0.03x our exit today (below where we sold; we sold at an implied ~$90,094 cap, now ~$3,012); highest print after our exit: ~$39,186 cap (0.43x our exit, 1m candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $$MRH on a graduation_ride trigger with 1.19% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

What should improve: The stop was set at -15% but the realized loss was -30.1%: the curve's liquidity collapsed faster than the position poller could quote and sell. A stop-loss is a request, not a guarantee. Smaller size in thin curves and a faster quote cadence are the real defenses.

#90 BDAG +20.4% take_profit 2026-07-23 02:53

Opened	2026-07-23 02:53 UTC, committed 0.01 SOL
Trigger	graduation_ride: bought at bonding-curve graduation to ride the Raydium migration flow
Firewall	allow (score 100), sell simulated on-chain before the buy
Peak	+20.6%
Exit	`take_profit` after 2m, realized +20.4% (+0.0020 SOL)
After we left	trades at 0.06x our exit today (below where we sold; we sold at an implied ~$95,584 cap, now ~$5,302); highest print after our exit: ~$46,016 cap (0.48x our exit, 1m candles)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $BDAG on a graduation_ride trigger with 1.25% price impact; firewall verdict allow (score 100). Committed 0.0100 SOL expecting a profitable exit.”

intel-quality rules

Wallet D5BG..Xypu · decision ledger · 6 trades · 0 wins (0%) · net -0.0413 SOL · avg winner +0.0% · avg loser -4.6% · median hold 30m

Policy at entry: 0.15 SOL per trade · take-profit +200% · stop-loss -35% · trailing stop 30% · 30m max hold · Oracle conviction ≥ 55 · socials required

#3 Enthuist -8.0% timeout 2026-07-18 00:21

Opened	2026-07-18 00:21 UTC, committed 0.15 SOL
Trigger	intel_confirmed: the intel engine watched the first minutes of trading and confirmed real two-sided demand
Firewall	warn (score 84), sell simulated on-chain before the buy
Peak	+0.0%
Exit	`timeout` after 37.5h, realized -8.0% (-0.0120 SOL)
After we left	trades at 0.42x our exit today (below where we sold; we sold at an implied ~$5,221 cap, now ~$2,169); never traded again after our exit (no later candles exist)
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $ENTHUIST on a intel_confirmed trigger with 0.51% price impact; firewall verdict warn (score 84). Committed 0.1500 SOL expecting a profitable exit.”

What it did well: Scratched out near flat (-8.0%) instead of riding a dead coin further down. The token never printed another trade after this exit; the agent was the last one out before the market vanished.

What should improve: The coin flatlined almost immediately; holding the full 30 minutes tied up budget for nothing. A liquidity-decay early exit would free the slot in a fraction of the time.

#5 Smoke -7.1% timeout 2026-07-21 00:00

Opened	2026-07-21 00:00 UTC, committed 0.15 SOL
Trigger	intel_confirmed: the intel engine watched the first minutes of trading and confirmed real two-sided demand
Oracle conviction	22/100 at entry
Firewall	warn (score 84), sell simulated on-chain before the buy
Entry intel	quality 30/100 · 4 buyers / 2 sellers observed · organic 49% · top-10 hold 100%
Peak	+0.0%
Exit	`timeout` after 30m, realized -7.1% (-0.0107 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,270 cap, now ~$2,170); highest print after our exit: ~$2,186 cap (0.41x our exit, 1m candles); lifetime label: `rugged`, ATH cap ~$3,566
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $SMOKE on a intel_confirmed trigger with 0.51% price impact; firewall verdict warn (score 84). Committed 0.1500 SOL expecting a profitable exit.”

What it did well: Scratched out near flat (-7.1%) instead of riding a dead coin further down.

What should improve: The coin flatlined almost immediately; holding the full 30 minutes tied up budget for nothing. A liquidity-decay early exit would free the slot in a fraction of the time.

#6 WET -2.5% timeout 2026-07-21 00:30

Opened	2026-07-21 00:30 UTC, committed 0.15 SOL
Trigger	intel_confirmed: the intel engine watched the first minutes of trading and confirmed real two-sided demand
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 45/100 · 0 buyers / 0 sellers observed · organic 45% · top-10 hold 0%
Peak	+0.0%
Exit	`timeout` after 30m, realized -2.5% (-0.0037 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,272 cap, now ~$2,170); never traded again after our exit (no later candles exist); lifetime label: `rugged`, ATH cap ~$2,200
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $WET on a intel_confirmed trigger with 0.54% price impact; firewall verdict allow (score 100). Committed 0.1500 SOL expecting a profitable exit.”

What should improve: The coin flatlined almost immediately; holding the full 30 minutes tied up budget for nothing. A liquidity-decay early exit would free the slot in a fraction of the time.

#7 YES YOU -2.5% timeout 2026-07-21 01:02

Opened	2026-07-21 01:02 UTC, committed 0.15 SOL
Trigger	intel_confirmed: the intel engine watched the first minutes of trading and confirmed real two-sided demand
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 45/100 · 0 buyers / 0 sellers observed · organic 45% · top-10 hold 0%
Peak	+0.0%
Exit	`timeout` after 30m, realized -2.5% (-0.0037 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,272 cap, now ~$2,170); never traded again after our exit (no later candles exist); lifetime label: `rugged`, ATH cap ~$2,204
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $YES YOU on a intel_confirmed trigger with 0.54% price impact; firewall verdict allow (score 100). Committed 0.1500 SOL expecting a profitable exit.”

What should improve: The coin flatlined almost immediately; holding the full 30 minutes tied up budget for nothing. A liquidity-decay early exit would free the slot in a fraction of the time.

#9 KYDEN -5.0% timeout 2026-07-21 01:34

Opened	2026-07-21 01:34 UTC, committed 0.15 SOL
Trigger	intel_confirmed: the intel engine watched the first minutes of trading and confirmed real two-sided demand
Oracle conviction	17/100 at entry
Firewall	warn (score 84), sell simulated on-chain before the buy
Entry intel	quality 4/100 · 1 buyers / 0 sellers observed · organic 28% · top-10 hold 100%
Peak	+0.0%
Exit	`timeout` after 30m, realized -5.0% (-0.0075 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,271 cap, now ~$2,170); highest print after our exit: ~$2,433 cap (0.46x our exit, 1m candles); lifetime label: `rugged`, ATH cap ~$2,709
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $KYDEN on a intel_confirmed trigger with 0.52% price impact; firewall verdict warn (score 84). Committed 0.1500 SOL expecting a profitable exit.”

What it did well: Scratched out near flat (-5.0%) instead of riding a dead coin further down.

What should improve: The coin flatlined almost immediately; holding the full 30 minutes tied up budget for nothing. A liquidity-decay early exit would free the slot in a fraction of the time.

#10 SADCAT -2.5% timeout 2026-07-21 02:05

Opened	2026-07-21 02:05 UTC, committed 0.15 SOL
Trigger	intel_confirmed: the intel engine watched the first minutes of trading and confirmed real two-sided demand
Oracle conviction	35/100 at entry
Firewall	allow (score 100), sell simulated on-chain before the buy
Entry intel	quality 45/100 · 0 buyers / 0 sellers observed · organic 45% · top-10 hold 0%
Peak	+0.0%
Exit	`timeout` after 30m, realized -2.5% (-0.0037 SOL)
After we left	trades at 0.41x our exit today (below where we sold; we sold at an implied ~$5,272 cap, now ~$2,169); never traded again after our exit (no later candles exist); lifetime label: `rugged`, ATH cap ~$2,200
Receipts	buy tx · sell tx · wallet

Ledger rationale, verbatim: “Sniped $SADCAT on a intel_confirmed trigger with 0.54% price impact; firewall verdict allow (score 100). Committed 0.1500 SOL expecting a profitable exit.”

What should improve: The coin flatlined almost immediately; holding the full 30 minutes tied up budget for nothing. A liquidity-decay early exit would free the slot in a fraction of the time.

21. Everything we learned

Model judgment beat frozen rules on expectancy, decisively, in this window. +0.0147 SOL across 33 LLM-judged trades vs -0.1182 SOL across 57 rules-driven trades, with the LLM cohort owning six of the seven biggest wins. The mechanism looks like selectivity: the model can decline launches a threshold cannot describe.
And 33 trades is still just a signal. A single +73.9% winner moves the whole cohort. We are scaling the sample, not declaring victory. If the effect is real it will survive more data; if it is not, we will publish that too.
Win rate and profitability are almost unrelated in this regime. The highest-win-rate arm lost money; the second-most-profitable arm wins a quarter of the time. Expectancy (win rate times average win, minus loss rate times average loss) is the only number that matters.
The take-profit is where the money is. The only net-positive exit category. Everything else is damage control. Configure it, and ladder it once the position size justifies it.
Stop-losses cap the common case, not the tail. Five positions blew through their stops by 50 to 85 points on liquidity collapse. Tail defense happens at entry (concentration screening) and sizing, not at exit.
Trailing stops must arm above breakeven. Armed underwater, they are a machine for realizing small losses. The data shows it clearly across 12 trailing exits.
The Oracle's conviction gradient is real and was mostly ignored. Sub-30-conviction entries funded most of the losses. Not one trade entered above 50. The gate now exists on every arm that opted in, and the empty top band is the next thing to fix: high-conviction coins exist, the arms just were not positioned to catch them.
At scale, the Oracle's top band is a 7x-the-base-rate signal that nobody traded. Across 42,995 labeled coins, conviction 50+ hit 77% pumped-or-graduated vs 12% for the sub-30 band, and the fleet bought zero of the 110 qualifying coins because its two Oracle gates (35 and 65) bracketed the profitable band (50 to 61) without covering it. Audit your thresholds against the live score distribution, not against intuition.
The fleet has no paperhands problem, verified against every token's full price history. 37 of 90 tokens never traded again after our exit; of the rest, zero printed 1.5x our exit afterward (median post-exit high: 0.43x) and none trades above our exit today. Every sell was vindicated. The losses came from entries, and the recoverable upside sits inside the hold (laddered exits), not beyond it.
The biggest winners were invisible at minute zero. The skipped coins that ran to eight and nine figures showed 3 visible buyers and mediocre quality during our observation window; their runs started after it closed. Minute-zero sniping structurally cannot catch them; a delayed second-look entry can.
The judges' skips prove the edge is selection. Across 3,850 labeled judged coins, buy verdicts went good 53.3% vs 11.7% for skips, on coins the fleet mostly never traded. That is a 4.5x lift measured with zero execution effects in it.
Model overconfidence is real and measurable. Judge verdicts at 0.7 confidence were the best band; 0.9-plus verdicts went winless. Confidence floors exist in the config; this argues for ceilings too.
Execution was never the problem. Median 3 seconds from token creation to our buy landing, median 0.04% entry impact, zero entries over 2% impact, fees a rounding error. Every SOL lost, was lost to what we chose, not how we filled it.
Autonomy without audited side effects is theater. Two closed loops looked alive for two days while doing nothing, behind green health checks. Count rows, not status codes.
The ledger pays for itself. This entire article, including every thesis quote and every grade, was generated from the decision ledger and the position table. An autonomous system that cannot produce its own honest postmortem is not finished.

22. What happens next

The experiment continues with the same public standard: real money, real receipts, published either way. Concretely, in flight right now:

Rebalance toward what works. The evolution layer reallocates budget across arms using confidence intervals on realized returns, so the LLM arms earn larger budgets by performing, not by us liking them.
Fix the model routing, then widen the model roster. The named-model arms should actually be piloted by their named models; the failover chain answering most calls muddied the model-vs-model comparison this round. Once the routing holds, the next trials add more pilots to the cockpit: new arms for additional frontier models, starting with Moonshot's Kimi K3, running head to head against the incumbents at equal budgets under identical rails. The LLM cohort's edge (section 8) is currently a "model judgment vs rules" result; a clean multi-model bracket is how it becomes a "which model trades best" result, and each arm's theses and trades will publish on its own ledger like everything else here.
An arm that trades the conviction-50 crossing: built. Section 11's 77%-good band is the strongest signal in this dataset, so the sniper now has an oracle_crossing trigger: a watcher polls the Oracle and buys a coin the first time it crosses the strategy's bar, instead of gating a launch-second buy on a score that does not exist yet. The formerly-idle strict arm is re-armed on it: bar at 50 (its old 65 was above the highest score ever observed), tiny size, and, because section 12's data says the median crossing coin eventually decays to 0.32x, a sell-the-initials ladder at 1.5x with a protected moonbag. It goes live with the next worker deploy and its trades will publish like everything else.
Trailing stops now arm only above breakeven: shipped. The exit ladder no longer fires a trailing stop on a position that was never green; the hard stop-loss keeps sole custody of the downside. Every below-breakeven trail in this dataset realized a loss that the stop would have capped anyway.
A second-window scanner. The crossing watcher is the first instance of the pattern; the general version re-scores surviving coins at 30 and 60 minutes so the fleet can catch the winners whose stories start after minute zero (section 12), instead of judging every coin exactly once at birth.
Arm trailing stops above breakeven only, and add a liquidity-decay early exit so dead coins stop squatting on concurrency slots for half an hour.
Keep the optimizer honest. Every applied tuning now writes an audit row and a ledger entry, and the fleet's tunings will appear in future editions of this dataset.

Everything above runs on the same platform any three.ws user can touch: the agents, the wallets, the sniper, the ledgers. If you want to arm your own agent with your own rules (or your own model), the tutorial is below. Size it like an experiment, not like a retirement plan.

Arm your own agent →

23. Methods, data, and glossary

Data sources

what	source	notes
trades, fills, exits	`agent_sniper_positions` (production ledger)	real fills only; the SIMULATED sentinel excluded
entry rationale, theses, confidence	`agent_decisions` hash-chained reasoning ledger	public per agent at `/ledger/<agent>`
judge verdicts, buys and skips	`sniper_llm_verdicts`	one verdict per (coin, model); answered_by records fallbacks
launch observation, entry signals	`pump_coin_intel` (intel engine)	82,994 coins in window
ground-truth outcomes	`pump_coin_outcomes`	graduated / pumped / flat / rugged + ATH cap
conviction scores + history	`oracle_conviction`, `oracle_conviction_history`	history retained 72h; captured at study close
candle price histories	pump.fun public swap API	USD per token; price times 1B fixed supply equals market cap
live coin state post-exit	pump.fun frontend API + DexScreener	keyless public endpoints
self-tuning audit trail	`agent_sniper_optimizer_runs`, `sniper_evolution_log`	every applied change journaled

Key methodological choices and limitations

Snapshot cutoff. Trades frozen at closes before 2026-07-23 04:00 UTC (the fleet keeps trading; live totals will drift from this article). Counterfactual window: 2026-07-20 00:00 to 2026-07-23 04:00 UTC.
Currency conversions use SOL at ~$190 (spot at time of writing). All PnL is computed natively in SOL/lamports and never round-trips through USD; dollars appear only in market-cap framings, where a few percent of SOL drift does not change any conclusion.
Implied market caps at our fills derive from our own fill prices times pump.fun's fixed 1B token supply.
Outcome labels straddle a regime change (pump.fun's BOOST buyback mechanic, mid-July); pre-BOOST baselines are not directly comparable.
Counterfactuals are bounds, not backtests. Where we report what skipped cohorts went on to do, we did not model entry slippage, fill competition, or our own market impact, which is also why the interactive replay (section 19) fixes entries to real fills and only varies exits.
Small-n candor. The LLM cohort is 33 trades; judge calibration bands hold dozens of coins, not thousands; the signal-separation table (section 16) is hypothesis-generating. Where n is small we say so in place.
Survivorship handling. The paperhands audit counts never-traded-again coins explicitly (37 of 90) instead of dropping them.

Glossary

Bonding curve: pump.fun's automated market: price rises deterministically with tokens bought; all early trading happens against it rather than an order book.

Graduation: a coin completing its bonding curve (~$69k market cap) and migrating to a full DEX listing; the strongest "it made it" signal in this market.

Rugged: the coin's value was pulled out from under holders (creator dump, coordinated exit); operationally, a collapse it never recovers from.

Mayhem mode: a pump.fun launch variant with adversarial mechanics; the fleet excludes it unconditionally, read from the chain.

BOOST window: pump.fun's five-minute post-graduation buyback; the boost-ride arm trades only this window.

Oracle conviction: our 0-100 fused score per coin (creator pedigree, market structure, narrative, momentum), recomputed as the coin trades.

Trade firewall: a real on-chain simulation of the full buy-and-sell round trip before spending; if the token cannot be sold back, the buy aborts.

Arm: one strategy in the A/B fleet: its own agent, wallet, budget, and entry policy on shared execution rails.

Expectancy: win rate times average win, minus loss rate times average loss; the number that decides profitability, as opposed to win rate, which does not.

Data availability

The full trade dataset (autonomous-trading-experiment.json) and the backtest price series (autonomous-trading-experiment-series.json) are published alongside this article. The execution stack is in the platform repository (see workers/agent-sniper/), and every agent's decision ledger and wallet is publicly auditable via the links in section 6.

Go deeper

The dataset behind every number in this post: autonomous-trading-experiment.json
The first trade and the five-gate walkthrough: the first-trade postmortem
The full engineering breakdown: the Agent Sniper doc
How to arm one yourself: the tutorial

← All posts