Drop7 Research
← Experiments

Stronger-teacher (D2 continuation) afterstate corpus: ranking gate and frozen override rule

completedtier CHECKalgorithmicpublic-developmentpublic-policyEX-20260821-afterstate-d2-teacher-58772f17
No plain-English explanation has been written for this experiment yet. Add web/content/research/EX-20260821-afterstate-d2-teacher-58772f17.mdx and it will appear here. The registered protocol is shown below.

The registered protocol

The iteration-3 valid negative used a weak phase-greedy D1 continuation for its H40 labels. Regenerating the successor-closed corpus with a stronger public teacher (fair D2, five chance samples) at K=64 over the same training roots produces labels whose trained model ranks fresh held-out siblings closer to fair D4 than the D1-teacher model did (top-1 gap to D4 shrinks by at least half), and the frozen override rule then passes on fresh origins.

Candidate
distributional-afterstate-ranker-d2
approaches/afterstate-learning/distributional-afterstate/common.hpp
Comparator
fair-d4
approaches/fair-expectimax/reference/fair-only-depth4.cpp

Primary metric

Fresh held-out top-1 accuracy vs fair D4 against the scenario-mean H40 D1-continuation target (K=256), compared with the D1-teacher model's iteration-3 result (0.4245 vs D4 0.4986).

Statistical unit: whole-origin

Pass criteria, fixed in advance

  • Corpus successor-closed at K=64 with the D2 teacher; completeness 1.0.
  • Ranking gate: model top-1 >= (0.4245 + 0.5 * (0.4986 - 0.4245)) = 0.4616 on each half-fold, i.e. the D1-teacher gap to D4 shrinks by at least half.
  • Override gate: eligible-root override regret <= D4 regret - 0.01 in EACH half-fold, with override rate >= 5%.
  • Decisive-root label stability >= 0.5 for both evaluation corpora (else inconclusive).
  • Quantile coverage within [0.66, 0.87].

On pass: Preregister a SCREEN gameplay experiment with the frozen override rule and the D2-teacher checkpoint.

On fail: Record a valid negative. If a stronger teacher does not shrink the gap, the label-quality hypothesis is weakened and the afterstate line is parked.

Data and reuse

Training: the same harvested training roots (seeds 0x5da70000-0x5da700ff) relabeled with the D2 teacher at K=64, H=40. Ranking gate: fresh seeds 0x5da70500-0x5da7053f (64 origins, D1-continuation K=256 target for cross-model comparability). Override gate: fresh seeds 0x5da70600-0x5da7063f (64 origins, same frozen rule and target as EX-20260820-d4-toptwo-override-gate-0bdb39a1).

seed leases: SL-20260820T083000Z-5da70000

What happened

valid run · outcome: failnot-supported-as-testedpilotRS-20260821T134500Z-4b9d2f68

The stronger-teacher hypothesis fails as tested. A successor-closed corpus relabeled with a fair-D2 (five-sample) continuation teacher at K=64 (2.88M rows over 6,535 roots; partial at the generator's 4h default wall stop) trained a model that, on the fresh 0x5da70500 ranking gate against the fixed D1-continuation H40 target, reached top-1 0.3365 - far below fair D4's 0.5020, below the D1-teacher model's 0.4245 from iteration 3, and only at exact-D1's own 0.3339. The frozen criterion (top-1 >= 0.4616 on each half, i.e. closing half the iteration-3 gap to D4) failed by a wide margin in both halves (0.342, 0.331). The frozen override gate on fresh 0x5da70600 roots also failed (eligible-root regret: half1 -0.0089, half2 +0.0110, pooled +0.0012). IMPORTANT CONFOUND, disclosed: the D2-teacher model was evaluated against D1-continuation outcomes (frozen for comparability with iteration 3), so part of its regression may reflect the teacher/target mismatch rather than teacher quality alone. Read narrowly, the result says a stronger-teacher corpus did not produce a better ranker of the fixed public-continuation outcome, and the afterstate line's ranking deficit is robust to the teacher choice within the tested configurations.

What it had to pass
  • Corpus successor-closed at K=64 with the D2 teacher, completeness 1.0 — observed: per-root completeness 1.0 over 6,535 fully labeled roots; the corpus is partial (6,535 of 8,192 planned) at the generator's 4h default wall stop
  • Ranking gate: model top-1 >= 0.4616 on each half-fold — observed: half1 0.3418, half2 0.3314
  • Override gate: eligible-root override regret <= D4 regret - 0.01 in EACH half-fold, rate >= 5% — observed: half1 -0.0089, half2 +0.0110, pooled +0.0012; override rate 35.9%
  • Decisive-root label stability >= 0.5 for both evaluation corpora — observed: 0.8058 (override corpus), 0.8010 (ranking corpus)
  • Quantile coverage within [0.66, 0.87] — observed: 0.7244
Recorded metrics
corpusD2Rows
2,884,224
corpusD2Roots
6,535
corpusD2Complete
false
epochsCompleted
30
rankingTop1Pooled
0.3365
rankingTop1Half1
0.3418
rankingTop1Half2
0.3314
d4Top1Pooled
0.5020
d1Top1Pooled
0.3339
iteration3D1TeacherTop1
0.4245
rankingPairwisePooled
0.6310
rankingRegretPooled
0.3120
overrideRegretGainHalf1
-0.0089
overrideRegretGainHalf2
0.0110
overrideRegretGainPooled
0.0012
overrideRate
0.3592
labelStabilityDecisiveSpearman
0.8058
quantileIntervalCoverage
0.7244
Limitations
  • Teacher/target mismatch confound: the model was trained on D2-continuation outcomes but evaluated against D1-continuation outcomes (frozen for comparability with iteration 3). A matched D2-continuation target was not generated (cost); the regression may partly reflect the mismatch.
  • The D2 corpus is partial (6,535 of 8,192 roots) at the generator's 4h default wall stop.
  • The compact 3.4M-param architecture and the D1-harvested root distribution are unchanged from iteration 3.
  • Single machine profile; FP32 on the shared-memory iGPU.