Drop7 Research
← Experiments

Stage D0: fair relabelling of oracle-visited vs matched fair-D4 public states

completedtier CHECKdiagnosticno-gameplayprivileged-teacherEX-20260823-hpool-stage-d0-e0ad1c65
No explanation has been written for this experiment yet. Add web/content/research/EX-20260823-hpool-stage-d0-e0ad1c65.mdx and it will appear here. The registered protocol is shown below.

The registered protocol

Generate 64 perfect-information-oracle games (approaches/oracle-curriculum/perfect-information-oracle semantics as implemented in approaches/oracle-curriculum/state-curriculum/oracle-curriculum.cpp: depth 4, beam 128, 500-move cap, corrected 17,000-point Hardcore scoring) on the fresh training lease 0xa52e0000-0xa52e003f and sample public states from move 50 onward to form pool O (target 2,000 states, capped by availability). Generate fair-D4 (reference, five strata) games on 0xa52e0100-0xa52e01ff and sample public states from move 50 onward to form pool F, matched 1:1 to O on rise phase, occupancy and maximum height (exact-match buckets as in approaches/oracle-curriculum/topology/oracle-topology-audit.cpp; unmatched O states are dropped and counted). Every state is reduced to the public tuple (board, next disc, moves-until-rise) before any label is computed. For each state compute R_fair = mean remaining moves over K=32 independent public futures derived from the public-state hash (domain-separated streams as in oracle-curriculum.cpp, never from the seed) under a fixed public depth-1 fair continuation with horizon 25 (moves capped at 25 count as 25, censor flag kept), and R_tape/R_real = the realised remaining moves on the state's own trajectory capped at the same horizon. Report R_fair(O), R_fair(F), R_tape(O), R_real(F), tau = (R_fair(O)-R_fair(F))/(R_tape(O)-R_real(F)), by origin-game split halves, plus the kBlocked flow-band fraction over O, and at every O root the fair-top-1 rate (under the same 32 common futures over all legal columns) of the oracle's column versus fair D4's column at that root.

Candidate
oracle-curriculum state pool relabelled under public randomness
Comparator
fair-d4 matched state pool
approaches/fair-expectimax/reference/fair-only-depth4.cpp

Primary metric

tau = (R_fair(O) - R_fair(F)) / (R_tape(O) - R_real(F)), pooled

Statistical unit: whole-origin

Pass criteria, fixed in advance

  • CHECK gates before any seed is read: public-state constructor privilege boundary (the relabel binary links no oracle tape accessor), determinism of R_fair across two runs and across thread counts, domain separation of restart streams, mirror invariance of R_fair within floating tolerance.
  • tau >= 0.25 pooled.
  • Sign of R_fair(O) - R_fair(F) agrees in both origin-game halves.
  • Oracle-column fair-top-1 rate at O roots >= fair-D4-column fair-top-1 rate at the same roots.

On pass: Record valid + pass at CHECK tier (diagnostic); register the pool-comparison follow-on (fair / optimistic / mixed pools with fair labels, 5 init seeds each) as a successor experiment. No policy claim.

On fail: Record valid + fail at CHECK tier; assess the theory not-supported-as-tested; the optimistic-state pool follow-on is not registered; audit-05's H-pool program closes.

Data and reuse

Role: training-data generation, no evaluation gameplay. Opens the new training-role lease 0xa52e0000-0xa52e01ff (oracle games 0xa52e0000-0xa52e003f, fair-D4 games 0xa52e0100-0xa52e01ff) exactly once. Seeds are privileged inputs to the oracle generator only; every exported state is public. No development, protected or final seed is read. The range appears nowhere in the tree (grep of 0xa52e on 2026-08-23) and lies inside the SEEDLEASE-A52 reserve 0xa5278000-0xa52fffff.

seed leases: SL-20260823T200000Z-a52e0000

What happened

valid run · outcome: failnot-supported-as-testedpilotRS-20260823T205143Z-ead14c9d

Stage D0 refuses the H-pool theory on two of its three preregistered criteria. Pool O was 1,984 public states sampled from 64 oracle games (depth 4, beam 128; 63 of 64 games reached the 500-move cap); 1,271 were matched 1:1 to fair-D4 states on exact (rise phase, occupancy/4, max-height/2) buckets over all 256 comparator seeds (match rate 0.641; 713 O states dropped and counted). Under 32 common public futures with fair D1 continuation at horizon 25, oracle-visited boards are only marginally better than matched fair boards: R_fair(O) = 24.783 vs R_fair(F) = 24.520 (difference +0.263 moves, cluster bootstrap 95% [+0.156, +0.367], same positive sign in both origin-game halves). The preregistered transferable fraction is negative, tau = -0.959 (95% cluster interval [-3.069, -0.390]), because its denominator degenerates at this horizon: R_tape(O) = 24.338 vs R_real(F) = 24.611 (94-95% of both realised remainders were capped at 25), so the realised-lifetime gap the fraction was defined against is -0.274 moves rather than positive. Independently of that degeneracy, the action-quality criterion fails outright: the oracle's own column is fair-top-1 at its own roots less often than fair D4's column at the same roots (0.766 vs 0.814 over 1,271 roots, difference -0.049, 95% [-0.069, -0.029]; on the 194 unique-maximum roots 0.268 vs 0.387). Blocked-flow-band fraction over matched O states is 0.0047 (F: 0.026). Per the preregistration the theory is assessed not-supported-as-tested, the pool-comparison follow-on is not registered, and audit-05's H-pool program closes.

What it had to pass
  • CHECK gates before any seed is read: privilege boundary (relabel binary links no oracle tape accessor), determinism of R_fair across two runs and thread counts, domain separation of restart streams, mirror invariance of R_fair — observed: ALL GATES PASS in runs/RUN-20260823T191900Z-b9f8f80d/d0/gates.log on probe seeds 0xa5278000-0xa527810f: 0 oracle symbols in d0-relabel, generate byte-identical at 16 vs 4 threads, relabel byte-identical at 1 vs 8 threads across two runs, 0 mirror/metadata/sibling/stream failures on 47 probe states (mirror invariance is exact: the restart plays in the canonical frame), label edits changed no relabel byte
  • tau >= 0.25 pooled — observed: tau = -0.9590, 95% cluster interval [-3.0691, -0.3897]; denominator R_tape(O)-R_real(F) = -0.274 moves is degenerate at horizon 25 (94-95% of both realised remainders capped) while the numerator R_fair(O)-R_fair(F) = +0.263 [+0.156, +0.367]
  • sign of R_fair(O) - R_fair(F) agrees in both origin-game halves — observed: half 0 (games 0-31): +0.2319; half 1 (games 32-63): +0.3146; both positive
  • oracle-column fair-top-1 rate at O roots >= fair-D4-column fair-top-1 rate at the same roots — observed: oracle 0.7655 vs fair D4 0.8143 over all 1,271 matched O roots under the same 32 common futures (no subsampling; ties count for every tied column); difference -0.0488, 95% cluster interval [-0.0691, -0.0288]; strict-maximum roots only: 0.268 vs 0.387
Recorded metrics
tauPooled
-0.9590
tau95ClusterBootstrap
  1. -3.0691
  2. -0.3897
tauByHalves
half0
-0.8357
half1
-1.1761
R_fair_O
mean
24.7826
seStates
0.0365
seOriginGames
0.0633
R_fair_F
mean
24.5201
seStates
0.0491
seOriginGames
0.1772
R_tape_O
mean
24.3375
seStates
0.0904
seOriginGames
0.0827
R_real_F
mean
24.6113
seStates
0.0525
seOriginGames
0.2163
fairDifference
0.2626
fairDifference95ClusterBootstrap
  1. 0.1561
  2. 0.3670
fairDifferenceByHalves
half0
0.2319
half1
0.3146
realisedDifference
-0.2738
horizon
25
scenariosPerState
32
poolSizes
oStates
1,984
oMatched
1,271
oUnmatched
713
matchRate
0.6406
fStates
1,271
oOriginGames
64
fOriginGames
101
bucketsMatched
63
censoring
R_tape_O_cappedAt25Fraction
0.9473
R_real_F_cappedAt25Fraction
0.9426
R_fair_O_survivedHorizonFraction
0.9624
R_fair_F_survivedHorizonFraction
0.9167
oracleGamesCensoredAt500
63
oracleOriginCensoredStates
62
flowBands
O_matched_blocked
0.0047
O_all_blocked
0.0030
F_blocked
0.0260
top1AtORoots
roots
1,271
subsampled
false
oracleColumnTop1Rate
0.7655
fairD4ColumnTop1Rate
0.8143
difference
-0.0488
difference95ClusterBootstrap
  1. -0.0691
  2. -0.0288
oracleEqualsD4ColumnRate
0.3895
uniqueMaximumRoots
194
oracleColumnTop1RateStrictRoots
0.2680
fairD4ColumnTop1RateStrictRoots
0.3866
bootstrap
method
cluster bootstrap over O origin games carrying matched F partners
resamples
10,000
seed
0xb0071eaf
clusters
64
sensitivityIndependentClustersTau95
  1. -9.3496
  2. 5.1228
wallSecondsMain
3,132
oracleGameScoresMeanMoves
495
Limitations
  • The preregistered tau is ill-conditioned as measured: with 63 of 64 oracle games censored at the 500-move cap and comparator states drawn from the same shape buckets, both realised remainders sit at the 25-move horizon cap (R_tape(O) 24.34, R_real(F) 24.61, 94-95% capped), so the denominator is -0.274 moves instead of the large positive gap the definition assumed. The tau < 0.25 refusal is therefore driven by a degenerate denominator, not by a negative numerator; the numerator (the fair-value advantage of oracle boards) is positive but small, +0.263 moves on a 25-move horizon.
  • The fair-top-1 criterion is unaffected by that degeneracy and fails on its own: the oracle's action is fair-best at its own roots significantly less often than fair D4's action at the same roots.
  • Match rate 0.641: the 713 dropped O states skew toward buckets fair D4 rarely visits (matched O states are shape-matched by construction), so the numerator is a lower bound on the raw O-vs-F fair-value difference over all oracle states, and the comparison is conditional on shape overlap as preregistered.
  • R_fair uses a fixed public D1 continuation and horizon 25 with 32 futures; 92-96% of futures survive the horizon, so both pools are near the measurement ceiling and the +0.263 difference is compressed by censoring at both ends.
  • F origin games contribute clustered states (mean 12.6 matched states over 101 of 256 games); uncertainty uses cluster bootstrap over O origin games carrying matched partners, with the independent-clusters sensitivity interval also recorded (it is much wider: [-9.35, +5.12]).
  • Diagnostic tier (CHECK): no gameplay evidence and no policy claim; training-role seeds only, opened once.