daily-fix: inline-round scientific duties

Overview / Motivation

Auto-filed by the /daily 2026-07-25 problem sweep. The user-chat inline free-analysis carve-out deliberately skips the planner + critic stack, so its scientific duties live as an explicit list in CLAUDE.md. Three failures yesterday show the list is missing three entries.

Goal

Add three inline-round duties to the CLAUDE.md § "User-chat inline free analysis" carve-out: an n-versus-d guard-rail on ridge/linear-map fits, a pre-launch diff of any re-implemented estimator against its in-repo reference, and a same-turn obligation when a round refutes a claim in a promoted task body.

Workflow gap

1. Two headline runs launched below the interpolation threshold. Session dffde9b6 @ 22:20:13Z launched a pilot and then a full 4-cell ladder against a 4-shard subset: n = 1,877 rows against d = 3,584 features. The within-regime ceiling came out at 0.099 against a published 0.625 — the subset sat below the n > d interpolation threshold, so every held-out R² in that run was estimator-degenerate rather than a signal read. Caught only after two launches, at 22:41:22Z, and it required writing a streaming extractor and re-staging to n = 4,724. Assistant verbatim: "my first full run would have produced a real ladder verdict on a wrong operator." This is the same degeneracy class another session had spent the same afternoon diagnosing for #825.
2. A re-implemented estimator diverged from the shipped reference, found mid-run. Same session @ 22:33:00Z, 2.5 minutes into the full ladder run: the round's rungs 8/9 fit the answer alignment B from the map's predictions → target answers, whereas both the note and the shipped issue1345_operator_comparison chain fit B on the answer clouds. The round's version is strictly more permissive — it absorbs map error into the "coordinate change" verdict. Killed, fixed against the shipped _chain, relaunched.
3. A round refuted a promoted body's Takeaway and filed nothing. Session 63122023 @ 18:56:45Z established that #825's "the user's next turn is linearly unpredictable" is a λ-selection artifact (all four M user cells flip from −1.43…−1.65 to +0.19…+0.25 under two independent selectors). The round explicitly flagged that #825's promoted body still bolds the false Takeaway and that the 12 banked role-map-comparison ROLE-GAP deltas are all computed unguarded — and then filed nothing. Verbatim: "One thing I did not fix: #825's promoted body still carries 'the user's next turn is linearly unpredictable' as a bolded Takeaway…". Counting method: every tool_use Bash command in both sessions was grepped for task.py new|file_infra_task|set-body|set-title|set-clean-result; two hits, both for creating #1689 — no corrective filing of any kind.

• Why the existing carve-out misses these: it already mandates a compute-character pre-launch statement, both mapping arms, and figure sanity. Items 1 and 2 are estimator-validity preconditions (is this fit even identifiable? is it the same estimator the repo already ships?), and item 3 is a record-integrity duty. The carve-out's same-turn completion contract covers committing artifacts and folding the finding into the parent — it does not cover a finding that invalidates a DIFFERENT task's promoted body.
• Confidence (emitter): high on all three (each quoted from the session's own text). Item 3 has the widest blast radius: a promoted clean-result asserting a refuted claim, with no owner.
• verified-at-filing: absence confirmed in the named target — searching CLAUDE.md's § "User-chat inline free analysis" block for the three concepts returns nothing: grep -c 'interpolation threshold\|n_train' CLAUDE.md → 0; grep -c 'reference implementation\|shipped estimator' CLAUDE.md → 0; the carve-out's same-turn contract enumerates (1) commit + push artifacts with the staged-index verification and the inline payload lint gate, and (2) fold the finding into the parent's clean-result or carry an explicit deferral — with no clause for refuting another task's promoted body. Incident text quoted from the two sessions' own messages; the "filed nothing" claim is the grep result stated above, not an inference. Landed-fix history check git log --oneline --since='7 days ago' -- CLAUDE.md → the wave touched CLAUDE.md via several merges; none adds an inline-round estimator-validity or refuted-body duty. (2026-07-25)

Proposed change (refine in planning)

  CLAUDE.md § "User-chat inline free analysis" — three added duties:
+ (i)  n-vs-d guard-rail: before any ridge / linear-map fit, compare n_train against
+      the feature dimension d and refuse (or loudly WARN, stated in the dispatch
+      note) when n_train < d — every held-out R² in that regime is
+      estimator-degenerate, not a signal read.
+ (ii) estimator-reference diff: an inline round that re-implements an estimator
+      already shipped in-repo records, in its dispatch note BEFORE launch, a diff of
+      the new estimator against the named reference (function + file) — not after.
+ (iii) refuted-promoted-body duty: when a round refutes a claim in ANY task's
+      promoted body (its own parent or a sibling), it must apply the body correction
+      or file a task for it IN THE SAME TURN; a chat-only "I did not fix X" is an
+      incomplete round.

Item (i) may be better as a mechanical assertion in the shared fit helper than as prose — the planner should check whether analysis/mapping_baselines or the ridge path has a natural home, since a guard-rail in code covers non-inline rounds too. If so, do both: the code assert and the one-line duty.

Scope / surfaces

• Primary target: CLAUDE.md § "User-chat inline free analysis" (and the parallel Step 9a-ter block in .claude/skills/issue/SKILL.md, which mirrors these duties — the two must not drift).
• Possibly src/explore_persona_space/analysis/ for the (i) assertion.
• Keep the additions terse: this carve-out is already one of the longest bullets in CLAUDE.md, and the file is loaded on every session.

Constraints / invariants

• (iii) must NOT authorize an inline round to rewrite another task's promoted body unilaterally when the correction is a scientific-meaning change — filing is the safe default, and the promotion classification stays user-only. Word it as "apply the correction or file", with filing as the presumption for anything touching a Takeaway.
• (i) must not block a deliberate under-determined fit that the round justifies — refuse-by-default with a stated override, matching the project's fail-fast posture.
• scripts/workflow_lint.py --check-references / --check-asks pass; ruff passes.
• This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — recursion guard applies.

Related

The #825 body correction itself is NOT this task — it is a scientific-meaning change tracked separately as a daily-held needs-human task from the same sweep. This task adds the DUTY so the next round does not repeat the omission.

Provenance

• workflow_fix_target: CLAUDE.md
• fingerprint: 46ff8f476147
• Source: /daily 2026-07-25 transcript sweep, sessions dffde9b6 @ 22:20:13Z / 22:33:00Z and 63122023 @ 18:56:45Z.