Rung (c) results: masked-contrast ANTI-SELECTS — the zero-training scorer family closes

2026-08-09 ~10:2xZ. Pre-reg: 2026-08-09-prereg-subgoal-mcselect.md (finalized 09:2xZ, run launched 09:12:36Z, complete 10:20Z — ~1.1 GPU-h of the 4.0 gate). Frozen read: mcselect_results.py → reports/analysis__subgoal_mcselect_q4_ar100k_k4l2.json. Everything below is that script’s output; the argmax, tie rule and CI machinery were landed and oracle-gated BEFORE the run.

Plain words. The model can write several candidate “next step” notes to itself, and we know (from rung (b′)) that the sampled set usually contains a genuinely better note than the one it writes by default — but the first scorer we tried (pick the note the model is most confident about) chooses badly. This rung tried the opposite philosophy, imported from MG-Select: pick the note that most changes what the model wants to do (measured against a “no-note-at-all” reference). Verdict: that scorer chooses badly too — in fact worse. Picking maximally-informative notes drags the policy toward disruptive, phase-wrong instructions. Two strikes means this whole family of free scorers is now closed for this model; anything further has to learn what a good note looks like.

The frozen reads

Agreement diagnostics: the scorer is decidedly not inert — the mc pick differs from the greedy text on 66.1% of rows (inert-guard bar was 5%; 244 tie rows) — yet it agrees with the oracle pick on only 14.4% of labeled rows and with SC’s pick on 14.5%: chance-level against the oracle at a 9-candidate width. Execution oracles all green pre-read (τ/sha echo, KL-finiteness == eligibility, state-copy byte-match on 4,301 joined rows; pred_masked-vs-banked flip count 1207/4301 = exactly the amendment-1 composition-noise figure, a free instrument confirmation).

What this means

Informativeness is anti-correlated with quality here, not merely uncorrelated. KL(p_cond ‖ p_masked^{1/τ}) rewards the candidate that most bends the action distribution away from the planner-less default. On our trunk the biggest benders are disruptive instructions — phase-wrong, over-specific, or off-task text the policy obediently follows — and the damage lands hardest exactly in the late-horizon slot where real subgoal signal lives (+0.385 last-10%, the mirror image of the ceiling’s −0.464). Self-certainty failed by rewarding plausibility; masked-contrast fails by rewarding impact. The better candidates the width provably holds (ceiling −0.250 vs bare) sit between those poles: right-phase text that shifts the policy moderately, which neither free axis ranks first.

The pre-registered kill rule executes: the zero-training scorer family CLOSES for this trunk. Two independent scorer philosophies (likelihood-flavored, contrast-flavored) both anti-select on the same frozen width with the same eligibility rules. Per the pre-reg, learned-verifier shapes (RoVer, Q-guided) now need their own affirmative case — cost, labels, and a reason to believe a learned critic escapes both failure axes — not a routing inheritance from this ladder. Candidate 2 (TOPReward-shape history phase estimation) was pre-named as the escalation only for a flat-late-horizon mc failure; the observed failure is active anti-selection, not phase-blindness, so it does not auto-open either.

What stays alive: the ceiling itself. Rung (b′)’s finding is unchanged — the sampled width holds genuinely better subgoals (−0.250 [−0.353, −0.148] vs bare). The gap is now firmly a scorer gap, measured against two closed attempts. Free exploratory follow-up banked as a queue item: a record-only post-mortem on the banked [N,C] KL + [N,C,S,D] per-candidate error dump — per-candidate KL-vs-quality correlation and the oracle pick’s KL-rank histogram — to see WHERE on the informativeness axis the good candidates actually sit before anyone prices a learned verifier.

Cost

Run ~1.1 GPU-h (68 f/min steady over 4,301 q4 rows × ~10 decodes + 9 reference forwards each) vs the 4.0 gate and the 2–2.5 projection — the capture-during-decode design (no logit re-forward) is what kept the conditioned side at single-decode price.

Addendum 2026-08-09 ~11:0xZ — the post-mortem map (record-only)

The queued post-mortem (mcselect_postmortem.py → reports/analysis__subgoal_mcselect_postmortem_q4_ar100k_k4l2.json) ran same-day on the banked dump. NOT pre-registered, no decision rides on it — it is the map to read before anyone prices a learned verifier. Oracle-gated like everything else (planted monotone fixture with exact hand arithmetic, tie/degenerate/contract abort branches). “Oracle-best” below = the eligible candidate with the lowest realized frame MAE — the frame-error oracle over all 4,301 rows, not the labeled ceiling pick.

The section above needs one correction of emphasis. The frozen read’s story was “informativeness is anti-correlated with quality.” At the per-row rank level that is not what the map shows — the truth is stranger and more useful:

KL is rank-noise, not a reversed compass. Within a row, informativeness carries zero rank information about candidate quality (Spearman +0.012, CI straddling 0; the good candidate sits uniformly on the axis, 0.498 vs the 0.5 null). You cannot fix MC by flipping the argmax to an argmin — the bottom of the KL axis is as uninformed as the top (bottom-1 excess 16.8% vs 12.6% null, the mirror of the top’s 17.4%: mild excess at BOTH extremes, i.e. the best candidate is merely somewhat atypical). The +0.313 harm is magnitude-driven, not order-driven: the value-level row-centered Pearson is +0.126 (vs the rank-level ~0), so when the argmax rides the far KL tail it occasionally lands on catastrophically wrong text, and those rows carry the read. That is the winner’s curse at an extreme of a noisy axis, amplified by heavy-tailed damage — consistent with the late-horizon +0.385 fingerprint.

SC was the better axis all along — ~6× too weak to matter. Its per-row signal is real and right-signed (−0.030, CI entirely below 0; the oracle-best is at SC-rank-0 in 30.1% of rows vs 12.6% chance) — and it still lost by +0.210 as a selector. A rank signal of |rho| ≈ 0.03 over ~8 candidates simply cannot survive an argmax. That is the calibration number this map exists for: the banked ceiling (−0.250 vs bare) is real, and the two free axes offer |rho| ≤ 0.03 toward it. A learned verifier has to clear a per-row rank signal perhaps an order of magnitude stronger than anything zero-training gave us — that is the affirmative case it must make before its GPU-hours are priced.

The family failed twice independently, not once twice. The two axes are essentially uncorrelated with each other (per-row Spearman(KL, SC) +0.032 [+0.017, +0.047]; the banked pick agreement was already chance-level at 14.5%). “Zero-training scorer family” was the right closure unit: likelihood-flavored and contrast-flavored signals are different measurements, and each is separately too weak. One curiosity for the record: MC’s pick is the frame-error-best candidate slightly MORE often than SC’s (25.6% vs 23.6%) while losing harder on MAE — more evidence that MC’s failure lives in the magnitude of its misses, not their count. (153 rows had a constant KL across all eligible candidates and are excluded from the rho reads; eligible width spans 4–9, median 8.)