Negative-result audits
The public ingestion job no longer has to replay every submitted negative task. New compact banks from accounts with at least 256 previously replay-verified tasks use independent 1-in-20 sampling. New accounts, existing pending banks and legacy full receipts retain full replay. Every supplied candidate identity receives an exact check regardless of sampling or queue budget.
An unreplayed claim earns no verified leaderboard credit, supplies no model-training counters or timings, and enters no exact completed-task or mathematical-coverage index. It is retained in the issue audit as unreplayed_tasks. A bank labelled sampled is processed, not certified as wholly correct. The leaderboard now ranks contributed inputs, with the replay-verified portion shown separately. Complete sampled banks with no rejected tasks and at least one matched random challenge receive provisional credit for unique unreplayed tasks. A bank with no random checks receives no provisional credit yet. Input counts come from fixed task bounds, not client counters. Claims are deduplicated across banks and accounts, and an exact replay replaces rather than adds to provisional credit. Any failed account audit revokes that account’s entire provisional balance, preserving exact verified work. Historical retained claims are treated by the same rule; no contributor receives a manual adjustment. Sampling does not multiply a sampled score by twenty.
Selection and failures
The verifier generates a fresh 256-bit random salt after receiving the immutable bank body, and durably records it before executing any task. SHA-256 of the salt, body digest and canonical task ID selects each negative task with probability approximately 1/20. A resumed job reuses its original challenge; it does not reroll. Selection is not a client-visible task-ID partition that can be avoided before submission. A small bank can receive no sampled tasks at all; that earns no verified credit.
A sampled digest mismatch disables sampling for the account, both for remaining new banks in that job and later jobs. Remaining tasks in the failed bank receive full replay. Previously skipped tasks remain unverified; a failure does not somehow certify or erase them. A later full replay can establish their result. If someone copies a retained public digest into another bank, a matching later replay credits the earliest retained matching claim, not the copier. A false digest cannot reserve that credit. The original audit is linked to the later verification. Public salts and recorded digests make each completed audit reproducible, but do not make the work attribution cryptographically provable. An edited issue is a new body, never an alteration to its previous audit.
What this saves, and what it costs
For eligible newly submitted negatives, expected replay task count falls by 95%. This is not a claim that total Action runtime falls by 95%: checkout, issue polling, ledger I/O, model fitting and publication remain. Replay now uses the Rust kernel with a Python cross-check (see NATIVE_KERNEL.md); issue polling, ledger I/O and publication remain. Measure the remaining cost rather than assuming linear scaling.
There is less new certified coverage per bank. Unsampled tasks can be selected again by another client because an unverified claim is not a valid exclusion. Counting them as completed would let a dishonest client suppress an actual solution. Submitted logical counts alone also cannot support a throughput model; training continues exclusively on exact replays.
Random audits detect widespread cheating more readily than isolated omissions. If a fixed fraudulent task has a 5% audit probability, its omission escapes that audit 95% of the time. For 100 independently selected fraudulent tasks the chance all escape is about 0.6%. These are audit-sampling calculations, not solution probabilities. Probation does not prove future honesty. This implementation deliberately does not convert a passed sample into a claim about all other tasks in that batch.
The shell-count and progression-conservation identities are useful additional checks, but a client can fabricate those counts. They cannot replace a proof of execution. GitHub issues remain the transport, with up to 256 tasks per bank and one automatic bank per ten seconds per runner, with persistent rate-limit backoff. There is no per-computer account restriction; deduplication and first verified attribution still apply.
Operations
python tools/ingest.py --issues --sample-negatives 20
Without --sample-negatives, ingestion retains full replay. The live workflow explicitly selects the policy, so local fixtures and manual full audits remain conservative. cluster.json separately publishes unreplayed_claims (submission claims, not deduplicated coverage), sampled_banks, and each bank's unreplayed count. All history and audit decisions remain on cluster-data.
A future probabilistic-coverage ledger would require its own labels, adversarial analysis, reservation/redundancy policy, and contributor-credit rules. It must not be silently merged into the existing exact ledger.