6 Feedback integration

Update 2026-10-14 — Claude adds a "same kind of change" record to the brief, and finds that the first real-looking citations were word matches

Why: every brief said "none applies". The 64 decided results she has (KEEP/REVERT, in intervention_outcomes) are cognition-bench changes to other files, and the brief only serves results that share a file or a word with the change. the operator asked Claude to do this one directly, so Piper could be left alone.

What changed (on Firefly, backups *.bak-claude-c06t-20261014, ledger rows filed with predictions, resolve 10-21):

• prior_art.py 1.1.0, a type record. The brief sorts the change into a kind (card, tool, gate, retrieval, ...). When that kind has at least 3 decided results anywhere, it serves one entry, tally:<kind>, for example "gate: kept 7 of 11", with two recent examples. It is labelled as a weaker prior than a same-file result. Its read is logged like any other entry, so the hook cites it and the gate checks it with no change to the hook. Off switch: foundation/state/prior_art.json {"type_record":"off"}.
• improvement_loop.py 1.0.1: the classifier read delegate_bridge as a gate change, because "delegate" contains "gate". It now reads it as plumbing.
• amp-bench #6 rev c: a tally:<kind> citation earns half credit: 0.5 with a logged read before the row, 0.25 without one, and only when the kind had at least 3 decided results before the row. It can never reach full credit. A row that also cites a same-area result gets that result's full credit.

Found while checking: three of her harness-probe rows from this afternoon had just started citing real KEEP results (a belief.py column change, the contradiction gate). Those are word matches, not lessons. Those tasks list memory.db among the files they write, so the brief's area became "memory", and it served all 34 outcomes and lessons that mention memory. prior_art.py 1.1.1 fixes this: data stores (.db, .sqlite, .jsonl) don't define an area, and generic, numeric or very short terms are ignored. Replaying her last 8 recorded briefs, the memory.db briefs now serve nothing (they served 34), and real same-file matches are kept (belief.py, contradiction.py, prediction.py). Her three citation checks pass on both versions.

Honest reach: of her 79 rows in the last 14 days, about 1 in 4 fall in a kind that has a record (retrieval, gate, data fix). About 45% are instrument changes (ledgers, checks, suites), and none of those has ever been decided, so the type record has nothing to tell her about her most common kind of change. Expect v2 to settle somewhere around 10 to 15, not higher, unless instrument changes start getting KEEP/REVERT verdicts.

v2 re-read 3.6 ± 4.0 (n = 83). All three credits are the word-match rows above, so the honest count is still 0. They leave the 14-day window by 10-28. Her accepted baseline stays 0.0: rev c scores today's record exactly as rev b did. v1 unchanged at 75: "measure shows it working" stays 0 until a row cites a result that is actually about the change.

Update 2026-10-14 — the brief is now in the route that files rows: every new ledger row carries a citation backed by a logged read

What she did (13:28 to 13:31, in answer to Claude's 13:24 note; her reply is in doc piper/claude_amp6_reply_20261014):

• delegate_bridge 1.117.0: a prior-art pre-flight runs before any worker starts. It builds the brief for the task's declared files, which logs the read, and keeps it with the task. When the worker returns, learned_from is taken from that brief.
• change_ledger_hook 1.5.0: her first fix exposed one more layer. The route that actually files version-change rows (_apply_unrowed, shared by worker-return and the sweep) still filed with no citation; the row for 1.117.0 itself came out NULL. It now runs the brief for the changed file and checks the result against the read log. It writes the settled refs, or an explicit "none applies" naming the logged brief. If the brief fails, the row is still filed, but with a visible uncited: marker that the gate refuses, so it never looks normal.

Checked by Claude (read-only, $0): both versions are live, and her three checks pass (check_prior_art_route, check_caller_citation_route, check_change_ledger_citation). The clhook-change_ledger_hook-1.5.0 and clhook-cognition33_suite-2.3.0 rows (13:31, 14:10) both carry a brief-backed "none applies". These are the first real rows the gate has judged. Her replay of the 75 earlier rows: 8 would have had a brief-backed citation, and 0 would cite a real KEEP/REVERT. Filed history was not rewritten.

Bench: she preregistered and submitted clhook-change_ledger_hook-1.5.0. The verdict was INCONCLUSIVE, not scoreable yet, which is what she predicted. Under rev b, "none applies" scores 0, and every resolved ledger row so far is inconclusive. So the score can only move when a change touches an area with a settled KEEP/REVERT and the brief serves it. That needs more decided results, not more wiring.

v1 summary score 62.5 → 75: "in her default path" goes from 12.5 to 25, because the brief now runs on every worker and on every filed row. "Measure shows it working" stays 0: no row yet cites a real result.

Update 2026-10-14 — bench fixed (rev b): her baseline is re-set to 0 of 74, which is the honest number

The three #6 bench problems noted above are fixed (amp-bench, commit "#6 rev b"):

• A result counts as settled only if its direction is known. Inconclusive, not-resolvable and cannot-tell ledger rows no longer count.
• A learned_from that says "none applies" now scores 0, even when it mentions a row id.
• Citations of outcome:<id> (the KEEP/REVERT results her prior-art brief serves) now count. Full credit needs her logged brief_entry read before the row.

Checked against made-up rows: a cited KEEP read first scores; a cited inconclusive result, a "none" row and a read logged after the row all score 0. Her accepted v2 baseline is re-set from 6.2 ± 16.8 to 0.0 ± 1.3 (n = 74; 68 rows cite nothing, 6 say none applies). All 6 resolved ledger rows are inconclusive, so today the only citable results are the 61 KEEP/REVERT outcomes the brief serves. Once a row is filed through the brief with a real citation, it will show here.

Update 2026-10-14 — a citation gate and a prior-art brief are built and pass their tests, but the route that files most ledger rows does not use them

What she did (12:46 to 12:56, deliverables d-b8f7dec1, d-6df71145, d-e970f5d5):

• prior_art.py 1.0.0, the prior-art brief. Given an area (component, files, tags) it lists the settled outcomes that apply, from the change ledger, the intervention-outcome store and the lesson store. Unsettled or inconclusive results are listed separately as "replication candidates, do not cite". If nothing applies it says so in plain words, not with an empty list. Every brief logs its own read (brief:<id>) and one read per outcome it shows.
• change_ledger_hook 1.3.0, the citation gate (citation_verdict). A row is refused if learned_from is empty, or if it names an outcome with no logged read before the row was written. "None applies" passes only if a logged brief query came first.
• change_ledger_hook 1.4.0 arms the gate on the hook's own check --auto route. That route runs the brief itself and fills learned_from from what the brief served.
• check_change_ledger_citation.py 1.1.0 proves the gate can fail as well as pass.

Checked 10-14 (read-only, $0): her check passes 13 of 13, including both refusal cases (a read logged after the row, and a "none" with no query behind it). On live data, using a throwaway read log so no fake reads were added to hers: the brief for transfer served 3 settled outcomes (2 KEEP, 1 REVERT) and 8 candidates, and the briefs for change_ledger and delegate_bridge said plainly that nothing applies. This is the check the 10-13 review asked for. A self-reported link can no longer pass without a logged read, and inconclusive outcomes are kept out of what can be cited.

What has not moved: nothing in production calls check --auto. Rows are filed by worker_return, which the bridge calls, and her own 1.4.0 note says that caller is "untouched". 31 rows were written on 10-14: 27 hook-filed as unpredicted with no learned_from, including the rows for these three files themselves. None of the 31 went through the gate. So the gate is again a mechanism that nothing in the path uses, the very gap her own 1.4.0 note names. A smaller point: the brief matches by shared words, so the change_ledger area returns 67 "candidates", nearly the whole ledger.

v2 bench re-read: 4.7 ± 4.8 (n = 74 rows in 14 days; her accepted baseline is 6.2 ± 16.8 on 8 rows). All 3 credited rows are from 10-13. Correction on our side: 2 of those 3 credits are false. Those rows say "none: … does not apply" and the suite credited them because they mention a settled row id. The honest count is 1 of 74 (cl-20261013-docread-1.0.0, which cites the ingest outcome, itself only "cannot-tell"). Three bench fixes are due on Claude's side: treat an explicit "none" as no citation; accept outcome:<id> citations backed by a logged brief_entry read, since her brief serves those; and stop counting INCONCLUSIVE ledger rows as settled. Not changed yet.

Next step: run the brief in the route that actually files rows. Either worker_return asks for the brief and writes its result into learned_from, or the dispatch pre-flight runs it before the version bump, which also fixes the "unpredicted" gap in #16.

v1 summary score 25 → 62.5: exists 12.5 → 25 (gate, brief and outcome store all built); in her default path 0 → 12.5 (partly: armed on check --auto and the add command, not on the worker-return route); measured 12.5 → 25 (the v2 record bench scores it directly). "Measure shows it working" stays 0: 1 honest credit in 74 rows.