E8 stops at scouting under its own width floor; BELLS question recorded as sent - #22
Merged
Merged
Conversation
…faces follow The owner posted the selection-rule ask on 2026-09-16 (America/New_York) as CentreSecuriteIA/bells_leaderboard issue #1. I verified the issue resolves and carries the Cubits11 author. The dispatch log now holds the message as sent, verbatim, with its permalink, its UTC timestamp, the reproduction commit its script URL pins, and an adjudication rule for both a substantive answer and silence. sent_at_commit stays null: the checkout at send time was not established, and a commit I did not observe is not a commit I record. The dossier moves from PREPARED to SENT, RESPONSE PENDING, and quotes the text that was actually posted rather than the draft it replaced. The qualified-outcome ledger counts a fourth technical interaction and states again that an issue the author opens is diagnostic, not an outcome. Qualified outcomes stay at 0. Derived surfaces regenerated against that count and against the merge: the research index, the reproduce page, /try/, the conformance record, the ledger snapshot, the flagship source map, the traction artifacts, the repository graph and sitemap lastmod. None was hand-edited. python3 scripts/verification_manifest.py exits 0, all 83 checks, run under .venv/bin/python — jsonschema is not importable from the system interpreter and the contract tests fail without it.
Generating distribution artifacts against an uncommitted `outcomes.yaml` records the metrics event as `pending_commit` with a null commit and a null URL. Once the source is committed the same regeneration resolves it to `baf018b` and a blob URL, and `distribute.py --check` fails until that resolution is recorded. So a change to a bound source takes two commits: the source, then the binding. The repository graph picks up this branch's topology in the same pass. I did not amend the first commit. Amending would move the sha the event now binds to and reopen the same drift.
…ore any score The contract declares the SESOI, the marginal-only width floor and its consequence before any pool is scored. The schema gains an optional informativeness group and the validator holds its floor at or above the SESOI, because |q_obs - q_ind| never exceeds the marginal-only width. The runner reports the realized width against the floor and applies the consequence: IDENTIFICATION-LIMITED, delta recorded and not claimed. scouting.json fixes the seed, the sizes, the middle band, the candidate order and the selection rule; every scouting item is burned for evaluation by hash.
…oor at a deployable budget Three candidate pools were scouted on 200-item slices at FPR 0.05 with thresholds from a separate 400-item benign calibration set. G2's threshold on OR-Bench benign is 0.999485 and it misses 93–96% of every candidate; the widest marginal-only width was 0.07 against a declared floor of 0.10, and no candidate had both marginals in the middle band. The rule written before scoring says NO-ELIGIBLE-POOL: no evaluation item was drawn or scored, the contract stays draft, the runner refuses it, and E8 has produced no row. PREREG.md is rendered from contract.json, scouting.json and the freeze by run/render_prereg.py and held to that rendering by tests/test_e8_runner.py. verify_prereg.py and the conformance record now count contract.json as the declaration correction C1 made canonical. Scouting scores and slices are committed; they are burned for any evaluation under a later contract.
…ion and freeze Same shape as b6e6a93: events.json pins each source to its committing commit, so it drifts until the source is committed. Not amended, for the same reason given there.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Six commits over
main(d3766ae), in order:baf018b,b6e6a93: the BELLS selection-rule question (How were the 170 prompts in data/non_adversarial_prompts.csv selected? CentreSecuriteIA/bells_leaderboard#1) is recorded as sent.outcomes.yamlgoes from 3 to 4 technical interactions (diagnostic only), and the traction event binds to the commit that carried it.0ae1f21: E8 declares its identification group first (SESOI 0.05, a marginal-only width floor of 0.10, consequence IDENTIFICATION-LIMITED). The scouting protocol is frozen before any score exists.ac2492b: E8's scouting stops the run under its own rule. At FPR 0.05 no candidate pool clears the width floor: Mosscap 0.040, SPML 0.005, in-the-wild jailbreaks 0.070. No evaluation item was drawn or scored, the contract staysdraft, and E8 has no observation row. The three scouting slices are burned.25a0684: the traction events bind to the commits that carried E8's preregistration and freeze.aea82aa: the cadence row for 2026-09-18.Gates at
aea82aa:verification_manifest.py, 83 checks, exit 0 (.venv, Python 3.14).evidence_guard.py --base d3766ae --candidate aea82aapasses. No trust-root or workflow file changes.0ae1f21andac2492brebuild2bc48c1and9f909c0with the co-author trailer removed. Their trees, author dates and committer dates are identical to the originals. Onlydistribution/traction/events.jsonpinned the old IDs, and25a0684regenerates it. Ignoring those IDs, the events are unchanged, and the repo graph is unchanged under its drift gate's canonical form.Merge as a merge commit: squash or rebase would orphan the commit IDs the traction events pin. Merging freezes nothing for E8, sends nothing, and does not touch E2 collection.