Track record

No call has reached the 72h mark yet. First outcomes land Oct 18, 10:45 PM ET (24h). Pending calls show their current state below, but it is not counted.

73 published calls logged · 45 outcomes scored · 13 "case reopened" re-verdicts and 4 unpublished call(s) logged but not scored.

24h after the call

Balanced accuracy 59.3% (n=38: 18 v4, 20 legacy) · false-positive rate 7/12 survivors flagged (58.3%) · false-negative rate 6/26 rugs/dumps passed (23.1%, 5 outright rugs)

In context: 25/38 calls matched the outcome (65.8%, 95% range 49.9–78.8%). Raw hit rate depends on how many coins rug, so it is not our headline. 7 call(s) excluded: already down 80%+ at call time.

vs RugCheck (same coins, same labels)

6 scored call(s) excluded from this comparison: no RugCheck report captured within 45 min of the call.

Stated chance vs what happened: the chance the price is down 50%+ within 24h

Brier score 0.182 vs 0.315 for a base-rate-only guess (lower is better; n=18). Mean stated chance of a 50%+ price drop 69.9%; actually RUGGED or DUMPED 66.7%.

These numbers are scored against our locked outcome labels (RUGGED or DUMPED), which also count liquidity drops, LP pulls and mint use, while the model only predicts the price, so a small gap is expected. In the model's training data the two targets disagree on 3 of 477 labeled snapshots at 24h and 5 of 355 at 72h. Base-rate guess = the share of training snapshots whose price fell 50%+ within 24h, for the exact model version each call used (mean 36.3%). It is out-of-sample for our calls and was known when we made the call.

72h after the call

No pass/flag calls have reached 72h yet (n=0).

Stated chance vs what happened: the chance the price is down 50%+ within 72h

No v4 call with a stated 72h price-drop chance has been scored yet (n=0).

7d after the call

No pass/flag calls have reached 7d yet (n=0).

We don't state a price-drop chance for 7d, so this horizon is graded on balanced accuracy only.

Not model-rated yet (bonding curve)

No calls yet. Coins still on the pump.fun bonding curve get no tier and no odds, and are not counted in any rate above.

How a call is graded

• Outcome labels are locked (locked-v1, Oct 17, 2026). They cannot change without a public note.
• RUGGED: price down 90%+ since our call, OR mint/freeze authority used after the call (supply up 5%+), OR LP pulled (liquidity down to 30% or less of the call level and far below what the price move explains, or pool under $100).
• DUMPED: price OR liquidity down 50%+ since our call (market cap when the post did not record a price; liquidity rules skipped on pump.fun bonding curves).
• SURVIVED: neither. UNKNOWN: coin missing from DexScreener, or data unavailable 12h after the mark (excluded, never counted as a rug). Calls made when the coin was already down 80%+ in 24h are excluded.
• Headline: balanced accuracy (average of the share of rugs/dumps we flagged and the share of survivors we passed), plus false-positive and false-negative rates, at 72h; 24h and 7d shown too.
• Pass calls (LOWER, SURVIVOR; before Oct 17 9:49 AM ET: CLEAN, NO RED FLAGS YET) are correct when the coin survives. Flag calls (REKT, EXTREME, HIGH; before: SUS, REKT) are correct when it rugs or dumps. Old calls keep the tier they were posted with.
• NOT MODEL-RATED YET (since rules v4.1, Oct 17, 2026): coins still on the pump.fun bonding curve get no tier and no stated chance (risk_p24/risk_p72 are null). They are left out of every headline rate and listed separately. Hard REKT rules still apply to them.
• ELEVATED (40–60% chance the price is down 50%+ within 24h) is a range, not a pass or a flag. It is left out of balanced accuracy and FP/FN and graded only on its stated chance.
• Stated chance: since rules v4, each call states the chance the price is down 50%+ within 24h and within 72h (risk_p24, risk_p72). The model was trained on 214 coins (1,155 snapshots) out of a 656-coin random sample of pump.fun graduations (Oct 9–14, 2026); coins under $3K liquidity were excluded. We grade the stated chances with the Brier score (mean squared error; 0 is perfect) against a guess that always says the training base rate, and with a calibration table per tier. These numbers are scored against our locked outcome labels (RUGGED or DUMPED), which also count liquidity drops, LP pulls and mint use, while the model only predicts the price, so a small gap is expected. In the model's training data the two targets disagree on 3 of 477 labeled snapshots at 24h and 5 of 355 at 72h. Calls with no modeled chance are counted and left out.
• False positive: we flagged it and it survived. False negative: we passed it and it rugged or dumped.
• RugCheck benchmark: RugCheck "flags" a coin if its report at call time (within 45 min) marks it rugged or has any "danger" risk (pre-registered; no score cut-off); otherwise it "passes". Same coins, same labels.
• Prices are from DexScreener (GeckoTerminal hourly close when scoring runs more than 1h late), authorities and supply from Solana RPC. RugCheck is only the benchmark; it never sets our outcome labels.
• Calls from Oct 16 to 17 were backfilled from our saved posts on Oct 17 (see logged_at_utc). Some metrics were parsed from the rounded numbers we posted.

Full checks we run: methodology.

Every call

Verify it yourself

Raw logs: calls.jsonl · outcomes.jsonl · record.json. Both logs are append-only sha256 hash chains: each line's hash = sha256(prev_hash + the line's JSON without hash, keys sorted, no spaces). Change any past line and every later hash breaks.

Calls chain: 90 records, intact, head 4c4f8af943fb8a1d…
Outcomes chain: 45 records, intact, head ed95e66a8f6fa597…

Not financial advice. Every verdict is a snapshot of on-chain risk at check time, not a guarantee. Small samples swing a lot. DYOR.