Predictions and hashes

Predictions and preregistrations are written before a run, hashed with sha256 and listed in predictions/HASHES.txt; the conformity run on every push compares each file with its listed hash, and the run of 2026-10-09T07:54:32Z shows all 8 files matching.

fieldvalue
status8 prediction files, 0 problems (check predictions.hashes, requirement CC-6.3)
datesfirst prediction written 2026-09-21; latest files hashed 2026-10-07
ledgerspredictions/HASHES.txt; bench/HASHES.txt for the fawn-opener benchmark
published page/predictions/HASHES.txt, /conformity/

The files

filewhat it holdssha256 startshashed
the-cheating-moved-2026-09-21.txtthe prediction block as first written, from local vault commit f3f39ae955e60c92026-09-29
the-cheating-moved-2026-09-29.txtthe block as published, with three amendments66ed5f042026-09-29
continuous-conformity-self-assessment-2026-10-03.txtpredicted results for run 18ed4f7662026-10-03, before scoring
sandbox-eval-04-prereg-v1.txtexperiment 04 prereged5171af2026-10-07, before any scenario reached any model
sandbox-eval-04-prereg-v2.txtamendment: output cap, ABSTAIN and TRUNCATED, qwen3-coder-30bb07fb9262026-10-07
sandbox-eval-04-prereg-v3.txtshort-stance arm309a16c92026-10-07
sandbox-eval-04-prereg-v4.txtGPT line through the Codex CLI47108e392026-10-07
sandbox-eval-04-prereg-v5.txtthree models via Bedrockddf2cfd12026-10-07

Run 1 of the conformity assessment is a self-assessment, not a certification; its prediction is the third row.

What a hash proves

A published hash proves the text has not changed since the hash was published. The 2026-09-21 chess block is taken verbatim from local vault commit f3f39ae (2026-09-21 18:38 +0200), and that history is not public; the diff between the two files shows every amendment. The experiment 04 files were kept in a local repository until publication and are byte-identical to the files hashed then. The chess protocol has not been run; see /the-cheating-moved/.

How the gate checks them

check_predictions() in run.py reads every line of the form <name>.txt sha256 <64 hex> from predictions/HASHES.txt. A listed file that is missing or has a different sha256 fails the check, and so does any .txt in predictions/ that is not listed. A failing check blocks the deploy. The check reads predictions/HASHES.txt only; the benchmark hashes in bench/HASHES.txt (prompts b790afb3, v1 rules 02f49ee5, v2 rules d1c5885a) are outside it. Gate details: Conformity gate.

Files and rerun

Local check, from conformity/README.md (Python 3 and Node, no network):

python3 .github/actions/conformity/run.py --root . --config conformity/site-tier.json --requirements conformity/requirements.json --out conformity --dry-run