Measurements · updated 2026-09-13

Accuracy ledger: what each risk band did next

Did the tokens we flagged actually behave worse than the ones we did not? 30,382 tokens, scored before we knew, split by graduation so the answer cannot be an artefact of who graduated. Misses first.

Every cell shows its n. "Now" figures are as of 2026-09-13 14:06 UTC and move; drawdown-from-peak does not.
30,382tokens, scored before the outcome was known
89.4%of never-graduated HIGH tokens sit ≥90% below their peak (LOW SIGNAL: 74.9%)
93.7%of graduated HIGH tokens (LOW SIGNAL: 88.9%)
1.6%of never-graduated HIGH tokens are up right now (LOW SIGNAL: 3.4%)

Read this first: what this ledger does not show

It does not show that the score predicts forward return. We tested exactly that on a held-out set in September and it did not replicate. Nothing on this page is a trading signal, and we will not publish a forward-return claim.

It does not show that LOW SIGNAL is safety. 74.9% of never-graduated LOW SIGNAL tokens also sit ≥90% below their peak. A low score means our coordination checks did not fire. That is all it means.

It never shows a pooled "ever 2×" figure. HIGH tokens graduate more often, and graduated tokens 2× more often, so the pooled number reverses once you split by graduation. Every table here is split for that reason.

The band is the score at first sight — coordination only. A token with a Token-2022 transfer fee and a coordination score of 0 sits in LOW SIGNAL here. Mint-level traps are a separate finding on the scan page and are not part of this table yet.

What each band did next

Every token was scored by the scanner before its outcome was known, then polled roughly every 15 minutes. Bands are the published score bands (<35 LOW SIGNAL, 35–64.9 ELEVATED, ≥65 HIGH). "Graduated" is our market-cap proxy for the bonding-curve event, not the event itself. Prices marked "now" are as of 2026-09-13 14:06 UTC and drift; the drawdown-from-peak columns use each token's own observed peak.

Never graduated

band at first sightn≥90% below peak now≥90% below first seenup right nowever 10×
LOW SIGNAL9,73374.9%73.3%3.4%0.1%
ELEVATED8,98370.7%69.5%2.1%0.2%
HIGH2,53189.4%89.0%1.6%0.7%

Graduated

band at first sightn≥90% below peak now≥90% below first seenup right nowever 10×
LOW SIGNAL5,34888.9%82.7%5.8%7.9%
ELEVATED1,84489.2%76.6%5.5%10.3%
HIGH1,94393.7%88.7%2.6%13.2%

The pattern that survives the split, in both cohorts: HIGH-band tokens collapse more completely, are less often up now, and 10× more often — an engineered spike and an engineered collapse. That is consistent with what the product claims to detect, and it is the only claim we make from this table.

How much of the peak is left

Measured from the stored price path only — the highest price within our observation window, then the price at three later points. Nothing here uses a live price, so it cannot change after publishing.

band+60 min+6 h+24 h
LOW SIGNAL63.9% n=15,03935.4% n=14,9208.3% n=14,467
ELEVATED66.2% n=10,81346.9% n=10,75918.8% n=10,456
HIGH57.1% n=4,46728.7% n=4,4510.6% n=4,382

Median share of the observed peak retained. The peak is the highest price we saw, which for a token found while already falling is the first observation — this understates true peaks and is stated so.

Method, so you can argue with it

Sample: checks>=5, first_price>0, max_price>0. Band basis: cabal_score at first sight: <35 LOW_SIGNAL, 35-64.9 ELEVATED, >=65 HIGH (score only). Graduation basis: outcomes.graduated (market-cap proxy, not the curve event).

Definitions: "≥90% below peak now" is current price ≤ 10% of the token's highest observed price; "≥90% below first seen" uses the first price we recorded instead; "up right now" is current price above first seen; "ever 10×" is observed peak ≥ 10× first seen.

Never published from this data: pooled ever-2x (Simpson's paradox); any forward-return claim (did not replicate out-of-sample).

Reproduce: scripts/ledger-build.py reads the outcome database and writes the JSON this page is rendered from. Recomputed 2026-09-13.