Drop7 Research
← Experiments

Offline gate: calibrated top-two near-tie override of fair D4 by the frozen afterstate model

completedtier PILOTalgorithmicpublic-developmentpublic-policyEX-20260820-d4-toptwo-override-gate-0bdb39a1

The third pilot established that the learned board-scorer cannot replace the four-move search as the thing that picks the move. This experiment asks a much smaller question: could it still be useful in the one situation where the search is least sure of itself?

Often fair D4 rates its top two columns almost identically. Those near-ties are where a four-move estimate is most at the mercy of which futures it happened to sample. The candidate here is unchanged fair D4 with one conservative addition: at a near-tie — the top two within 500 points, a threshold fixed from already-read data before this experiment existed — consult the frozen third-iteration network, and swap to the search's second choice only if the network's advantage survives a resampling test. The comparator is unchanged fair D4. Nothing else about either player differs, and again no games were played: both were graded on how well they ordered columns at fresh held-out positions.

"Pass" required the override to cut average regret by at least 0.01 in each half of the held-out set, to fire on at least 5% of eligible positions, and to reproduce byte-identically across two runs.

What happened. A valid run with a fail outcome — but a narrow one, and the record says so explicitly. The rule fired often (38.3% of positions were near-tied; the override triggered on 37.0% of those), it was deterministic, and it improved regret in both halves: by 0.0351 in one and 0.0075 in the other. The frozen bar was 0.01 in each, so the second half missed by a quarter of a percentage point and the gate fails. The record's own reading is that the mechanism works as designed and that this is the first preregistered held-out test here where a learned model's intervention improved on fair D4 at all, with the checkpoint's ranking quality as the bottleneck. It also warns, in the limitations, that the miss "should be read as a narrow failure, not as evidence of no effect."

resultvalid · failpilotRS-20260820T184500Z-63c0a8e2

The run was valid; the outcome was fail, at the a pilot — a small run to find bugs and project cost, not a strength claim level. Of 7 preregistered checks, 6 passed and 1 failed.

The calibrated top-two near-tie override of fair D4 by the frozen iteration-3 afterstate model FAILED its frozen gate by a narrow margin, with genuinely positive signal. On 2,689 fresh held-out roots, 1,030 (38.3%) were near-tied at the frozen 500-point threshold and the override fired on 37.0% of those. On eligible roots the override reduced mean normalized regret versus unchanged D4 in both half-folds (half1 0.2484 vs 0.2835, +0.0351; half2 0.2343 vs 0.2418, +0.0075; pooled +0.0214), but the frozen criterion required at least +0.01 in EACH half-fold, and half2 fell short. On the 84 decisive eligible roots the improvement was larger (+0.0426: 0.0880 vs 0.1306). Label stability passed (decisive Spearman 0.814), model quantile calibration passed (0.861), the gate is deterministic (byte-identical repeated reports), and the corpus is successor-closed. Interpretation: the conservative-override MECHANISM works as designed (fires often, calibrated, positive in both halves) and this is the first preregistered held-out test in this repository where a learned model's intervention improved on fair D4 at all; the CHECKPOINT's ranking quality is the bottleneck, consistent with the model being undertrained (11 of 20 epochs) and trained on a weak D1-continuation teacher. Per the frozen failure action, no further override variants of this same checkpoint are opened; a retry requires new evidence such as a fully trained or stronger-teacher model under the same frozen rule.

Limitations recorded with the result
  • The evaluation target is the H40 phase-greedy-D1 continuation outcome - the same label family the model was trained to predict; D4 is compared against the same target, but the offline gate cannot measure real gameplay interaction, which only a SCREEN tier can.
  • The model was undertrained (11 of 20 epochs) and used a weak D1-continuation teacher; both are documented as the likely bottleneck and motivate any retry.
  • The near-tie threshold (500) and regret margin (0.01) were frozen choices; the half2 miss (0.0075) is close to the margin and the result should be read as a narrow failure, not as evidence of no effect.
  • Single machine profile; FP32 on the shared-memory iGPU.

Full record →

What it closes: no further override variants of this checkpoint without new evidence. What it leaves open: the same frozen rule applied to a fully trained model, or one taught by a stronger continuation player. And what it cannot speak to at all is real play — an offline ranking test cannot measure how these choices interact over a whole game, which only a paired gameplay tier could.

The registered protocol

At roots where fair D4's top two actions are near-tied (Q gap <= 500 points, the bottom ~37% of the top-two gap distribution measured on already-read development labels), the frozen iteration-3 distributional afterstate model can identify the better action more reliably than D4's own ordering, because near-ties are where D4's 4-ply estimates are most noise-sensitive while the model's calibrated H40 outcome distributions carry independent long-horizon information. The override policy (D4, plus swapping to the second action only when the model's paired 256-scenario advantage has a positive 95% bootstrap lower bound) will show lower mean normalized regret than unchanged D4 on eligible fresh held-out roots.

Candidate
d4-toptwo-override
approaches/afterstate-learning/distributional-afterstate/override-gate.py
Comparator
fair-d4
approaches/fair-expectimax/reference/fair-only-depth4.cpp

Primary metric

Mean normalized regret of the override policy versus unchanged fair D4 on override-eligible fresh held-out roots, against the scenario-mean H40 outcome ranking (K=256, phase-greedy D1 continuation), reported per origin-hash half-fold.

Statistical unit: whole-origin

Pass criteria, fixed in advance

  • Corpus is successor-closed: every non-trivial harvested root has labels for every legal sibling under all 256 scenarios.
  • Label stability floor: decisive-root scenario-half Spearman >= 0.5; below this the outcome is inconclusive.
  • On eligible roots, override-policy mean normalized regret <= D4 mean normalized regret - 0.01 in EACH half-fold.
  • Override rate >= 5% of eligible roots (a rule that never fires proves nothing).
  • Nominal 76.5% quantile interval coverage of the frozen model on fresh afterstates within [0.66, 0.87] (sanity that the model is in-support).
  • The gate script is deterministic: two runs produce byte-identical reports.
  • All existing CHECK-tier self-tests pass before any label is inspected.

On pass: Preregister a SCREEN gameplay experiment: 32 paired public-development games, native fair D4 with the frozen override rule (model inference over stratified afterstate samples at near-tied roots only) versus unchanged fair D4, with per-game result rows and the standard cohort summary.

On fail: Record a valid negative (or inconclusive) result. A failure here closes the direct-override use of this model: no further override variants of the same checkpoint without new evidence.

Data and reuse

The model is the frozen iteration-3 checkpoint (trained only on seeds 0x5da70000-0x5da700ff, training-role). The near-tie threshold 500 was fixed from already-read iteration-2/3 comparator labels before this experiment's data existed. Evaluation roots come from fresh seeds 0x5da70300-0x5da7033f (64 origins), read for the first time by this experiment's single gate evaluation. No protected or final data is read.

seed leases: SL-20260820T083000Z-5da70000

What happened

valid run · outcome: failnot-supported-as-testedpilotRS-20260820T184500Z-63c0a8e2

The calibrated top-two near-tie override of fair D4 by the frozen iteration-3 afterstate model FAILED its frozen gate by a narrow margin, with genuinely positive signal. On 2,689 fresh held-out roots, 1,030 (38.3%) were near-tied at the frozen 500-point threshold and the override fired on 37.0% of those. On eligible roots the override reduced mean normalized regret versus unchanged D4 in both half-folds (half1 0.2484 vs 0.2835, +0.0351; half2 0.2343 vs 0.2418, +0.0075; pooled +0.0214), but the frozen criterion required at least +0.01 in EACH half-fold, and half2 fell short. On the 84 decisive eligible roots the improvement was larger (+0.0426: 0.0880 vs 0.1306). Label stability passed (decisive Spearman 0.814), model quantile calibration passed (0.861), the gate is deterministic (byte-identical repeated reports), and the corpus is successor-closed. Interpretation: the conservative-override MECHANISM works as designed (fires often, calibrated, positive in both halves) and this is the first preregistered held-out test in this repository where a learned model's intervention improved on fair D4 at all; the CHECKPOINT's ranking quality is the bottleneck, consistent with the model being undertrained (11 of 20 epochs) and trained on a weak D1-continuation teacher. Per the frozen failure action, no further override variants of this same checkpoint are opened; a retry requires new evidence such as a fully trained or stronger-teacher model under the same frozen rule.

What it had to pass
  • Successor-closed corpus (all legal siblings x 256 scenarios) — observed: 4,858,880 rows over 2,793 roots; completeness 1.0
  • Label stability floor: decisive-root scenario-half Spearman >= 0.5 — observed: 0.8144
  • Eligible-root override regret <= D4 regret - 0.01 in EACH half-fold — observed: half1 +0.0351 (passes); half2 +0.0075 (below the frozen 0.01 margin); pooled +0.0214
  • Override rate >= 5% of eligible roots — observed: 37.0%
  • Quantile interval coverage within [0.66, 0.87] — observed: 0.8606
  • Gate script deterministic (byte-identical repeated reports) — observed: two runs byte-identical after moving wall time out of the report
  • All existing CHECK-tier self-tests pass before any label is inspected — observed: SELFTEST OK (9 checks)
Recorded metrics
roots
2,689
eligibleRoots
1,030
nearTieRate
0.3830
overrideRateEligible
0.3699
d4RegretEligiblePooled
0.2629
overrideRegretEligiblePooled
0.2414
regretGainHalf1
0.0351
regretGainHalf2
0.0075
regretGainDecisive
0.0426
d4RegretWholeSet
0.1932
overrideRegretWholeSet
0.1850
labelStabilityDecisiveSpearman
0.8144
quantileIntervalCoverage
0.8606
Limitations
  • The evaluation target is the H40 phase-greedy-D1 continuation outcome - the same label family the model was trained to predict; D4 is compared against the same target, but the offline gate cannot measure real gameplay interaction, which only a SCREEN tier can.
  • The model was undertrained (11 of 20 epochs) and used a weak D1-continuation teacher; both are documented as the likely bottleneck and motivate any retry.
  • The near-tie threshold (500) and regret margin (0.01) were frozen choices; the half2 miss (0.0075) is close to the margin and the result should be read as a narrow failure, not as evidence of no effect.
  • Single machine profile; FP32 on the shared-memory iGPU.