the record

Bot performance

Every published signal, measured. Follow call quality over time, then look inside the decisions that produced it.

source updated
just now

Published calls

Sent and alerted signals only. Screening passes and declined candidates stay in the diagnostics below.

published signals: 1,831; 764 distinct tokens · 1,811 measured
reached 2×: 36.6%; 662 of 1,811 measured published signals
median observed peak: 1.59×; 1,811 measured signals · from signal entry
median last return: -54%; 1,811 last observations · not a live quote

Call quality by day

Share of measured published signals that reached 2×, grouped by call date in UTC, and split by whether the token had already been called a day or more earlier.

all published signalsfresh callsre-alerts, 24h+ after the first callstill maturing, under 48h observed

Latest stored outcomes. A re-alert repeats a token the bot already called while discussion continues; the two populations perform differently, so read the combined line together with both. Shaded days have had less time and their rates are still rising.

daily figures

Call date	Signals	Tokens	2× / measured	Fresh calls	Re-alerts	Fold-ins	Observed
Sep 3maturing	1	1	0/1 · 0%	0/1 · 0%	—	0	—
Sep 4	90	47	36/89 · 40.4%	36/89 · 40.4%	—	1	70.1h
Sep 5	196	107	95/195 · 48.7%	86/180 · 47.8%	9/15 · 60%	49	67.8h
Sep 6	203	107	95/200 · 47.5%	73/156 · 46.8%	22/44 · 50%	52	68h
Sep 7	230	119	95/230 · 41.3%	72/189 · 38.1%	23/41 · 56.1%	70	62.5h
Sep 8	158	93	73/157 · 46.5%	56/121 · 46.3%	17/36 · 47.2%	51	54.2h
Sep 9	188	100	55/185 · 29.7%	35/124 · 28.2%	20/61 · 32.8%	46	64.1h
Sep 10	146	84	34/146 · 23.3%	29/111 · 26.1%	5/35 · 14.3%	33	67.9h
Sep 11maturing	232	126	73/232 · 31.5%	63/153 · 41.2%	10/79 · 12.7%	52	45.5h
Sep 12maturing	241	109	72/239 · 30.1%	62/173 · 35.8%	10/66 · 15.2%	70	20.2h
Sep 13maturing	146	78	34/137 · 24.8%	29/97 · 29.9%	5/40 · 12.5%	54	7.2h

Is call quality changing?

Compare adjacent 3-day mature cohorts. This comparison window is independent of the totals filter.

-7.3 pp

Change in observed 2× hit rate. This is a descriptive change, not evidence of a guaranteed improvement.

previous cohort: 45.5%
recent cohort: 38.2%

Both cohorts contain at least 10 graded positive signals aged 72h or more, with at least 48h observed. Observation lengths still vary; these are descriptive comparisons.

Minimum age 72h; at least 48h observed. Newer signals are held back. Observation lengths still vary, so these are not matched fixed-horizon returns.

Median last return change: 7 pp
Median peak change: -0.1×

What happened after the call

Published signals with valid, recorded outcomes.

reached 3×: 21.4%
still ≥1.2×: 18.7%
halved or worse: 52.9%
median time to peak: 2.3h
last value below peak: 76.3%
median observation: 49.6h

Peak outcomes are opportunities observed by the bot, not assumed exits. Last value below peak is not a pathwise maximum drawdown.

Published-call grades

Stored peak grades for published signals, including untradeable and ungraded records.

38721.1%

27515%

53329.1%

61733.7%

100.5%

—

90.5%

A ≥3× · B ≥2× · C ≥1.3× · D <1.3× · X untradeable · — ungraded. Percentages use all 1,831 selected signals.

By source

Published signals only; source is the bot’s signal type.

Group	Signals	2× / measured	2× rate	Median peak	Last return
Thesis spikes	1,222	430 / 1,210	35.5%	1.59×	-52%n=1,210
Interesting theses	410	153 / 404	37.9%	1.56×	-78%n=404
Research	199	79 / 197	40.1%	1.7×	-41%n=197

2× rate = hits / measured positive decisions. Last return uses available last observations. Missing values are shown as —.

By chain

Published signals only; denominators are shown in each row.

Group	Signals	2× / measured	2× rate	Median peak	Last return
Solana	922	312 / 918	34%	1.59×	-23.5%n=918
Robinhood	772	300 / 758	39.6%	1.61×	-86%n=758
BNB Chain	124	44 / 122	36.1%	1.59×	-88%n=122
Base	13	6 / 13	46.2%	1.93×	-81%n=13

2× rate = hits / measured positive decisions. Last return uses available last observations. Missing values are shown as —.

Research verdicts

Published research signals, grouped by their stored verdict.

Group	Signals	2× / measured	2× rate	Median peak	Last return
SPECULATIVE LONG	199	79 / 197	40.1%	1.7×	-41%n=197

2× rate = hits / measured positive decisions. Last return uses available last observations. Missing values are shown as —.

inside the bot

Pipeline & Telegram diagnostics

All selected pipeline signals, including screens, rejections and research triggers. These are a different population from published calls.

Telegram /track

The bot’s original outcome counts and formulas, preserved separately from the stricter web measures.

precision: 50.3%
recall: 32.5%
median peak: 1.53×

Hits / missed winners: 1,205 / 2,497
False positives / flat: 1,192 / 983
Correct passes: 4,519
Reached 2× / 3×: 3,702 / 2,225
Still ≥1.2× / halved: 1,749 / 5,596
Ungraded / untradeable: 80 / 4,448

Data coverage

Missing and excluded observations are visible here, never silently treated as losses.

Signals / distinct tokens: 14,923 / 7,434
Valid measurements: 10,290 / 14,923
Grading coverage: 69%
Missing last observation: 0
Ungraded / excluded: 80 / 4,554
Inconsistent price basis / invalid values: 824 / 530
Estimated / missing entry anchors: 8,933 / 37
Research failures / grading notes: 12 / 90
Overdue active grading: 364

Pipeline decisions

Graded positive decisions are passed, sent, alerted or researched. Counts may repeat the same token at different stages.

Group	Signals	2× / measured	2× rate	Median peak	Last return
denied	6,170	0 / 0	—	—	—n=0
skipped	4,534	0 / 0	—	—	—n=0
alerted	1,633	583 / 1,615	36.1%	1.59×	-55%n=1,615
researched	1,394	468 / 1,354	34.6%	1.55×	-35%n=1,354
not sent	703	0 / 0	—	—	—n=0
sent	199	79 / 197	40.1%	1.7×	-41%n=197
passed	126	41 / 125	32.8%	1.32×	-92%n=125
added to the earlier	89	0 / 0	—	—	—n=0
dropped	63	0 / 0	—	—	—n=0
failed	12	0 / 0	—	—	—n=0

2× rate = hits / measured positive decisions. Last return uses available last observations. Missing values are shown as —.

Pipeline grades

Distribution across all selected signals. A grade measures the observed peak, not the realized result.

2,22514.9%

1,4779.9%

2,64117.7%

4,05327.2%

4,44829.8%

—

790.5%

A ≥3× · B ≥2× · C ≥1.3× · D <1.3× · X untradeable · — ungraded. Percentages use all 14,923 selected signals.

Missed winners to review

The largest observed peaks among declined signals. Multiple decisions can refer to the same token.

Token	Decision	Signal date	Observed peak	Last observed
MEMERobinhood · New-pair screen	denied	Sep 3	535.27×	137.63×
FLYBRAINRobinhood · New-pair screen	denied	Sep 10	407.42×	63.16×
KNOTSSolana · Watched accounts	skipped	Sep 6	306.64×	239.83×
KNOTSSolana · Watched accounts	skipped	Sep 6	284.59×	230.38×
UBERSolana · Watched accounts	skipped	Sep 6	244.36×	41.47×
S3Solana · New-pair screen	denied	Sep 7	174.15×	3.87×
FLYBRAINRobinhood · Watched accounts	skipped	Sep 10	169.11×	26.21×
TENDIESSolana · Watched accounts	skipped	Sep 11	165.7×	76.55×

Up to eight signals, ranked by peak multiple. Excludes untradeable pools and inconsistent price measurements.

Research processing

Available processing log: Sep 5, 2026, 07:50 AM UTC – Sep 13, 2026, 05:13 PM UTC. This log is independent of the filters above.

Queue & completion

Counts reflect the available research-processing log, not the entire history.

median queue wait: 17s
pending: 26
sent: 7
added to the earlier post: 8
not sent: 19

retained automation state

research runs: 60
tracked pairs: 1,101
screening triggers: 100

saved experiment

Nightly rule lab

Report 2026-09-12 · 06 Sep 00:11 → 11 Sep 23:48 (5.98 days covered, last 7.0 d, signals at least 24 h old). This saved report uses its own evaluation window, independent of the filters above.

screened: 1,014; 1,010 graded · 4 ungraded
runners in the sample: 373; 253 reached 3× · 823 duds
actual precision: 24.2%; 16 runners / 66 graded passes
actual research per day: 11; Recorded screening workload

Candidate rules

The bot’s saved counterfactual comparison. Candidate outcomes describe the evaluation sample, not future performance.

Candidate	Passes	Runs/day	Precision	Recall	Extra runners	Extra duds	Extra precision
callouts_walletscandidate only	181	30.2	35.4%	17.2%	49	88	41.5%
crowdcontrol rule	472	78.9	37.5%	47.5%	164	346	39.5%
rh_structurecandidate only	160	26.7	31.9%	13.7%	48	132	33.1%
score5candidate only	564	94.2	36%	54.4%	187	422	37.6%
wallets_onlycandidate only	113	18.9	36.3%	11%	41	95	38.7%

saved decision

No change: no candidate cleared every gate and no promoted clause decayed.

rules & promotion constraints

Recorded main rule: sub_t>=1 & corrob & holders>=200 & score>=6 & social_or_two | sub_t>=2 & holders>=200 | theses>=8 & light_t>=3 & kol>=8 & holders>=500

callouts_wallets: substance from a thesis OR a callout, plus a notable wallet and 150 holders Rule: sub_any & wallets_lite & holders>=150. Constraints: watched 8.0 d · held back: 19.7 extra runs/day > 6
crowd: control: 8+ people posting, no substance needed Rule: people>=8. Constraints: watched 8.0 d · held back: control candidate, never promoted; 69.3 extra runs/day > 6
rh_structure: Robinhood, under $400k, 3+ KOL and 20+ smart wallets, 200-1000 holders, whatever the text (automation research 2026-09-05) Rule: rh & small_cap & kol>=3 & sm>=20 & mid_holders. Constraints: watched 7.9 d · held back: extra precision 33% < 35%; 24.2 extra runs/day > 6
score5: screen score 5 or more Rule: score>=5. Constraints: watched 8.0 d · held back: 83.2 extra runs/day > 6
wallets_only: ticker plays: 8+ KOL, 40+ smart wallets, 500+ holders, whatever the text Rule: kol>=8 & sm>=40 & holders>=500. Constraints: watched 8.0 d · held back: 17.7 extra runs/day > 6

Definitions, denominators & limitations

Published-call measures include only signals marked sent or alerted. Pipeline diagnostics also include screening and research decisions. Counts are signals, not unique positions or executed trades.
Signals are decisions, not unique trades. One token can appear at several pipeline stages. Token counts are deduplicated by chain and address.
Peak hit rate is the share of graded positive decisions that reached 2×. Positive decisions are passed, sent, alerted and researched; screen passes are not necessarily Telegram calls.
Telegram precision is HIT / (HIT + false positive), excluding flat outcomes. Telegram recall is HIT / (HIT + MISS). The separate Telegram summary preserves its original denominator and upper-middle median.
Web performance excludes untradeable pools, inconsistent price bases, invalid multiples and unknown decisions. Missing observations are not zero returns.
Return, peak and drawdown summaries use graded positive decisions. Current means the last graded observation, not a live quote. Drawdown is the last observation below its observed peak, not maximum drawdown along a price path.
A: peak ≥3×; B: ≥2×; C: ≥1.3×; D: below 1.3×. X pools are untradeable. Still up means last observed ≥1.2×; halved means ≤0.5×.
Day groups use signal date in UTC and the latest stored outcome. They are not a price history or portfolio equity curve. Mature comparisons use adjacent equal calendar windows and a shared minimum observation period, not identical fixed-horizon returns.
The source grades for up to 72h on a widening schedule. Estimated anchors, missing grades, failed research and overdue active grading are reported separately. Private pump.fun follows and vamp watches are excluded.
Fresh calls are rows posted within 24h of the token's first published call; re-alerts are later rows on the same token, which the bot posts whenever discussion continues. The two populations perform differently, so the daily chart shows them separately as well as combined.
The stored peak is recomputed from coarser bars as a signal ages and is not a running maximum, so a brief spike counted as a hit one day can be un-recorded the next. A day's rate is therefore uncertain by a few points even after it has fully matured.
Silent thread additions (a spike folded into an open post, with no new message) are collected as spike rows and count as published signals; their count is shown beside each day.
Coverage: Sep 3, 2026, 05:59 PM UTC through Sep 13, 2026, 06:25 PM UTC. Last grading attempt: Sep 13, 2026, 05:41 PM UTC. A grading attempt does not guarantee a new valid price observation.