Seed-free corpus gate, then a paired 256-game SCREEN of fair D4 (seven strata, memo engine) with reveal-construction leaf terms against the unchanged leaf
web/content/research/EX-20260823-reveal-construction-screen-371fd638.mdx and it will appear here. The registered protocol is shown below.The registered protocol
With the fast fair-D4 search at seven strata, the one-entry leaf memo (bit-exact, finding audit-06), cache 60,000, work bound worstCaseWork(4,7)+1 = 11,892,399, terminal utility -1,000,000, policy seed 0xd7075eed, 2,000-move cap, three arms play the same 256 ordered seeds 0xa52d0000-0xa52d00ff: 'frozen' (unchanged leaf, zero extra weights, gated bit-identical to FastSearch), 'A' (frozen + aligned_double_hit at +300), 'B' (frozen + aligned_double_hit +300 + chain_to_crack_cracked +150 + chain_to_crack_solid +300, all three multiplied by danger_gate = 1 if max height <= 4, 0.5 at height 5, 0 at height >= 6). Term definitions are fixed in approaches/lifetime-objective/chain-reveal-leaf/extra-terms.hpp as specified in runs/RUN-20260823T091530Z-cbe65468/kimi-k3-theory-design.md section 2 (i)-(iii). Before any leased seed is read, a seed-free corpus gate on runs/RUN-A51D-corpus/all.states (training-role data, already opened) must pass for aligned_double_hit; if it fails, the screen is not run and the record is valid + fail at mechanics tier. The primary contrast is A minus frozen; B minus A is a declared-underpowered secondary. A shadow unchanged search runs at every decision of arms A and B to measure coverage (root-argmax divergence), doubling their cost but changing no decision.
approaches/lifetime-objective/chain-reveal-leaf/run.cppapproaches/fair-expectimax/reference/fair-only-depth4.cppPrimary metric
paired mean whole-game score delta, A minus frozen, 256 games, depth 4 seven strata memo engine
Statistical unit: whole-game
Pass criteria, fixed in advance
- Corpus gate for aligned_double_hit, read once, thresholds fixed here: held-out partial correlation of the feature with log1p(remaining lifetime) beyond the 18 leaf features, occupancy and rise clock >= +0.05 AND incremental held-out R^2 >= 0.005 AND prevalence (feature > 0) >= 5% of depth-4 positions AND uncollected-setup rate (live setups with joint readiness >= 0.5 not followed by a reveal of that gray within 2 moves in the same game) >= 30%. If this fails the screen is not run. chain_to_crack is reported on the same gate but rides in arm B regardless (declared secondary).
- CHECK gates passed before the lease is opened: zero-weight bit parity of the augmented leaf with fastFairLeaf on every visited board, search parity (column, work, nodes, cache hits, completed depth) with FastSearch at d4s5 and d4s7, mirror invariance of every extra feature, thread determinism at live weights, metadata blindness, memo-vs-fresh extra-feature identity.
- incompleteDecisionsTotal 0 and illegalDecisionsTotal 0 in every arm.
- A minus frozen: bootstrap 95% lower bound > 0 AND Student-t 95% lower bound > 0.
- A minus frozen: candidate Q25 >= comparator Q25 AND both halves > 0.
- Arm A coverage >= 2% of its decisions; below 2% the result is inconclusive by rarity regardless of the deltas.
- Mechanism check (reported, not a pass condition): A's reveals per move exceed frozen's; a score gain with reveals per move flat is recorded as 'pass, mechanism unsupported'.
On pass: Record valid + pass; register a replication on the reserved 0xa52d0100-0xa52d01ff block and a seed-free dose sweep of the weight as successor experiments; register the arm as a playground leaderboard policy. No tier above SCREEN is claimed.
On fail: Record valid + fail for these exact terms and weights; report B and the corpus statistics as diagnostics; open no further cohort. If the corpus gate killed both H2 terms, or arm A recorded coverage < 2% with |delta| < 25,000, the predeclared fallback diagnostic (entombed_high soft penalty at -150, 128 paired games on a new lease, declared a non-measurement for score, judged on coverage and sign pattern) is registered as a successor experiment; it is not run under this protocol.
Data and reuse
The corpus gate reads runs/RUN-A51D-corpus/all.states (training-role seeds 0xa5200000-0xa52152ff, already opened; whole-origin held-out split from afterstate-net/dataset.py) and reads no new seed. CHECK gates read previously opened development probe seeds 0xa5278000-0xa52784ff only. The screen opens the new, never-read lease 0xa52d0000-0xa52d00ff exactly once; 0xa52d0100-0xa52d01ff of the same lease stays reserved for a replication or the predeclared extension and is not read by this experiment. The lease is disjoint from every range in docs/exploratory/lease-map.md; a full-tree grep for 0xa52d found no constant.
seed leases: SL-20260823T100000Z-a52d0000
What happened
Seed-free corpus gate for the reveal-construction leaf terms, read once on the depth-4 non-explored subset of runs/RUN-A51D-corpus/all.states (62,831 rows, 768 games, whole-origin split, base held-out R^2 0.6483): the gating term aligned_double_hit FAILS the preregistered four-part gate. Held-out partial correlation with log1p remaining lifetime beyond the 18 leaf features, occupancy and rise clock is -0.0443 (threshold >= +0.05) and incremental R^2 +0.00047 (>= 0.005); prevalence 19.75% and the uncollected-setup rate 60.47% (56.51% excluding ambiguous-empty outcomes; 6,969 setups) pass. chain_to_crack_cracked (partial r -0.0325, R^2 +0.00026, prevalence 28.92%) and chain_to_crack_solid (partial r -0.0385, prevalence 1.77%, uncollected 62.85%) also fail; entombed_high on depth 4 alone reads partial r -0.0595, incremental R^2 +0.00125. Per the protocol the 256-game screen was NOT run and no leased seed was read. All CHECK gates passed at the final term source (0 bit, parity, mirror, determinism, metadata mismatches at d4s5 and d4s7; runs/RUN-20260823T091530Z-cbe65468/gates.log). The value-criterion part of the gate rejects these terms as lifetime predictors; the action-level statistic says the depth-4 (s5, epsilon 0.03) behaviour policy leaves 60% of live same-wave double-hit setups uncollected within two moves.
- ✕aligned_double_hit held-out partial r >= +0.05 — observed: -0.0443
- ✕aligned_double_hit incremental held-out R^2 >= 0.005 — observed: 0.00047
- ✓aligned_double_hit prevalence >= 5% of depth-4 positions — observed: 19.75%
- ✓aligned_double_hit uncollected-setup rate >= 30% — observed: 60.47% (excluding ambiguous 56.51%)
- ✓CHECK gates (bit parity, search parity, mirror, determinism, metadata blindness, memo identity) at d4s5 and d4s7 — observed: 0 mismatches in every gate; runs/RUN-20260823T091530Z-cbe65468/gates.log
- –Screen gate (A minus frozen lower bounds, Q25, halves, coverage) — observed: not evaluated: the screen was not run because the corpus gate failed
Recorded metrics
- records
- 5,257,181
- depth4Rows
- 62,831
- depth4Games
- 768
- depth4Train
- 50,003
- depth4Test
- 5,528
- baseHeldOutR2
- 0.6483
- partialR
- -0.0443
- incrementalR2
- 0.0005
- prevalenceDepth4
- 0.1975
- uncollectedRate
- 0.6047
- uncollectedRateExcludingAmbiguous
- 0.5651
- setups
- 6,969
- revealedWithin2
- 2,750
- partialR
- -0.0325
- incrementalR2
- 0.0003
- prevalenceDepth4
- 0.2892
- partialR
- -0.0385
- incrementalR2
- 0.0001
- prevalenceDepth4
- 0.0177
- uncollectedRate
- 0.6285
- uncollectedRateExcludingAmbiguous
- 0.5775
- setups
- 324
- revealedWithin2
- 120
- partialR
- -0.0595
- incrementalR2
- 0.0013
- prevalenceDepth4
- 0.1678
- partialR
- -0.0504
- incrementalR2
- 0.0007
- prevalenceDepth4
- 0.1698
- partialR
- -0.0277
- incrementalR2
- 0.0003
- prevalenceDepth4
- 0.2383
- partialR
- -0.0064
- incrementalR2
- 0.0000
- prevalenceDepth4
- 0.0151
- leafBitMismatches
- 0
- parityMismatches
- 0
- mirrorMismatches
- 0
- determinismMismatches
- 0
- metadataBlindnessMismatches
- 0
- liveDivergence
- A_d4s5
- 2/600
- B_d4s5
- 5/645
- A_d4s7
- 7/370
- B_d4s7
- 6/310
- Mechanics-only tier: no gameplay was run; this rejects the terms as held-out predictors of remaining lifetime beyond the frozen leaf on depth-4 corpus positions, not as action-changing leaf terms in play.
- The corpus behaviour policy is the reference depth-4 search at FIVE strata with epsilon 0.03 exploration, not the seven-stratum memo configuration the screen would have run; the uncollected-setup rate is therefore inflated by epsilon moves and by s5's coarser chance model (Kimi K3 review, written before the result was read: runs/RUN-20260823T091530Z-cbe65468/kimi-k3-prereg-review.md section 3).
- The same review argued, blind to the result, that a partial-correlation kill criterion is a value criterion and can reject a term whose purpose is to re-rank sibling actions at the horizon boundary; the preregistered gate is nonetheless applied as written. Any successor that drops the value criterion must disclose that this gate was read first.
- The uncollected-setup rate counts a reveal of a different gray as not collecting this setup; a high rate can mean the search had better lines.
- The formal freeze command ran after the corpus gate was read; the thresholds were fixed by file content 9.5 minutes before the read (experiment amendment).
- The implementer's support-disjoint rule R2 (skip adjacent-side pairs whose completion paths share a cell) excludes configurations where one dropped disc completes both runs in the same wave; the term as frozen is narrower than the design's prose.