Occams · reviews
The inference review of 2026-10-03, verified by rerun
Before anything in the review was planned or built, its claims were checked against this repository at 2e86d58 (TASKS-v4.md M16.0). Every claim read in the code holds at the file and line it names. Its probes were rerun independently, from its descriptions, in a scratch environment that touched no Register.
| Claim (review id) | The review's figure | Rerun here |
|---|---|---|
| position-boxed beats-null on a world with no effect, false positives at 0.01 and 0.05 (F01) | 0.140, 0.225 | 0.140, 0.225 (200 seeds) |
| the same null resampled at the winner's trade count | 0.015, 0.100 | 0.015, 0.100 |
| the fifth check on that world | 0.065, 0.135 | 0.065, 0.135 |
| day-boxed beats-null under a common factor of 0.5 (F02) | 0.090, 0.200 | 0.093, 0.213 (150 seeds) |
| the same with no common factor | 0.010, 0.050 | 0.000, 0.020 |
| a coin flip passes the fifth check on the day-boxed null control (F04) | passes | passes: make null --engine=day_boxed has printed it since ADR-0043 |
| a forged and re-chained Register verifies; a truncated one verifies (F11) | 37, 20 | 37, 20 |
| the required counts and bounds of its appendix (A5) | — | match; one count differs by one |
What the reruns mean
- On the engine that measured Q-005 and both programme 2 questions, the beats-null check fires fourteen times as often as its declared 0.01 on a world with nothing in it.
- On the day-boxed engine the null is exact when trades are independent and far too narrow when they share a date and a market.
- The fifth check, added so that a supported verdict attests the entry and not the side, can be passed by a coin flip when the paths drift.
- The chain shows that one line was not changed without the rest. It does not show that the file is the one that was written.
Every one of these errs toward a false positive. The four null verdicts therefore stand under every correction, and the one supported verdict, Q2-001, is where a correction can bite.
Three corrections to the review
- A survey's readiness cannot be recomputed from recorded results. A survey cell file holds summaries and no per-date series, so survey-wide inference needs every cell run again.
- The lab imports four more vendored modules than the review lists:
audit,result,privacyandcharset. - Its cross-check rule misfires at small values. Refusing when two p-values differ by more than a factor of two refuses harmless disagreements far below alpha; this lab refuses when the two fall on opposite sides of the corrected alpha.
What was not rerun
The verdicts themselves. They are re-scored, as diagnostics beside the record and never in its place, by M16.18 under ADR-0047.