Lab · RatLab

The lab tests trading ideas on paper before any may touch Ratt Damon's real wallet. None has earned it yet: 0 of 13 swarm ideas proven, and a random exit still beats every exit rule. Everything here is paper or simulated unless marked real.

The first two months How the lab works Real money: the Desk

The first two months

Mixed17 Sep → 17 Nov · so far, day 5703:34:50 UTC

A bot whose coin pick comes from a simulated brain has traded a real wallet in public since 17 Sep. What grew: the lab's data, its testing method and the safety rules. What did not: the money (−⁠109.1% against what was put in), and no trading idea has been proven so far.

Week by weekreal money:

Week by week, bot ledger
Week of	Buys	Closed	Won	Realized	Running total	Typical bet	Paused
17 Sep	49	46	33	+$14.66	+$14.66	$15	0 h
24 Sep	128	128	80	+$36.26	+$50.91	$15	0 h
1 Oct	165	168	59	−$65.51	−$14.59	$20	3 d 15 h
8 Oct	87	87	46	−$8.89	−$23.49	$15	4 d 14 h
15 Oct	0	0	–	–	–	–	7 d
22 Oct	0	0	–	–	–	–	7 d
29 Oct	52	51	23	−$21.54	−$45.03	$10	5 d 8 h
5 Nov	4	5	4	+$0.90	−$44.12	$10	6 d 19 h
12 Nov (so far)	0	0	–	–	–	–	4 h

Measured 3 NovBot ledger: sale minus bet, per closed trade. Until 7 Oct it booked quoted prices, not landed fills, and it leaves out fees, token-account rent and SOL's price, so it overstated September (30 Sep: ledger +⁠$42.31, wallet +⁠$13.71 against what was put in). Over the whole period the two still differ; the wallet line is the ground truth.

Every stop a person had to liftreal money:

9 stops; 8 lifted by a person, one still in force. Buying was paused 34.5 of 55.5 days: the breaker 23.0, the rolling stop 11.6.

Every stop a person had to lift
Last buy	Lifted	Rule	Wallet when lifted	The bot's reason
3 Oct 08:47 UTC	6 Oct 15:22 UTC (3 d 6 h after the last buy)	Circuit breaker	$125.69	circuit breaker: the real book fell under $120
7 Oct 15:19 UTC	8 Oct 14:57 UTC (23 h 38 min after the last buy)	Rolling stop	$106.30	rolling stop: the real wallet lost $6.85 over its last 14 days (20 trades) — only sells until a person clears "stopped"
9 Oct 05:50 UTC	9 Oct 13:10 UTC (7 h 20 min after the last buy)	Rolling stop	$105.52	rolling stop: the real wallet lost $0.04 over its last 14 days (21 trades) — only sells until a person clears "stopped"
10 Oct 06:41 UTC	13 Oct 00:36 UTC (2 d 17 h after the last buy)	Rolling stop	$97.22	rolling stop: the real wallet lost $2.95 over its last 14 days (21 trades) — only sells until a person clears "stopped"
13 Oct 18:02 UTC	14 Oct 03:01 UTC (8 h 58 min after the last buy)	Rolling stop	$99.00	rolling stop: the real wallet lost $1.34 over its last 14 days (27 trades) — only sells until a person clears "stopped"
14 Oct 11:39 UTC	2 Nov 21:38 UTC (19 d 9 h after the last buy)	Circuit breaker	$79.92	circuit breaker: the real wallet is worth $83.57, under $85
3 Nov 12:23 UTC	3 Nov 16:12 UTC (3 h 48 min after the last buy)	Rolling stop	$81.20	rolling stop: the real wallet lost $1.65 over its last 14 days (20 trades) — only sells until a person clears "stopped"
4 Nov 02:13 UTC	4 Nov 07:58 UTC (5 h 44 min after the last buy)	Circuit breaker	$69.50	circuit breaker: the real wallet is worth $68.12, under $70
5 Nov 04:12 UTC	still in force	Rolling stop	–	rolling stop: the real wallet lost $9.70 over its last 14 days (20 trades) — only sells until a person clears "stopped"

Gaps in buying that no person had to lift (such as 18–21 Sep) are not stops. They are on the Timeline.

What changed, and why

What changed, and why: 13 changes, oldest first
When	What happened	What we changed	Since then	Kind
14 Sep	The first ~1,000 paper trades: chasing a pump lost 8.2% a trade, pools under $50k lost 14%, curve coins 11–16%.	Entry filters for every rat: no chasing, no thin pools, no curve coins.	Applies to every real buy: 485 so far.	paper, real money:
18 Sep	Shadow book: 9 of 11 coins under $40k fell 80% within 30 minutes.	Adding to a coin already held is allowed, with a cap on how much one coin can take.	Not measured separately.	paper, real money: Mixed effect.
30 Sep	The owner's rule: judge against the random rat; if the last 14 days are negative, cut.	Rolling stop: buying stops once 20+ recent real trades are net negative, until a person clears it.	Fired 6 times, pausing buying 11.6 days; the breaker (3) paused it 23.0. In all, buying was paused 34.5 of 55.5 days.	real money:
30 Sep	On the trades so far, high-conviction buys (z ≥ 1.5) had won 78%.	High-conviction buys got double bets ($30–$40), 30 Sep – 3 Oct.	On trades opened after 30 Sep 22:39 UTC: 37.5% won over 88, −⁠2.21% a trade on average (−⁠1.08% without GOIF). Hindsight: the doubling cost −⁠$9.81 net: −⁠$19.19 on GOIF, +⁠$9.38 on the other 32 doubled bets.	real money, simulated, hindsight: Raised risk.
30 Sep	The owner wanted more real trades: "one trade every 10 minutes minimum, real, with the filters" (30 Sep).	Minimum pace: when the wallet had not bought for a while, the top pick was bought even under the conviction bar.	22 real buys under the bar, 30 Sep 23:10 UTC – 1 Oct 10:02 UTC: −⁠$2.21, 10 won; none since.	real money, simulated: Raised risk.
30 Sep	The owner's aim (30 Sep): "the real one learns now", so real closes, not only paper ones, reach the brain.	Every real close is sent to the simulated brain as a lesson.	312 real closes sent. Each trace shrinks by about two-thirds every 8 brain-minutes and is nearly gone within about 40 (measured 3 Nov).	simulated, real money, measured 3 nov:
1 Oct	September's bot ledger read +$50.91 (quoted prices).	Base bet raised from $15 to $20 (1 Oct 00:49); back to $15 on 6 Oct, up to $20 again on 9 Oct 03:03.	Base bet by run: $6 (17 Sep) → $15 (21 Sep) → $20 (1 Oct) → $15 (6 Oct) → $20 (9 Oct) → $10 (13 Oct). The week from 1 Oct: −⁠$65.51, 35% won.	real money: Raised risk.
2 Oct	The owner wanted the wallet to copy the simulation exactly after the lowered trail fired fifteen winners in a row.	The hold floor (in place since 18 Sep) was switched off; it came back on 7 Oct.	Off for 4.5 days, until 7 Oct.	real money: Raised risk.
6 Oct	GOIF, 3 Oct: a $38.66 bet went to zero; its pool held 0.23% of the supply.	Holder check before every buy (300+ holders, top 10 hold ≤ 50%; no data means no buy) and a breaker on the true wallet value.	Worst trade since: −⁠100.0% over 184 trades; worse than −50%: 2 (before: 1 of 301). Too rare to prove. The average trade since: −⁠1.79% (before, without GOIF: +⁠0.65%).	real money:
7 Oct	BINDER: holds of 8–30 s lost 5.8%; the fill landed 2.9% under the quote.	Book the landed fill, not the quote; 1% slippage (3% when urgent); the hold floor back (20 minutes); confirm a stop before selling.	Holds under a minute: 2 while the floor was off (2 Oct to the fix), 1 since.	real money:
13 Oct	Losses after the relaunches of 6, 8 and 9 Oct.	Bet size: $15 → $20 (1 Oct) → $15 (6 Oct) → $20 (9 Oct) → $10 (13 Oct).	$10 a bet since 13 Oct (earlier $6–$40); worst trade at that size: −⁠$10.00 (99 trades).	real money:
14 Oct	Swarm, exit lab, tournament and slow lab tested ideas against random controls.	Nothing promoted to the real wallet: no idea met its own bar.	Nothing promoted. Swarm: 8 rejected, 5 waiting. Exit rules: none ahead of a random exit on pooled trades. Tournament: no candidate clears the control at the 14-trial threshold (z 2.68).	paper:
3 Nov	Coins born on pump.fun lost money; the others made a little. Found after looking at the data.	No filter added. New trades are the check.	Since 3 Nov 01:06 UTC: pump.fun 31 trades −⁠$19.15; others 21 trades −⁠$1.08. Far too few to decide. Over the whole period, −⁠$38.66 of the pump.fun loss is GOIF alone.	real money, hindsight:

The brain, audited 3 NovSimulated

simulated, real money: It picks the coin. Among the coins the safety rules allow, he buys the one his simulated striatum scores highest, once it clears a bar: 485 of 485 closed real trades trace back to that pick.
simulated, real money, measured 3 nov: The pick is no better than a coin flip. Its score moves like random noise (measured 3 Nov: spread 0.70 against 0.72 for pure noise), and on real trades it does not predict the result: correlation +0.01 over 485 trades.
simulated, measured 3 nov: It never sells, but its score can hold a buy back. A time limit, a trailing stop, a stop loss, a take-profit or the callers leaving sells; brain exits are switched off. Nothing is bought until the top allowed coin clears a bar (z 0.9), except the 22 buys the minimum-pace rule forced on 30 Sep – 1 Oct. Since the score behaves like noise, the bar changes when he buys, not what. Its only other veto is on a strategy the real wallet does not use.
simulated, real money, measured 3 nov: Lessons fade. 312 real closes were sent to it as lessons since 30 Sep; each trace shrinks by about two-thirds every 8 brain-minutes and is nearly gone within about 40, and its plastic wiring is within ±0.3% of how it was built.
simulated: It never grew and was never reset. 55,583 neurons since 14 Sep, one run since 16 Sep: 45.6 simulated brain-days in 56.7 calendar days.

No living animal. Its activity readouts (spikes, firing rates) are on the Lab page →

The lab's recordpaper:

paper: Tokens logged47,182

paper: Looked up16,934

paper: Traded1,482paper or real, any rat

paper: Lookups (findings)344,530

paper: Holder snapshots107,444

paper: Paper trade rows48,469all rats, incl. the random control (24,283)

paper: Lessons7,862all rats, incl. the random control (3,603)

paper: Lesson profiles with 6+ trades: 60; 10 say avoid, 0 say favor. The worst: token <10m old −⁠40.6% a trade (18), coin that just migrated off its curve on Solana −⁠21.4% a trade (22), curve coin with <50 holders −⁠20.2% a trade (19).
paper: LLM analysts gave 598 verdicts on 121 coins (1 Oct–7 Oct), all rejections.
paper: The LLM meta brain made 0 picks; its last read was 13 Oct.

Nextpaper:

paper: Slow lab: first paper decision 4 Nov 00:00 UTC; 8 days evaluated.
paper: Exit lab: its newest rule has 129 of the 40 exits it needs on pooled paper and real trades.
paper: Swarm: 5 ideas waiting for data; no paper trade since 29 Oct.

Swarm

Paper614 paper trades ·last trade 29 Oct 08:13 UTC

Research rats each test one trading idea, written down before their first trade, with $10 paper bets. A skeptical judge decides; nothing reaches the real wallet without proof.

13 ideas0 proven0 promising8 rejected5 still collecting614 paper trades closed, 0 openlast paper trade 29 Oct 08:13 UTC

Every swarm hypothesis with its paper-trade record and the judge's verdict
Rat · idea	Trades	Won	Avg / trade	95% range	P(gain)	Verdict
TuringH04 A top-10 pump.fun trader putting $3,000 or more on one coin (pool ≥ $25k) signals a rise.	0	—	—	—	—	collecting
KeplerH08 After 2 top traders converge, buy only if the price confirms +15% within 30 min.	41	19.5%	−1.0%	−⁠14.9% … +⁠15.2%	46%	collecting
LaplaceH07 After 2 top traders converge, wait for −20% and buy the dip.	52	38.5%	−2.9%	−⁠13.0% … +⁠7.5%	34%	collecting
CurieH02 3 top pump.fun traders on the same coin within 60 min: a stronger signal than 2.	23	13.0%	−14.4%	−⁠25.9% … −⁠1.1%	7%	collecting
NoetherH03 2 top traders converge, pools ≥ $50k only, trailing exit to keep the winners.	26	19.2%	−9.0%	−⁠16.0% … −⁠2.1%	5%	collecting
GaussH12 Hyperliquid perps: take the opposite side of the top traders (same threshold as Kolmogorov).	170	46.5%	−0.1%	−⁠0.6% … +⁠0.5%	41%	rejected
KolmogorovH11 Hyperliquid perps: follow the side where top traders put ≥ 75% of the dollars (≥ 5 wallets).	49	38.8%	−0.3%	−⁠1.5% … +⁠0.8%	32%	rejected
HopperH10negative control Known negative control: copy every buy of a top pump.fun trader.	42	23.8%	−5.9%	−⁠19.3% … +⁠8.9%	26%	rejected
LovelaceH05 2 kolscan KOLs buying the same coin within 60 min: it rises by more than the fees.	40	25.0%	−11.9%	−⁠26.5% … +⁠3.6%	13%	rejected
ShannonH06 A top pump.fun trader AND a KOL on the same coin within 60 min: two independent sources.	41	24.4%	−11.6%	−⁠24.8% … +⁠3.4%	11%	rejected
FisherH01 When 2 of the 50 best pump.fun traders buy the same coin within 30 min, it rises by more than the fees.	40	22.5%	−10.9%	−⁠21.6% … +⁠1.0%	7%	rejected
DarwinH09 2 top traders converge on a coin under $300k market cap, held 24 h to aim for ×3.	41	12.2%	−24.7%	−⁠46.4% … +⁠1.4%	6%	rejected
BernoulliH13 Hyperliquid perps: follow only a strong consensus (≥ 8 wallets, ≥ 85%), held up to 48 h.	49	42.9%	−1.6%	−⁠2.8% … −⁠0.3%	3%	rejected

How strict the judge is (checked 11 Nov): on 9,000 simulated rules with no edge at all, it called at most 0.18% “proven”. A rule that truly makes +5% a trade is called proven 79% of the time when its trades look like his real ones (485 trades), but only 0.16% when they look like a lottery (rare, huge wins), which need many more trades to prove. Ideas translated from the lab's French notebook; results are per paper trade, after fees.

Exit lab

Paper03:35:04 UTC

Other ways out, replayed on Damon's own entries since 10 Oct: same coin, same entry price, then each rule decides when to sell. “vs his exit” is dollars per trade against the exit he actually took.

The random exit leads at +$0.09 a trade: no exit rule has beaten chance yet.

Exit rules replayed on 98 of his real trades, dollars per trade vs his own exit
Exit rule	vs his exit	95% range	Verdict
Out at 45 min unless +5% · 53 so far stop −10%; at 45 min out unless ≥ +5%; then trail 20% under peak; 24 h max	+$0.09	−⁠$0.20 … +⁠$0.52	collecting
Random exit (control) control: out at a random time 10 min – 6 h	+$0.09	−⁠$0.23 … +⁠$0.48	collecting
Half out at +30%, trail the rest stop −10%; half out at +30%, rest trails 20% under peak; 24 h max	−$0.19	−⁠$0.67 … +⁠$0.30	collecting
Stop to +2% once +15% stop −10%; once +15%, stop moves to +2%; then trail 20%; 24 h max	−$0.27	−⁠$0.78 … +⁠$0.32	collecting
Trail 25% from +10% stop −10%; from +10% trail 25% under peak; 24 h max	−$0.34	−⁠$0.83 … +⁠$0.20	collecting
Trail 3× the coin’s volatility stop −10%; from +10% trail 3× the coin's 1-h volatility under peak (10–35%); 24 h max	−$0.38	−⁠$0.87 … +⁠$0.15	collecting
Trail 15% from +10% stop −10%; from +10% trail 15% under peak; 24 h max	−$0.45	−⁠$0.86 … −⁠$0.07	discarded
Trail 15%, wider stop stop −20%; from +10% trail 15% under peak; 24 h max	−$0.59	−⁠$1.01 … −⁠$0.19	discarded
Hold 24 h (baseline) baseline: stop −25%, else hold 24 h	−$1.32	−⁠$1.80 … −⁠$0.87	discarded

His own exits on these 98 trades made −$24.67 in total (real money); every rule above is a paper replay.

A rule becomes a candidate for the real wallet only when its whole 95% range is above $0 and it beats the random exit. It is discarded when its whole range is below $0.

Slow lab

Paper03:05:27 UTC

Slower strategies on Hyperliquid markets: tested on past data first, then followed forward on paper. None of them trades.

Forward test: since 3 Nov · 8 days scored
Last scored close: 12 Nov 00:00 UTC

Funding carrySharpe 7.13candidate
HL spot-perp funding carry (long spot / short perp, F7 >= 10%/yr)
597 return days 2025-03-16..2026-11-02; candidate label assumes no short leg was auto-deleveraged on 2025-10-10 (unregistered stress)
Now: running.
Cross-sectional momentumSharpe 0.20inconclusive
Cross-sectional momentum R14 / 7-day hold, dollar-neutral, top-20 HL perps
914 days 2024-05-03..2026-11-02; registered twin judged uninformative (vol mismatch)
Now: running.
Trend followingSharpe −⁠0.09negative
Donchian trend-following ensemble, long-only, top-20 HL perps
Backtest: negative (discard rule triggered), PROVISIONAL. 946 days 2024-04-01..2026-11-02; listing-age reading dependent; pending measured funding for Q1 2024
Now: running.

Sharpe is from the backtest, not from money made. Nothing moves to real money from here without a person deciding.

Meta brain

Paperlast read 13 Oct 05:35 UTC

Last read 13 Oct, 29 days ago.It has not finished a read since, so what follows is a snapshot of that day, not of today's market.

The meta brain read what was trending on pump.fun, named the themes (the “metas”) and judged new coins riding them. Each call was to be graded on real candles against the coins it passed on. It only judged: it never bought its picks.

As read 13 Oct 05:35 UTCMarket attention: hot

Themes the meta brain named in its last read
Theme	Phase	Strength	Leader
Launchpad and AI-agent launch platforms · emerging	emerging	55	CRAWLER
AI company name parodies · peak	peak	65	NVIDIA
Super Intelligence / SI memes · emerging	emerging	40	SI276
Black Bear/Bull/Inu market sentiment memes · emerging	emerging	38	Bear

Its summary that day: Launchpad and AI-company-parodies remain the clearest active themes with CRAWLER and NVIDIA showing real flow. Black-sentiment and super-intelligence continue at lower intensity. Many single high-volume launches (NI, DragonBall, CROWD, DOT) but no shared story across two or more.

Graded record: 0 picks, so no evidence either way.Leader coins are as the brain named them, not links or tips.

Brain internals

Simulated03:35:28 UTC

A simulated rat brain, not a living animal. For trading, the lab reads one thing from it: each channel's striatum score (D1 − D2). Among the coins the safety rules allow, he buys the top-scoring one once it clears a bar; measured on 3 Nov, that score behaves like random noise. It never sells. These numbers show activity, not learning.

Brain internals · simulated · 55,583 neurons · awakeShow

Neurons55,583

Synapses17.1M368k of them plastic

Spikes per second269,984

Speed vs real time0.81×

Mean firing rate6.00 Hz

Simulated brain-days45.6simulated time

Firing rate by region

VPM · thalamus5.2 Hz

S1 · rat cortex6.9 Hz

MICrONS · mouse cortex5.1 Hz

Fly connectome8.0 Hz

Striatum3.2 Hz

Striatum score per channel (D1 − D2) · behaves like noise (measured 3 Nov)10 above zero

7,862 lessons in the lab's database (all rats, incl. the random control's 3,603)47,182 tokens in the lab's database

How the lab works

One idea per rat
Each research rat tests one trading idea, written down and fingerprinted before its first trade. Changing the rule makes a new idea.
No hindsight
Paper trades enter at the first price read after the signal, never at the signal's own price, and pay fees both ways.
A random twin
Every paper trade has a twin: a random coin from the same feed at the same moment. An idea has to beat its twin, not just a rising market.
A skeptical judge
Verdicts only at fixed trade counts, a starting belief that there is no edge, and a correction for testing many ideas at once.
Proof means twice
A promising idea is copied and has to win again on fresh trades before it is called proven. Losers stay in the notebook.
Then a person decides
Nothing in the lab can touch the real wallet. Moving a proven idea to real money is a person's decision, in a small step.

So far no idea in the lab is proven, so nothing has moved to the real wallet from here.

The RatLab research lab