The bar was set first
Campaign 2 ran 2026-08-21 to 2026-08-22 against a bar sealed before the first look was spent. What follows is that bar, as a definition, and the counts of what cleared it. A refusal is the bar working: each one names the mechanism that made it.
The pre-registered bar
1. The universe and the criteria were sealed before any look was spent: the pre-registration commit carries the manifest hashes, and the launcher refuses an unsealed or dirty engine tree.
Recorded in the engine’s docs/CAMPAIGN_2_PREREGISTRATION.md; docs/INDEPENDENCE.md sessions 38 and 40.
2. Universe floors first: an instrument had to clear the sealed liquidity, history and price floors, with a loadable tape, before any strategy family ran against it. An instrument that did not is a refusal with that floor named as its mechanism.
Recorded in the sealed floors, graded in the adjudications; the desk’s refusal panel.
3. One algorithm or one refusal per instrument: every family determination is graded on the Deflated Sharpe Ratio, which counts every look taken, and the ladder gates on that number. The probability of backtest overfitting is reported beside every verdict and does not gate — pre-registered, not re-posed.
Recorded in the pre-registration; the limitation declared beside the verdicts.
4. Six pre-registered criteria, graded from the campaign artifacts by the grading tool and not by hand; both universes were required to pass all six.
Recorded in the engine’s tools/campaign2_grade.py; docs/INDEPENDENCE.md session 40.
5. Every determination is replayable: 683 of 683 replays matched, and 15,103 honest looks were recorded across seven verified chains, so the trial count behind every deflated number is the count that was actually spent.
Recorded in docs/INDEPENDENCE.md session 41.
The census
One algorithm or one refusal per instrument; a determination is one of those verdicts.
| Universe | Determinations | selected | candidate | refusal | error | selected share |
|---|---|---|---|---|---|---|
| US equities and funds (us_all) | 18,677 | 180 | 425 | 18,067 | 5 | 0.96% |
| crypto against USD (crypto_usd) | 273 | 1 | not recorded | not recorded | not recorded | 0.37% |
us_all: the four outcomes sum to 18,677, which is the determination count.
crypto_usd: the four outcome counts are not all in the source log (the candidate, refusal and error counts — the log records the total and the one selection only), so no sum is shown.
Read the refusal count precisely
Of the 18,677 determinations in us_all, 15,797 were made by a universe floor (liquidity, history, price, unloadable tape) before any family ran. 2,880 instruments ran the slate, and 605 of those received a family determination (selected or candidate).
“18,067 strategies examined and rejected” would be false; “18,067 determinations, 85% of them made by a floor before any strategy was tried” is what the artifact says.
Declared beside the verdicts
Two limitations were declared with the results, in advance of any table built from them. They are quoted from the record.
The Deflated Sharpe Ratio (DSR) board is led by cash proxies
“the DSR-ordered board’s top twenty are entirely cash proxies — money-market and T-bill ETFs — because a Sharpe divides by realised volatility and the sealed floors cover price and ADV only; 34 of 180 SELECTED have sub-5% volatility.”
Median probability of backtest overfitting (PBO) 0.527
“Median PBO 0.527; 62 of the 140 equity-like SELECTED are above 0.5 [three per-symbol values elided: this page names no symbol]. The ladder gates on DSR and not on PBO — pre-registered, not re-posed — so the consequence is stated beside every verdict, never omitted.”
Elided from the verbatim record: three per-symbol PBO values. This page names no symbol.
Why this is published
Because the operator ran his own strategies through the same audit and published the result. The overfitting auditor and the honest leaderboard hold every backtest to a stated trial count; a research campaign run by the people who built those tools is held to the same arithmetic, and its census is published the same way.
The mechanisms are counted here; the instruments are not.
Source: AGENTS.md (the repository’s progress-log pointer) and docs/INDEPENDENCE.md sessions 38–41. The graded adjudications are in the engine repository, docs/CAMPAIGN_2_ADJUDICATION_{us_all,crypto_usd}.md, which this web tree does not contain.