AFF_results — offline SN120 duel evals (Reason v4)

Offline duel evals of three 35B-A3B checkpoints against the reigning SN120 king, run on 2×H200 on 2026-08-25. Contains the full per-turn rollouts (thought + action) and raw forced-logprob pairs for every side, plus two ablations: cueswap300 (does the king's cue explain its edge?) and teacher_cue300 (how does the teacher itself score, and what does a cue do to it?).

Results

King: tojointhecommunity/affine-5efg6cm3yl-king@7c1c94cd (reign #41), scored reason = 0.017715 on this slice. Teacher: zai-org/GLM-4.5-Air-FP8. Artifact: chal-01103.

tag	model	n	reason	margin	SE	z	wins?
test7_full1290	lenikonate/qwen36-35b-a3b-2408-test7	1285	0.015278	−0.002500	0.001206	−2.07	❌
full1290	lenikonate/qwen36-35b-a3b-2408-test4	1287	0.015032	−0.002724	0.001151	−2.37	❌
qlora_full1290	vuhaian/qlora_35b_18k_3e	1288	0.014559	−0.003175	0.001130	−2.81	❌
qlora_300	vuhaian/qlora_35b_18k_3e	299	0.015448	−0.000875	0.002080	−0.42	❌ (underpowered)

None crowns. All three pass the two non-margin gates (median thought ≥ 80 chars, B-gate rate ≥ 0.30) and lose purely on margin.

qlora_300 is a cautionary datapoint: on the first 300 turns qlora looked near-parity (z = −0.42). At the full 1290 the same checkpoint is the worst of the three (z = −2.81). At n=300 the SE is ~1.8× the full run's — do not read a direction from |z| < 2.

The shape of the loss

All three models show the same pattern: positive median turn-diff, negative mean, and they win a majority of individual turns while losing overall.

	test4	test7	qlora
exact teacher action	32.2%	30.0%	33.2%
command head	86.6%	89.3%	86.8%
king exact / head	37.6% / 78.9%	37.7% / 79.1%	37.6% / 79.0%
turns won	721/1287	718/1285	736/1288
diff, matched a ref	+0.008567	+0.010800	+0.007838
diff, no ref match	−0.008098	−0.008211	−0.008636

Every model picks the right command more often than the king (83–89% vs ~79%) but emits the teacher's exact action less often (30–33% vs ~38%). Under v4's tempered log-mean-exp the turn score is dominated by the best-matched reference, so missing all k=3 modes leaves no hit to recover from. This is a distillation-alignment deficit, not a coding-ability deficit — do not translate it into a benchmark-capability claim.

Cue-swap ablation — is the king's edge the cue or the thought?

Every side ends its thought with a near-deterministic cue: the king's at 99.5% of turns, the miners' at 99.8-100%.

king cue : "I could either verify the current state or proceed with the change;
            the usual choice is to check first.\nThe analysis is complete. Next command:"
miner cue: "The analysis is complete. Output the next bash command only,
            with no explanation. Next bash command:"

cueswap300 re-scores already-generated rollouts on 300 random paired turns (seed 1103): each side's own trailing cue is stripped and a common cue appended as body.rstrip() + "\n" + CUE. Only the teacher is re-run — no student engines, and lp_empty is reused since it does not depend on the thought.

Within-side cue effect (same model, same turn, only the cue changes — the pairing cancels most variance, so this is the load-bearing result):

side	Reason(king cue) − Reason(miner cue)	SE	z
king	+0.000097	0.000806	+0.12
test4	−0.000221	0.000804	−0.27
test7	+0.001408	0.000984	+1.43
qlora	+0.002404	0.001076	+2.23

The king does not win because of its cue. Dress it in the miner cue and its Reason is statistically unchanged (z = +0.12). The asymmetry runs the other way: the miner cue is a liability for test7 and qlora, which score meaningfully better wearing the king's. Equalising on the king's cue closes ~70% of their gap — because the challengers rise (test7 0.01452→0.01657, qlora 0.01284→0.01595), not because the king falls (0.01712→0.01733).

Paired margin vs king under a common cue:

model	own cues	both king cue	both miner cue
test4	−0.001802	−0.001980	−0.001662
test7	−0.002603	−0.000764	−0.002076
qlora	−0.004282	−0.001383	−0.003690

So rewriting the cue is a real but partial win: best case ~+0.0024 against the ~+0.0048 needed to clear the crown bar on the full slice. The remainder is exact-action mode alignment, not cue phrasing.

Limits. At n=300 every unpaired margin above has |z| < 1.1 and is not a result on its own; only the paired within-side effects clear the noise. Equalising cues also flips the median turn-diff negative (+0.0008 → ~−0.0003) and drops turns-won from ~175/300 to ~130–142/300 — the cue was giving the challengers a broad small edge on typical turns. A pipeline check recomputed 240 stored echo values: mean Δ = −3.4e−04 (unbiased), mean |Δ| = 2.8e−03, i.e. teacher echo nondeterminism is comparable to these effects per call and averages down only across turns. B-gate echoes were computed but are not reported.

Files: results/cueswap300_report.log, results/cueswap300.json, results/cueswap300_cells.json (per turn × side × condition).

Teacher-as-challenger — what does distilling teacher thoughts actually buy?

teacher_cue300 hosts the teacher itself (zai-org/GLM-4.5-Air-FP8, 1×H200) and scores its own thoughts as if they were a miner's, on the same 300-turn slice and seed (1103) as cueswap300. 4,500 fresh forced-logprob echoes, 0 errors. Two questions: what score is a teacher-distillation run training toward, and does a trailing cue make the teacher more confident?

The self-reference trap — read before quoting any teacher number

The artifact's teacher_refs[i].lp_own is the diagonal only: reference i scored under reference i's own thought. Valid v4 scoring needs thought j against all k references. Leave the teacher's own generating reference in and one a_i becomes self-prediction, which then dominates the tempered log-mean-exp. A live duel resamples teacher refs independently of the miner, so leave-one-out (LOO) is the honest regime; full-k is inflated.

side	full k=3	LOO (honest)
teacher, no cue	0.033165	0.005460
teacher + miner cue	0.030258	0.012816
teacher + king cue	0.029115	0.013901
test7	0.014697	—
king	0.013471	—

Read full-k and the teacher "wins" at z = +6.7. Drop the self-scored reference and its no-cue score collapses to 0.005460 — below both test7 and the king, losing to test7 by −0.009270.

Q1 — distilling teacher thoughts is legal, but it is not a ceiling worth chasing

Scored honestly, the teacher's own thoughts earn 0.005460, worse than test7 (0.014697) and the king (0.013471) on this slice. Overfitting to them perfectly would move a checkpoint backwards. The gates are fine (median stripped thought 196 ≥ 80; B-gate 0.427 ≥ 0.30) — it is simply that the teacher's thoughts are optimised to produce the teacher's own next action, not to make a held-out sample of teacher actions predictable, and v4 scores only the latter.

Q2 — both cues make the teacher substantially more confident

Paired, teacher vs itself, same turns, LOO:

cue appended	Δ Reason	SE	z	improved
miner cue	+0.007355	0.001965	+3.74	136/299
king cue	+0.008440	0.002338	+3.61	123/299

Both exceed the entire test7-vs-king margin (−0.002500 on the full 1290). The king's cue is nominally stronger; the two overlap within noise.

Mechanism. The cue barely moves the best a_i (0.04570 → 0.04103 / 0.03953) but nearly triples the plain mean (0.00483 → 0.01416 / 0.01546). It is not improving the top match — it lifts the references that would otherwise go negative, exactly the multi-modality failure affine.toml [duel] describes. Under a tempered log-mean-exp that floor-raising is worth a lot.

Sign flip. On full-k the same cue looks harmful (−0.0029, −0.0041): the self-scored reference already saturates the LME, so the cue can only dilute it. Measuring this from the published artifact alone gets the direction backwards.

Relation to cueswap300

cueswap300 asked whether the king's edge comes from its cue (it does not, z = +0.12). This asks how the teacher scores and what a cue does to it, on the same slice and seed. Complementary, and they agree in direction for the challengers: the king's cue is worth slightly more, but cue phrasing is not what decides the crown. test7's residual deficit is exact-action mode alignment (30.0% vs king 31.7% exact, while leading on command head 89.0% vs 76.3%).

Files: results/teacher_cue300_report.log, results/teacher_cue300.json, results/teacher_cue300_cells.json (per turn × condition, incl. per-ref a_i), results/teacher_cue300_rows.jsonl (raw echoes), scripts/teacher_cue300_*.py.

Files

Per tag <T>:

file	contents
results/<T>_eval.log	human-readable verdict + telemetry
results/<T>.json	verdict, contract knobs, telemetry, v3/v4 regime check
results/<T>_rows.json	challenger rollouts — per turn, k pairs with z_a, y_a, lpC_yc_za, lpC_yc_e, lpC_ya_za, lpC_ya_e, lpC_yc_zc, eta
results/<T>_king_rows.json	king rollouts, same schema (lifted from the artifact)
results/<T>_diagnostics.json	per-turn mode-alignment breakdown
artifact/turns.json	raw duel artifact: king pairs + teacher_refs per turn
artifact/series.json	duel slice metadata + per-turn aggregates
results/cueswap300*	cue-swap ablation: report, summary, per-cell scores
results/teacher_cue300*	teacher-as-challenger + cue ablation: report, summary, per-cell, raw echoes
scripts/	the runner, teacher-forcing client, analysis, and bias check used for teacher_cue300

Method

Reason v4, weight_version_key = 9: a_i = lpC(y_i|z_A) − lpC(y_i|∅) per teacher reference, Reason = tau·log((1/k)·Σ exp(a_i/tau)) with tau = 0.03, k = 3; miner score is the mean over turns. Crown needs margin > max(2·SE, 0.002) and median stripped thought ≥ 80 chars and B-gate pass rate ≥ 0.30.

The king is not re-run — its per-turn rows and the teacher's reference rollouts are lifted verbatim from the published duel artifact, so both sides score on identical turns against identical references. Scoring uses the production modules (affine.score, evalsrv.terms.miner_terms) directly.

King rows are byte-identical across all four runs (*_king_rows.json for the 1290-turn tags are the same 6,347,516 bytes; king reason 0.017715 in each), so the models are directly comparable to each other.

Caveats — please carry these

1. Teacher references are public (they ship in the artifact). A live duel resamples them fresh, so a checkpoint trained on published refs scores optimistically here. Development signal, not a live verdict.
2. The slice is a subsample of another duel's slice, seeded by that challenger's hotkey, not any of these. Smaller n ⇒ larger SE than live.
3. Both sides are rescored under v4. Published artifacts report per-turn reason as the v3 plain mean, not the v4 tempered log-mean-exp, despite advertising tau: 0.03. Verified here: the published per-turn mean reproduces the v3 recompute exactly (0.009718358326047678), while the v4 recompute is 0.017715 — a +0.008 shift, far larger than any margin. These numbers are therefore NOT comparable to the artifact's published score_king / margin.
4. The teacher-bias check (EVAL.md §12) has now been run (2026-08-25), against the teacher_cue300 local teacher. Fresh echoes vs the artifact's published values, n=36 each: lpC(y|z_own) mean −8.4e−05 (SE 6.0e−04), lpC(y|∅) mean +1.4e−03 (SE 1.9e−03), lpC(y_i|z_K) mean −9.5e−04 (SE 8.9e−04) — reproducing the historical reference (mean −8.9e−05, SE 3.0e−04, n=448) within noise. This licenses mixing fresh echoes with the artifact's stored rows. Per-call stdev ~3.6e−03 averages down across turns. (Applies to the teacher_cue300 teacher; the four duel runs above were scored before this check existed, but use the same pinned vLLM 0.22.1.)
5. A handful of turns per run (1–5 of 1290) produced no parseable action from the challenger and are excluded; this is model-side, not infrastructure.