Pilot iteration 2: K=64-scenario action-complete H40 corpus with fresh held-out origins
The first iteration could not grade anything, because the answer key it graded against was too noisy to reproduce itself. This experiment changed exactly one thing: instead of averaging each column's future over 8 shared runs of luck, average it over 64. Everything else — the network, the positions, the 40-move continuation, the thresholds — was held fixed, and the held-out positions came from a fresh block of game seeds that had never been read.
Same two contestants as before. The candidate is the learned board-scorer that reads only a resolved board, the visible next disc, and the rise clock; the comparator is exact fair D4, the four-move search that is this repository's reference. Neither played a game: this is an offline test of who orders the columns better at a position. "Pass" required the split-half agreement of the answer key to reach 0.50, and then the network to come within two percentage points of the search on top-choice and head-to-head accuracy, in both halves of the held-out set.
What happened. More luck samples helped, in the direction and roughly the size the noise arithmetic predicted — split-half agreement rose from 0.246 to 0.446 — but 0.446 is still under the frozen 0.50 floor, so the verdict is inconclusive for a second time. The network's confidence intervals were well calibrated this time (0.807 against a target band), and it clearly beat the weak one-move player whose games produced its labels, but it trailed the four-move search in both halves. The record also flags that training stopped at epoch 15 of 20 when the GPU budget ran out, with the ranking loss still falling. No gameplay tier was opened.
The run was valid; the outcome was inconclusive, at the a pilot — a small run to find bugs and project cost, not a strength claim level. Of 7 preregistered checks, 3 passed and 4 failed.
K=64 improved label stability from 0.246 to 0.446 mean split-half Spearman on fresh held-out origins, but still below the frozen 0.5 floor, so the preregistered verdict is again inconclusive. Quantile calibration passed (0.807 coverage of the nominal 76.5% interval). Against the still-noisy target the model trailed fair D4 consistently in both half-folds (pooled top-1 0.342 vs 0.433; pairwise 0.637 vs 0.697; regret 0.303 vs 0.232) while beating its D1 continuation teacher (top-1 0.342 vs 0.299). Training stopped at epoch 15 of 20 on the GPU time budget with the ranking loss still decreasing. No gameplay tier was opened.
Limitations recorded with the result
- Labels remain H40 returns under a phase-greedy D1 continuation (a weak fixed teacher).
- Model undertrained: 15 of 20 epochs at the GPU budget stop.
- Iteration-1 held-out roots were folded into training data here; the gate read only fresh origins 0x5da70100-0x5da7013f.
- Single machine profile; FP32 on the shared-memory iGPU.
Nothing about the theory was ruled in or out. The useful output is that the noise model held: sample count moved the key's reproducibility about as predicted, which is what justified paying for a much larger third iteration.
The registered protocol
Raising aligned chance scenarios from K=8 to K=64 (the only change from EX-20260820-afterstate-pilot-h40-29b8588a) lifts scenario-half label stability above the frozen 0.5 Spearman floor, and the unchanged distributional afterstate ranker then ranks fresh held-out whole-origin siblings non-inferiorly to exact fair D4 on top-1 accuracy, pairwise accuracy, and normalized regret against the scenario-mean H40 outcome ranking.
approaches/afterstate-learning/distributional-afterstate/common.hppapproaches/fair-expectimax/reference/fair-only-depth4.cppPrimary metric
Fresh held-out whole-origin top-1 action accuracy of the model's scenario-averaged ranking against the scenario-mean H40 outcome ranking, compared with exact fair D4's top-1 accuracy on the same roots.
Statistical unit: whole-origin
Pass criteria, fixed in advance
- Corpus is successor-closed: every non-trivial harvested root (>=2 legal actions) has labels for every legal sibling under all 64 scenarios.
- Label stability floor: mean per-root Spearman correlation between scenario-half (32 vs 32) action rankings >= 0.5 on held-out roots; below this the outcome is inconclusive (labels too noisy), not a pass or a theory rejection.
- Model top-1 accuracy >= fair D4 top-1 accuracy - 0.02 on each held-out half-fold.
- Model pairwise accuracy >= fair D4 pairwise accuracy - 0.02 on each held-out half-fold.
- Model normalized regret <= fair D4 normalized regret + 0.02 on each held-out half-fold.
- Nominal 76.5% quantile interval coverage on held-out afterstates within [0.66, 0.87].
- All CHECK-tier mechanics, legality, determinism, reflection, information-boundary, and resource-bound tests pass before any label is inspected.
On pass: Register a successor SCREEN experiment: 32 paired public-development games of the frozen 1-ply chance-averaged afterstate policy with exact fair-D4 fallback versus unchanged fair D4. The candidate, corpus, and thresholds are not retuned on the held-out fold first.
On fail: Record a valid negative (or inconclusive, if the stability floor fails again) result for this exact configuration. If stability passes but the model trails D4, the theory is not supported as tested; keep both corpus manifests for diagnosis and open no gameplay tier.
Data and reuse
Training corpus: seeds 0x5da70000-0x5da700ff (the same 256 origins as iteration 1, including its former held-out roots, which are now training-only and are never evaluated again). Evaluation corpus: fresh seeds 0x5da70100-0x5da7013f (64 origins), all held-out, read for the first time by this experiment's single gate evaluation. Iteration 1's gate outcome was inconclusive on label stability; this experiment changes only the scenario count and the held-out origins. No protected or final data is read.
seed leases: SL-20260820T083000Z-5da70000
What happened
K=64 improved label stability from 0.246 to 0.446 mean split-half Spearman on fresh held-out origins, but still below the frozen 0.5 floor, so the preregistered verdict is again inconclusive. Quantile calibration passed (0.807 coverage of the nominal 76.5% interval). Against the still-noisy target the model trailed fair D4 consistently in both half-folds (pooled top-1 0.342 vs 0.433; pairwise 0.637 vs 0.697; regret 0.303 vs 0.232) while beating its D1 continuation teacher (top-1 0.342 vs 0.299). Training stopped at epoch 15 of 20 on the GPU time budget with the ranking loss still decreasing. No gameplay tier was opened.
- ✓Successor-closed corpus (all legal siblings x 64 scenarios) — observed: 6,167,936 rows over 14,228 roots across both corpora; completeness 1.0
- ✕Label stability floor: mean scenario-half (32 vs 32) Spearman >= 0.5 — observed: 0.4462 on 2,750 fresh held-out roots; inconclusive per the frozen rule
- ✕Model top-1 >= D4 top-1 - 0.02 on each half-fold — observed: half1 0.3401 vs 0.4365; half2 0.3442 vs 0.4299 (moot given the stability failure)
- ✕Model pairwise >= D4 pairwise - 0.02 on each half-fold — observed: half1 0.6341 vs 0.6970; half2 0.6391 vs 0.6974 (moot)
- ✕Model regret <= D4 regret + 0.02 on each half-fold — observed: half1 0.3031 vs 0.2388; half2 0.3028 vs 0.2247 (moot)
- ✓Nominal 76.5% quantile interval coverage within [0.66, 0.87] — observed: 0.8069
- ✓All CHECK-tier tests pass before any label is inspected — observed: SELFTEST OK (9 checks) after the K-parameterization change
Recorded metrics
- Labels remain H40 returns under a phase-greedy D1 continuation (a weak fixed teacher).
- Model undertrained: 15 of 20 epochs at the GPU budget stop.
- Iteration-1 held-out roots were folded into training data here; the gate read only fresh origins 0x5da70100-0x5da7013f.
- Single machine profile; FP32 on the shared-memory iGPU.