issue-1310 body correction: per-turn instruct anti-prediction is a ridge under-regularization artifact, not a near-duplicate-context effect

Overview / Motivation

A user-directed 0-GPU inline estimator audit on #1310 (label nd-estimator-audit, 2026-07-30, commit a0ad4cea19) REFUTES the causal attribution in one of #1310's promoted ## Takeaways bullets and requires a prose correction to that promoted body. Filed per the inline-round record-integrity duty (CLAUDE.md § Routing, "Inline estimator-validity + record-integrity duties" item 3: filing is the presumption for anything touching a bolded Takeaway; a classification flip stays user-only).

Sibling: task #1887 hardens the shared fit-core selector DEFAULTS. This task is the RECORD half — correcting an already-promoted body — and touches no code.

Goal

Apply a prose correction to #1310's promoted body: the per-turn instruct "anti-prediction" it attributes to within-scene near-duplicate contexts is a ridge-lambda under-regularization artifact, and the dof cap the body credits as the mitigation is insufficient at per-turn row counts.

The refuted claim (verbatim from #1310's promoted ## Takeaways)

• Round 1's per-turn instruct anti-prediction (−0.10 to −0.19, swap inverted) was a within-scene near-duplicate-context fold artifact — real at the per-turn grain, not a property of the character map.

The same attribution carries the ### The per-turn instruct anti-prediction is confined to mid layers and traces to within-scene near-duplicate contexts result heading, and the ### Aggregating each prefill scene to one point turns every instruct cell positive and null-clearing in both models heading credits scene aggregation with the fix.

The refuting evidence

Audit artifacts (committed on main @ a0ad4cea19): eval_results/issue_1310/nd_estimator_audit/corrections_table.{json,md} (16 per-cell JSONs), figures figures/issue_1310/nd_audit_published_vs_corrected.png

• nd_audit_selector_spread.png, code scripts/issue1310_nd_estimator_audit.py + scripts/issue1310_nd_audit_report.py.

Substrate: the same persisted onpolicy prefill store, layer 19, the SAME fold assignment (fit825._cv_folds, seed 0, K=5) and the SAME Gram-space ridge the committed fits used. Every published capped-GCV value was reproduced EXACTLY before any other read (16/16 reproduction gates, |delta| = 0.0000 at 4 dp).

Per-turn instruct prefill cells at layer 19 — published capped-GCV vs the inner-group-CV selector (fit825.LAMBDA_SELECTION="inner-group-cv", 4 inner GROUP folds, the deployed-generalization criterion, no test-fold information):

cell	n	published	inner-group-CV	delta
onpolicy_instruct_Wren	1798	−0.1783	+0.2760	+0.454
onpolicy_instruct_HELIOS	1800	−0.0996	+0.3134	+0.413
onpolicy_instruct_Dana	1738	−0.1888	+0.2512	+0.440
onpolicy_instruct_Vex	1798	−0.1762	+0.2366	+0.413

All four sign-flip, at the PER-TURN grain, on the SAME folds, with the SAME within-scene near-duplicate contexts present. The anti-prediction therefore cannot be a property of the per-turn grain or of near-duplicate contexts.

Mechanism, pinned by the selected lambdas (grid np.logspace(-2, 4, 13)):

• Ambient pure-GCV pins lambda at the 0.01 grid FLOOR in 15/16 audited cells and reads −1.56 to −5.51 on the per-turn family (the #1345 degeneracy class, reproducing the range #1310 already quarantined at eval_results_onpolicy_gcvdegenerate/).
• The GCV_DOF_CAP = 0.9 mitigation rescues lambda only to 100 at per-turn row counts, because 0.9 × n_train ≈ 1,026 is a LOOSE constraint there.
• Inner-group-CV selects lambda 3,162–10,000.
• The published per-turn values equal the forced-lambda = 100 read EXACTLY (e.g. onpolicy_base_Wren +0.1140 published = +0.1140 at lambda 1e2), and the forced 1e3/1e4 reads (+0.18 to +0.37) bracket the inner-group-CV reads — so the published per-turn number IS "lambda = 100", and lambda = 100 is demonstrably ~30× too small.
• At the scene-aggregated row count (n = 300, n_train = 240) the same cap is BINDING (0.9 × 240 = 216), so it forces lambda to 3,162–10,000 — which coincides with inner-group-CV's principled choice. That is exactly why the aggregated family is unaffected.

So scene aggregation did not fix the instruct cells; adequate regularization did. Aggregation reached that regularization incidentally, by making the dof cap binding at the smaller n.

Scope — what is NOT refuted (do not over-correct)

• The scene-aggregated Takeaway ("base 0.13–0.22, instruct 0.23–0.40") is CONFIRMED: inner-group-CV reproduces all 8 aggregated cells to <= 0.0025.
• #1639 stands. Its within-cell ceilings ARE the aggregated cells, which the audit confirms robust; no correction is indicated there. (Named residual: the audit re-fit the within-cell R^2 only, not #1639's transfer-matrix / lattice / operator fits — those inherit the same n = 300 capped-lambda regime the audit found robust, but were not themselves re-selected.)
• The #1310 headline claim (a fixed-label fiction character supports a character-specific context→dialogue map) is NOT threatened. Every corrected read moved UP or stayed equal; the audit found no cell where a principled selector read LOWER than published.
• The run-2 SCRIPT-format cells (published 0.106–0.148 base / 0.188–0.253 instruct, ambient pure-GCV, no gcv_dof_cap field) are INDETERMINATE, not refuted: their activation store was lost with its instance, so they could not be recomputed at 0 GPU-h. Direction-of-error evidence says they are, if anything, understated.

Proposed change (prose only — no Takeaway classification flip)

Via task.py set-body on #1310, keeping runs.classification untouched (user-only):

1. Rewrite the third ## Takeaways bullet to attribute the per-turn instruct anti-prediction to ridge under-regularization at n_train < d rather than to within-scene near-duplicate contexts, citing the audit artifacts.
2. Fold a correction sentence into the ### The per-turn instruct anti-prediction ... result prose (v4 has no separate corrections heading), and note in the aggregation result that the mechanism is regularization reached incidentally through the smaller n, not the scene grain itself.
3. Extend the method-caveat bullet: the dof cap is insufficient at per-turn row counts (rescues lambda only to 100) and binding only at n = 300.
4. Add the audit artifacts + commit a0ad4cea19 to the **Repro:** footer.
5. Consider retitling the H1 only if the user judges the headline moved; the corrected reads strengthen rather than weaken it, so the default is to leave the title and confidence tag alone.

Whether the corrected per-turn values become the QUOTED per-turn numbers in #1310's body, or are reported as an audit addendum beside the published ones, is a judgement call for the plan — the audit's own arms are all layer-19 single-layer reads, not a full 28-layer re-fit with re-run nulls.

Constraints / invariants

• Body/prose only. Do NOT run task.py promote, do NOT flip runs.classification, do NOT edit ## Takeaways bullet COUNT out of the 3–6 spec band.
• verify_task_body.py --issue 1310 must PASS after the edit; set-body WITHOUT --snapshot (the body is already a clean-result and original-body.md exists).
• The audit's numbers are read from the committed corrections_table.json at a0ad4cea19 — re-grep them at compose time rather than re-typing from this body.
• No new compute: 0 GPU-h.