← ResearchTheories
Registered falsifiable claims from research/theories/. Each theory names its mechanism and the criteria that would refute it; experiments and results link back to these claims. Historical theories that predate the machine-readable registry live in the experiment ledger and on their approach pages.
not-supported-as-testedassessedpilotpublic-policyTH-20260820-distributional-afterstate-ranker-7aba7fb3
Claim
A single action-free public-afterstate evaluator, trained only on successor-closed labels (every legal sibling at each root, aligned chance scenarios, a fixed public continuation policy), ranks the legal root actions of disjoint whole-origin games at least as accurately as fair D4, and therefore can serve as the value core of a direct 1-ply chance-averaged policy with exact D4 fallback.
Mechanism
Prior learned policies in this repository failed for a diagnosed reason: they were trained on the action the behavior policy played (or a sparse sibling sample), so deployment required ranking actions whose successors were never labeled; low value error on visited states coexisted with poor sibling ranking. Making the evaluator action-free (its input is the fully resolved public afterstate: board, next visible disc, moves-until-rise) and the corpus successor-closed (every legal sibling's afterstates receive labels under common event-keyed chance scenarios) removes both the action-identity shortcut and the coverage gap by construction: every candidate action at deployment is scored by the same function on in-support states. Distributional (quantile) targets plus a within-root ranking loss train the relative ordering the policy actually needs rather than only absolute calibration, and the multi-move continuation labels supply the long-horizon signal that a one-ply heuristic lacks.
Falsification criteria
- On preregistered held-out whole origins, the model's scenario-mean action ranking underperforms exact fair D4's ranking of the same roots in top-1 accuracy against the H40 long-outcome target by more than the frozen margin.
- Within-root pairwise accuracy or normalized regret fails the frozen thresholds on any held-out origin fold.
- Action stability between independent scenario halves is below the frozen threshold, showing the labels themselves are too noisy to learn the ranking.
- The corpus cannot be built successor-closed (any legal sibling missing a label) at the pilot scale within the frozen compute budget.
Evidence: EX-20260820-afterstate-pilot-h40-29b8588a (completed)EX-20260820-afterstate-pilot-h40-k256-65a05477 (completed)EX-20260820-afterstate-pilot-h40-k64-d7a9faf5 (completed)EX-20260820-d4-toptwo-override-gate-0bdb39a1 (completed)EX-20260821-afterstate-d2-teacher-58772f17 (completed)EX-20260821-afterstate-full-train-override-ec7f28f4 (completed)result: inconclusiveresult: inconclusiveresult: failresult: failresult: failresult: fail
mixedassessedpublic-developmentpublic-policyTH-20260821-depth-gradient-beyond-fair-d4-034314fc
Claim
With seven chance strata (one stratum per disc value, so every individual chance marginal is exact), adding a fifth ply of look-ahead to the fair expectimax search raises mean score on the shared 64-game development cohort by a paired margin whose one-sided 95% whole-game bootstrap lower bound is above zero, as the third-to-fourth ply does (+86,172 with a lower bound of +26,468, finding-05). With five strata, where the third-to-fourth ply gradient is absent (-7,723, not significant), no fourth-to-fifth ply gradient is expected either. Depth and chance-estimator exactness are therefore complements, and the fourth ply is not a special stopping point.
Mechanism
finding-05 measured a depth gradient that exists only at the higher stratum count: at five strata the search's chance estimate is biased enough that a deeper horizon compounds the bias faster than it adds information, while at seven strata the estimate's marginals are exact and the extra ply pays. If that is the whole mechanism, the same exchange should continue at the fifth ply, which is also the first depth that always contains a row rise (a rise happens every five placed discs), so a five-ply search is the shallowest one that can never be surprised by the rise schedule. The competing account is that the fair search's terminal utility has no death-depth shaping - bestFutureValue returns the same -1,000,000 whether the game ends at ply 1 or ply 5 - so a deeper horizon finds more within-horizon deaths, flattens the sibling value surface, and hands the decision to column-order tie-breaking. The two accounts differ in sign at the fifth ply, which is what makes this worth the compute.
Falsification criteria
- At seven strata, the paired depth-5 minus depth-4 score delta over the complete 64-game cohort has a one-sided 95% whole-game bootstrap lower bound at or below zero.
- At five strata, the paired depth-5 minus depth-4 score delta over the complete 64-game cohort has a one-sided 95% whole-game bootstrap lower bound at or below zero.
- Numbered clears per move, covered reveals per move and mean occupancy do not all move toward the survival requirement (2.400 clears and 1.400 reveals per move) from depth 4 to depth 5 at the same stratum count.
- A depth-5 arm cannot complete its declared depth within its work bound (any decision below the requested depth voids the arm rather than falsifying the theory).
Evidence: EX-20260821-depth5-chance-exactness-factorial-a6a604fd (running)result: failresult: inconclusive
not-supported-as-testedassessedpublic-developmentpublic-policyTH-20260821-factored-chance-depth-compounding-aca01725
Claim
Factoring the fair search's chance node into independent next-disc and reveal draws pays at every depth, so applying it on top of the fourth ply raises mean score over the same four-ply search with a single reveal sample by a paired margin whose one-sided 95% whole-game bootstrap lower bound is above zero. The depth-3 gain finding-09 measured (+64,116 with a lower bound of +7,475 from one to six reveal samples) is therefore an increment that adds to depth rather than an alternative way of buying the same thing.
Mechanism
finding-09 diagnosed a specific defect: the reference search indexed the next-disc draw and the covered-disc reveal with the same counter, so with seven samples it saw seven of the 49 joint outcomes and those seven were perfectly correlated. A move whose value depends on the combination could not be evaluated. That defect is a property of the chance node, not of the horizon, and the frozen four-ply search contains exactly the same defective node. If the defect is what the extra reveal samples repair, repairing it inside a deeper search should be worth at least as much as repairing it inside a shallower one, and the two axes should add. The competing account is that depth and chance-estimator quality are exchangeable ways of spending one budget - both ultimately buy a less biased estimate of the same expectation - in which case paying for both at once buys little beyond paying for either.
Falsification criteria
- The paired depth-4 (N=7, M=2) minus depth-4 (N=7, M=1) score delta over the complete 64-game cohort has a one-sided 95% whole-game bootstrap lower bound at or below zero.
- The depth-3 reveal ladder stops ordering with joint chance coverage: taking reveal samples from six to twelve (85.7% to 100% joint coverage) does not raise the paired mean over the six-sample arm.
- Any arm completes a decision below its requested depth, which voids that arm.
- The joint-coverage attribution fails: an arm with higher joint (disc, reveal) coverage scores below one with lower coverage at the same depth and the same work bound.
Evidence: EX-20260821-reveal-sampling-unfinished-arms-470677b5 (running)result: failresult: fail
untesteddraftproposalpublic-policyTH-20260821-search-guided-self-play-at-scale-299ed02f
Claim
A search-guided self-play loop — actors that run a fair public-information search at every visited position, record the value of every legal sibling column under aligned chance scenarios, and train a public board evaluator on those successor-closed labels, which then becomes the leaf of the next iteration's search — produces a frozen public student whose mean whole-game score on fresh paired development cohorts rises with the compute spent per iteration, and at sufficient scale exceeds fair D4's mean by a preregistered margin.
Mechanism
Most learned evaluators in the ledger failed at ranking siblings because training data labelled only the move that was played (status.md §4-§7; sibling extrapolation). The distributional-afterstate pilot (TH-20260820-distributional-afterstate-ranker-7aba7fb3) removed that defect — it labelled every legal sibling under chance scenarios shared across siblings — and its four recorded pilot experiments relocate the bottleneck rather than resolving it: label stability required raising aligned scenarios from 8 to 256, and the resulting model still lost to fair D4 on ranking, with its own result record naming an undertrained checkpoint (11 of 20 epochs) and a weak D1-continuation teacher as the likely causes. This theory claims the missing element is the loop: using a search to produce the labels makes them stronger than the evaluator that guided the search (policy improvement), and feeding the trained evaluator back as that search's leaf and repeating is approximate policy iteration. Each iteration's labels then come from a strictly stronger searcher, which is the mechanism by which additional compute - more actors, more aligned scenarios per label, deeper or wider search per label, more iterations, longer training - turns into longer-horizon play rather than into a more confident copy of the same bias. It is the AlphaZero pattern adapted to a single-player stochastic game with a hard information boundary. No iteration of that loop has been run in this repository.
Falsification criteria
- Pilot scale (this workstation): after a preregistered number of iterations, the evaluator's held-out sibling ranking on a locked every-sibling panel (top-1 and pairwise agreement with the fair-D4 gold standard, as in the 477-root panel) does not improve across iterations, or never reaches fair D4's own agreement.
- Compute response: doubling actor compute between two consecutive iterations, with the data pipeline and learner held fixed, produces no paired SCREEN-tier (32-game) mean-score improvement beyond the whole-game bootstrap interval.
- Deployment: the frozen public student of any iteration, evaluated through the public interface on a fresh paired 64-game development cohort, does not exceed fair D4's mean with a one-sided 95% whole-game bootstrap lower bound on the paired difference above zero.
- Information boundary: any student that reads seed, hidden values, future discs, score, level, or move number is disqualified regardless of score.
Evidence: EX-20260821-afterstate-d4q-stage1-40136e9e (completed)result: fail
not-supported-as-testedassessedpublic-developmentpublic-policyTH-20260822-depth-native-leaf-weights-caa6f8ba
Claim
For the corrected-score fair depth-4 search with the frozen five-stratum chance model (the contract comparator fair-d4), there exists a vector for the eighteen active fair-leaf weights whose mean whole-game score on fresh development seeds exceeds the frozen vector's by more than a 64-game paired cohort's detection floor, and a derivative-free optimiser that evaluates candidates by complete depth-4 games on a fresh seed block every generation can find such a vector within 40 generations of 16 candidates.
Mechanism
The leaf prices what four plies of search cannot see (finding-14). The frozen weights were hand-set for a depth-3 search under historical 7,000-point scoring and have never been refitted under corrected scoring or at the depth they are used at (approaches/fair-expectimax/fair-policy/README.mdx). The two prior weight optimisations failed under conditions that each independently explain failure without implying optimality: the CEM run tuned at depth 3 on rotating three-game batches and transferred its vector to depth 4 unchanged; the heuristic-search evolutions fitted one-ply policies on a fixed seed set and regressed on disjoint seeds. Fresh seeds every generation remove the seed-overfitting channel by construction; depth-native fitness removes the depth-transfer confound; common random numbers within a generation let small perturbations of one policy be ranked with far less variance than two unrelated policies. If the frozen point were a local optimum of mean score at depth 4, the population mean would show no paired gain over the control on fresh seeds and the frozen candidate would fail the held-out screen.
Falsification criteria
- The preregistered held-out paired screen (64 never-read development games at depth 4, five strata) of the frozen final candidate against the frozen leaf has a one-sided 95% bootstrap lower bound on the mean score delta at or below zero, a Student-t lower bound at or below zero, a lower-quartile regression, or opposite-sign halves: the claim is not supported as tested.
- The paired population-mean-minus-control score delta on fresh training blocks does not exceed +20,000 in any of the last ten completed generations: no measurable gradient exists at this budget, and the claim is not supported as tested.
- Any generation artifact or screen artifact records an incomplete decision, an illegal decision, or a score-identity failure: the run is invalid and cannot support or reject the claim.
- Out of scope and explicitly not tested by this theory: transfer of a five-stratum-evolved vector to the seven-stratum search is measured as a secondary, diagnostic quantity only.
Evidence: EX-20260822-leaf-cmaes-d4s7-4f5f462a (completed)result: fail
not-supported-as-testedassessedmechanics-onlydiagnosticTH-20260822-entombed-disc-hazard-61faa529
Claim
A numbered disc of value n whose column already holds more than n discs and whose horizontal run already exceeds n (an 'entombed' disc; for n >= 3 unpriced by the frozen fair leaf, which has such terms only for 1s and 2s) is an early predictor of death for the fair depth-4 policy: it is present at most deaths, it typically appears well outside the four-ply horizon (more than two rises before death), and it carries lifetime information beyond the eighteen leaf features and occupancy.
Mechanism
In a gravity-packed column the vertical run through every disc equals the column height, so a disc n in a column taller than n can only clear horizontally, and if its row run already exceeds n it cannot clear at all until neighbours are removed. Such a disc also blocks vertical cracking of every cover beneath it. The harm (unrevealed covers, a column that only grows, rise pressure) arrives many moves later, outside a four-ply search, and the leaf's height terms are value-blind (a 7 and a 3 at height 5 score the same) while its dead-disc terms stop at value 2. The search is therefore not short of fear of death (finding-04: the terminal utility is saturated) but of foresight: the decision that entombs the disc looks harmless at depth 4. A trigger that fires when such a disc appears is a candidate for the 'emergency mode' the owner asked about, and a leaf term for it is a candidate 19th coordinate.
Falsification criteria
- In complete depth-4 behaviour games of the corpus, fewer than half of deaths have an entombed disc of value >= 3 on the board at death: the disc is not the usual way the policy dies.
- The median lead time between the first persistent entombed >= 3 disc and death is under 5 moves (one rise): the signal arrives too late to act on and is not foresight.
- On a whole-origin held-out split, the partial correlation of the entombed count with log remaining lifetime, after the eighteen frozen leaf features, occupancy and the rise clock, is above -0.02, or the incremental held-out R^2 is below 0.005: the leaf already prices it.
- Out of scope: whether pricing it changes play. That is a separate preregistered gameplay experiment.
Evidence: EX-20260822-entombed-disc-corpus-analysis-38be404e (preregistered)result: fail
supported-as-testedassessedmechanics-onlyengineeringTH-20260822-leaf-memo-semantics-preserving-3bb0ac6b
Claim
Memoising the fast fair leaf's eighteen board-only features across consecutive leaf calls with the same (board, moves_remaining), recomputing only next_disc_vertical_options, produces bit-identical leaf values, identical actions, logical work, nodes, cache hits and completed depth, and reduces the wall time of a depth-4 decision by at least 1.5x at five and seven chance strata.
Mechanism
Within one chance node the strata share the post-move board whenever the cascade consumed no reveal draw; only the sampled next disc differs, and fast-leaf.hpp reads next_disc in exactly one term. Identical inputs give identical doubles; the dot product is re-run in the frozen order; the memo sits below the search's work increment so no recorded observable changes (audit-06 and its adversarial review).
Falsification criteria
- Any leaf-value bit mismatch, any action/work/node/cache-hit/completed-depth mismatch against FastSearch, or any thread-count-dependent result on the probe games: not semantics-preserving.
- Interleaved timing ratio below 1.5x on real roots at either strata setting: the speedup claim is not supported as tested.
Evidence: EX-20260822-leaf-memo-gates-b31e4e7f (completed)result: pass
supported-as-testedassessedmechanics-onlydiagnosticTH-20260822-nnue-leaf-capacity-saturation-9c45e42e
Claim
On the existing 5.26M-record training corpus (runs/RUN-A51D-corpus/all.states, whole-origin split of dataset.py), the NNUE-shaped survival leaf's held-out lifetime correlation does not improve by more than 0.01 when hidden width is raised from 64 to 512, the middle layer from 32 to 64, epochs from 10 to 20, or the learning rate is changed: its accuracy is saturated by the feature space and the data, not by capacity or training length.
Mechanism
finding-08 reports held-out lifetime correlation 0.8564 for the 572k-parameter h64 student against 0.8646 for a 3.0M-parameter convolutional teacher, and concludes that 'the accuracy is not there to be had'. That conclusion rests on one pair of architectures. If the 135-active-feature first layer is the bottleneck, wider hidden layers and longer training will not move the held-out correlation; if the student is under-parameterised or under-trained, they will. Either answer is cheap (about 35 s of GPU per run) and neither requires a game.
Falsification criteria
- The best sweep configuration's held-out lifetime Pearson exceeds the h64/m32/10-epoch/lr 3e-3 baseline by at least 0.02 and by more than three times the seed-to-seed spread of the baseline: the claim is not supported as tested (capacity or training length matters).
- A delta between max(0.005, 2 x seed spread) and that threshold: inconclusive at this corpus size.
- Out of scope: whether a more accurate leaf plays better. finding-08 already shows a 0.008 correlation gap did not separate the CNN from the student in play; a gameplay test is a separate preregistered experiment.
Evidence: EX-20260822-nnue-leaf-capacity-sweep-79632fd9 (preregistered)result: pass
mixedassessedpublic-developmentpublic-policyTH-20260822-survival-instinct-vertical-clear-26f173b5
Claim
Survival instinct: when the next disc (value n >= 3) would land as the (n+1)-th or higher disc of a column it can never clear vertically, and refusing such placements at the root of the fair depth-4 search — leaving the search to choose among the remaining legal columns — extends mean paired lifetime and score on fresh development games relative to the unchanged search, by more than a 128-game paired cohort's detection floor.
Mechanism
Proposed by the repository owner from watching the depth-4 search: it places 3s and 4s above the row where they could ever clear vertically and dies under them a few rises later, while a human treats such a placement as a danger that must be cracked horizontally and avoids making it. The corpus analysis (RS-20260822T051732Z-ceb257d7) established that an entombed disc is present at 75% of depth-4 deaths with a median 11-move lead, and that the search's leaf already VALUES such positions correctly; what is untested is the ACTION: whether the move that creates the disc was avoidable at acceptable cost. A root-level refusal is the minimal form of an 'emergency mode' — a trigger computed from public state and a corrective constraint that overrides the search's preferred column — and its coverage (how often it overrides) is measured alongside its effect.
Falsification criteria
- The strict-filter arm's paired score delta against the unchanged search on 128 fresh development games has a one-sided 95% bootstrap lower bound at or below zero, or a lower-quartile regression, or opposite-sign halves: refusing entombing placements does not help at this budget.
- The filter overrides the unchanged search's column in fewer than 2% of decisions: the rule rarely applies and cannot explain the observed deaths.
- If both arms lose on the mean but raise the lower quartile, the rule trades the upper tail for the floor, which finding-14 showed is the wrong trade for the million-point target; recorded as fail, not mixed.
- Coordinator's prior, stated before the run: the occupancy-matched corpus numbers (about two moves of life per entombed disc) predict an effect below the 128-game floor; a null here is expected and would be a non-measurement for effects under roughly 33,000 points.
Evidence: EX-20260822-survival-instinct-root-filter-8531cbd9 (completed)result: inconclusive
supported-as-testedassessedpublic-developmentpublic-policyTH-20260823-d3-reveal-quadrature-beats-d4s5-paired-4ee1d5fe
Claim
The depth-3 factored-chance fair search with seven disc strata and six reveal samples (D3 N7M6, fair leaf, finding-16) beats the frozen fair depth-4 five-stratum search (D4 s5, N5 M1, finding-05) on the shared development cohort 0xa51d1000-0xa51d103f by a paired per-game score margin whose one-sided 95% whole-game bootstrap lower bound and Student-t lower bound are both above zero.
Mechanism
finding-16 recorded 376,442 for D3 N7M6 and finding-05 recorded 297,327 for D4 s5 on the same 64 seeds, a +79,115 difference of means, but the two arms were never contrasted as paired per-game deltas: every published bound for D3 N7M6 was taken against D3 M=1, D3 M=12 or D4 N7 arms. The K2 program (runs/RUN-20260823T191900Z-b9f8f80d, section 1.4 and section 8 row C0) leans on this +79k as the head start of its fallback route (c). The mechanism claimed is that decorrelating the reveal draw at depth 3 buys more than the fourth ply buys at five strata (finding-05 measured the fourth ply at -7,723 NS at s5), so the shallower, better-quadratured search should dominate pairwise and not merely in means. If the paired bound does not clear zero the +79k is a heavy-tail artefact of 64 games and cannot be quoted as a head start.
Falsification criteria
- The paired D3 N7M6 minus D4 s5 score delta over the 64 shared seeds has a one-sided 95% percentile bootstrap lower bound (20,000 resamples, seed 0xb0071eaf) at or below zero.
- The paired D3 N7M6 minus D4 s5 Student-t one-sided 95% lower bound is at or below zero.
- The two cohort halves in seed order disagree in sign, or W-T-L shows the mean is carried by fewer than a quarter of the games.
Evidence: EX-20260823-d3n7m6-vs-d4s5-paired-reanalysis-ea66f4ec (completed)result: pass
not-supported-as-testedassessedpilotdiagnosticTH-20260823-hpool-optimistic-states-fair-labels-bb707bc9
Claim
Public board states visited by the perfect-information oracle after move 50 have a higher expected remaining lifetime under FAIR, public-randomness continuation than fair-D4-visited states matched on rise phase, occupancy and maximum height. Defining R_fair(O) and R_fair(F) as the mean remaining moves over K independent public futures under a fixed public continuation policy, and R_tape(O), R_real(F) as the realised remaining moves on the states' own trajectories, the transferable fraction tau = (R_fair(O) - R_fair(F)) / (R_tape(O) - R_real(F)) is at least 0.25, with the same sign in both split halves of the O pool, and the oracle's column at oracle-visited roots is fair-top-1 under common public futures at least as often as fair D4's column is.
Mechanism
docs/exploratory/audit-05-optimistic-curriculum.md sections 4-7 classify every oracle-target failure as class (i) information gap and argue the removable cause is optimism in the LABEL (a function of the realised tape). The salvageable core, H-pool, is optimism in the STATE DISTRIBUTION with fair labels: oracle-visited states are publicly distinguishable (manifold classifier held-out AUC 0.925/0.915) and survival-type targets are the only targets this corpus ever fitted well. The crux number - the fair-relabelled value of an oracle-visited state - was computed by approaches/oracle-curriculum/state-curriculum/oracle-curriculum.cpp (7 public-derived H25 restarts per state, domain-separated streams, privilege terminated at the PublicState constructor) but never reported. This theory is a measurement, not a model: if tau is small, the oracle reaches its long lifetimes by knowing the tape, not by standing on better boards, and no student can inherit that by training on those boards. Design: runs/RUN-20260823T191900Z-b9f8f80d/kimi-k3-open-items.md (K4-B, Kimi K3).
Falsification criteria
- tau < 0.25 on the pooled O set (K >= 32 public futures per state, fixed public D1 continuation, horizon 25, as in the curriculum code) refuses H-pool.
- Split halves of the O pool (by origin game) disagree in the sign of R_fair(O) - R_fair(F).
- At oracle-visited roots the oracle's column is fair-top-1 under common public futures less often than fair D4's column.
- Any one refusal ends the optimistic-teacher program; no follow-on student is trained. A pass licenses exactly one follow-on (pool comparison A1 fair / A2 optimistic / A3 mixed with fair labels), registered separately.
Evidence: EX-20260823-hpool-stage-d0-e0ad1c65 (completed)result: fail
not-supported-as-testedassessedpilotpublic-policyTH-20260823-nnue-class-holds-d4-ordering-c7b397a5
Claim
A LeafNet-shaped student (approaches/lifetime-objective/learned-leaf/leaf_features.py features, EmbeddingBag(8902,64)->32, ~572k parameters, ~1.3 us per state) trained with a within-root listwise plus pairwise loss on the existing successor-closed exact D4 sibling-value labels (runs/RUN-20260821T085042Z-c5cf0e71/d4q-labels, 291,890 afterstates over 8,639 roots) reproduces fair D4's within-root ordering on whole-origin held-out roots at the frozen d4q thresholds: top-1 >= 0.60, pairwise >= 0.78, normalised regret <= 0.13, in each half-fold.
Mechanism
RS-20260821T104500Z-77d21e90 showed a 3.4M-parameter afterstate CNN reaches only top-1 0.375 on these labels, below exact D1 (0.486), and status.md relocated the obstacle to the capacity of a compact board evaluator. The leaf-affordable NNUE class has never been trained on these labels: it was trained on played-action lifetime (saturated at Pearson 0.855-0.857, RS-20260822T024228Z-94090db1) and on planner values (finding-11). The draft self-play theory TH-20260821-search-guided-self-play-at-scale-299ed02f needs a student that both holds D4's ordering and runs at leaf cost; this probe decides whether that student can exist in the only leaf-affordable model class, on data already on disk, for about one CPU-hour. A fail is a valid negative that routes the self-play loop to a root-prior redesign. Design: runs/RUN-20260823T191900Z-b9f8f80d/kimi-k3-open-items.md (K4-D probe, Kimi K3). Designer's prior: P(pass) about 0.08.
Falsification criteria
- Held-out (whole-origin, the d4q experiment's own split) top-1 < 0.60 OR pairwise < 0.78 OR normalised regret > 0.13 in either half-fold, for the best of 5 initialisation seeds judged on the validation origins only, refutes the claim for this model class at this size; the held-out panel is already-read diagnostic data, so even a pass is not evidence and licenses only a fresh-manifest confirmation.
- A pass that fails a fresh whole-origin label manifest confirmation refutes it at one tier higher.
Evidence: EX-20260823-nnue-d4q-ordering-probe-0ca09bb1 (completed)result: fail
not-supported-as-testedassessedpublic-developmentpublic-policyTH-20260823-reveal-construction-leaf-e8f7c78d
Claim
Adding to the frozen fair-D4 leaf a term that prices the JOINT same-wave readiness of two support-disjoint numbered neighbours of a solid gray (the cheapest reveal in the game: two hits in one cascade, no cracked intermediate), optionally with a second gray-terminating release step (a chain whose release delivers a hit to a cover) and a danger gate that switches these construction terms off when max height exceeds 4, increases cover reveals per move and mean paired lifetime and score against the unchanged search on fresh development games at depth 4 / seven strata, by more than a 256-game paired cohort's detection floor.
Mechanism
Owner's hypotheses (2026-08-23): (H2) the search does not try to maximise gray reveals, and lining up a chain reaction that fully exposes a gray disc is especially valuable; (H1) the policy should track discs needing resolution and build long chains when no emergency is pending. Kimi K3's design (runs/RUN-20260823T091530Z-cbe65468/kimi-k3-theory-design.md) locates what is structurally missing rather than mis-weighted: the chance model already enumerates reveals inside the 4-ply horizon and the leaf already pays ~+600 to +5,000 per reveal through cell counts and height relief; every prior reveal arm priced the EVENT (fair-reveal-reward) or the MARGINAL neighbour readiness (t1-exposures, -67,719) and achieved fewer reveals. No term expresses the joint same-wave condition of fast-engine.hpp:349-357 (solid_exposure is best*0.35+second*0.65 of marginals), no term extends latent_chain_potential a second release step, and no sum of linear terms expresses the 'no emergency -> build' product. The term pays for a configuration one move before a reveal, at the horizon boundary, adds no chain-depth incentive (finding-06: chain seeking trades lifetime), and is hard to farm without being one move from a reveal. H1's ledger clause is demoted: the entombed-3..7 count has held-out partial r -0.023 beyond the leaf and the hard action filter changed 1.97% of decisions (RS-20260822T233343Z-12becce9); it survives only as a declared-diagnostic fallback soft penalty.
Falsification criteria
- Seed-free corpus gate (runs/RUN-A51D-corpus/all.states, entombed-discs pipeline, thresholds fixed before the dump is read): the candidate term is killed if its held-out partial correlation with log1p remaining lifetime beyond the 18 leaf features, occupancy and rise clock is below +0.05, or incremental R^2 below 0.005, or prevalence below 5% of positions, or (aligned_double_hit) the fraction of live setups (joint readiness >= 0.5) the depth-4 policy fails to convert into a reveal within 2 moves is below 30%. Gameplay gate: arm A (aligned_double_hit +300) minus frozen on 256 fresh paired games at d4s7 has a one-sided 95% bootstrap or Student-t lower bound at or below zero, or Q25 regression, or opposite-sign halves, or root-argmax coverage below 2% (inconclusive by rarity). Predeclared mechanism directions: reveals/move up, clears/move not down, mean occupancy not up; a score gain with reveals/move flat is not support for this mechanism. Designer's prior stated before the run: P(arm A passes) about 0.15; modal outcome a fourth reveal-pricing null.
Evidence: EX-20260823-reveal-construction-screen-371fd638 (completed)EX-20260823-reveal-construction-screen-v2-63d73b6a (completed)result: failresult: fail
untestedproposedproposalpublic-policyTH-20260823-sibling-outcome-labels-fix-leaf-0c3d480d
Claim
A leaf-cost student (the existing 572k-parameter finding-08 NNUE, architecture unchanged) trained on ALL legal siblings per root, with labels that are long-horizon outcome distributions (per-rise hazard vector plus censored restricted-mean lifetime, Kaplan-Meier over K CRN-paired continuations under a fixed cheap public continuation policy) and a within-root pairwise ranking loss on the deployed scalar (KM expected lifetime), achieves within-root discrimination that played-action outcome training cannot provide (coverage) and exact-search-value distillation cannot provide (objective): offline, top-1 agreement with exact D4's argmax >= 0.55 and >= the played-action incumbent + 0.03; deployed as the existing blended leaf, it improves the paired 256-game d4s7 mean with a positive one-sided 95% lower bound without giving back the incumbent's d4s5 gain.
Mechanism
Design: Kimi K3, runs/RUN-20260823T191900Z-b9f8f80d/kimi-k3-main-design.md (P-SOL-v1), grounded in tonight's three results. M1 coverage: one row per played move gives no within-root gradient; complete panels supply 21 ordered pairs per root. M2 objective: exact D4 values encode search depth a 1.33us leaf cannot represent (RS-20260823T194142Z-946e3cd1: top-1 0.296/0.301 at 572k, 0.375 at 3.4M, vs exact D1 0.486); outcome distributions under a fixed cheap continuation are properties of the afterstate and leaf-representable. M3 label economics: under within-root CRN pairing, sibling outcome differences are driven by afterstate structure, so a cheap continuation engine (D1/D2) can preserve the ranking of D3 N7M6 (the C0-validated strong cheap policy, RS-20260823T194200Z-42b113db) - measured by the G0 fidelity ladder before any corpus spend, never assumed. Single-factor discipline: model, features, head weights and deploy path held identical to finding-08; only sibling-complete data, the K-sample CRN estimator and the ranking term change.
Falsification criteria
- F1: no cheap continuation engine achieves within-root label Kendall-tau vs D3 N7M6 with 95% LB >= 0.85 on the G0 ladder (label-fidelity failure; stop at ~14 CPU-h).
- F2: offline gate G1 fails (student top-1 vs exact-D4 argmax < 0.52, or < incumbent LeafNet + 0.01) while the G2 decomposition shows target quality adequate (label-argmax top-1 vs D4 >= 0.60): sibling-complete data did not fix discrimination.
- F3: G2 shows label-argmax top-1 vs D4 < 0.60: the outcome-label family itself does not track decision quality; falsifies the label mechanism regardless of fit.
- F4: gameplay co-primary fails: paired 256-game d4s7 lower bound <= 0 (successor screen experiment).
- A negative rejects only the tested configuration {continuation in D1/D2, H=48, K=16, 572k student, ranking+regression mix}; D3-continuation labels at corpus scale remain open via the scale-out path.
- Status 2026-08-23: F1 is triggered in its v2 form - no affordable continuation engine could be certified (the fast M=1 proxy failed the D3 N7M6 guardrail: mean tau 0.370, top-1 4/6) - so the core claims M1/M2 remain untested rather than refuted; the theory is blocked pending the fast-engine M=6 port (E-FAST-M6) or a scale-out lease for native-M6 labels.
Evidence: EX-20260823-fast-m6-reveal-sampling-port-be23e203 (completed)EX-20260823-sol-corpus-and-offline-gate-4d3d86e4 (superseded)EX-20260823-sol-corpus-and-offline-gate-v2-f95f56e4 (completed)result: failresult: pass
untestedproposedproposalpublic-policyTH-20260824-m6-depth-second-order-for-ordering-7de1e79e
Claim
With the reveal quadrature fixed at M=6, continuation-search depth is second-order for WITHIN-ROOT sibling ordering: fast d1-M6 or d2-M6 KM-lifetime orderings agree with fast d3-N7M6 at mean within-root Kendall tau LB95 >= 0.75 and top-1 agreement LB95 >= 0.80 at the corpus operating point (K=8, H=40, CRN continuations shared across siblings and engines), and labels from the certified cheap engine train the 572k NNUE leaf to the P-SOL G1/G2 gates (student top-1 vs exact D4 >= 0.55 and >= incumbent + 0.03; label-argmax vs D4 >= 0.60; fit >= 0.7x the split-half label ceiling).
Mechanism
Design: Kimi K3, runs/RUN-20260823T191900Z-b9f8f80d/kimi-k3-psol3-design.md (P-SOL-3). The v2 guardrail showed the REVEAL axis is first-order for sibling ordering (fast-M1 vs native D3 N7M6 mean tau 0.370, top-1 4/6, RS-20260823T225753Z-0fbd48c3) while E-FAST-M6 made M=6 affordable at shallow depth (RS-20260824T010000Z-8f3e9b4f; measured 0.177 CPU-s/move at d3 N7M6 continuation duty; d1/d2-M6 estimated ~2/~54 ms/move, measured first in T0). If depth is second-order once M=6 is fixed, a d1-M6 corpus prices at ~17.8 CPU-h for 14,336 whole-origin roots x 7 siblings x K8 x H40 - affordable - while the d3-M6 equivalent is ~180 roots. Falsification (a) both cheap engines miss the ladder thresholds: the shallow-depth label claim is dead; (b) certified-engine corpus fails G2 target-quality: the ladder reference is an inadequate D4 proxy; (c) G2 passes and G1 fails: the bottleneck is the student or training, not the labels.
Falsification criteria
- Ladder (R=64, two cohorts 16 then 48, early stop): both d1-M6 and d2-M6 below mean tau LB95 0.75 or top-1 LB95 0.80 vs fast d3-N7M6 refutes the depth-second-order claim at K8 H40.
- G2 label-argmax vs exact D4 < 0.60 on the 4,096-root gate set refutes the certified engine as a label source.
- G1 fail with G2 pass narrows the failure to student/training and leaves the label claim standing.
- S0: if T0 re-pricing leaves fewer than 6,000 affordable corpus roots, the experiment records no-run rather than running underpowered.
Evidence: EX-20260824-psol3-m6-ladder-corpus-gate-2d0167ad (preregistered)
untesteddraftproposalengineeringTH-20260824-rust-bitboard-engine-f68fcbfd
Claim
A 28-byte column-major board (7 x u32, 4 bits per cell) makes gravity, disc placement and legality constant-time bit operations (PEXT compaction on Zen 5, portable nibble loop elsewhere); combined with a packed-key open-addressed transposition table and per-thread searchers, the fair expectimax search runs bit-identical to the frozen C++ reference at higher throughput and lower memory, scaling near-linearly to all available cores.
Mechanism
Column nibble packing turns gravity into PEXT(word, expanded-nonzero-mask); occupancy/numbered/covered bitboards are derived with SWAR nibble tests and PDEP scatters; popper detection uses the proven 128-entry run-length table on row/column masks; cover hits are counted for the whole board at once with a 4-way bitboard adder; the leaf, chance stratification, canonicalisation, column order and accumulation order are copied character-for-character from the proven C++ fast engine.
Falsification criteria
- Any mismatch in the trajectory gate (Rust vs C++ reference playHeadlessMove over leased seeds), the leaf gate (uint64 bit patterns of leaf values), or the search gate (per-column values and chosen actions at fixed depth/strata); or end-to-end throughput at identical work not exceeding the C++ fast engine; or worker-count dependence of chosen actions in complete-depth mode.
Evidence: EX-20260824-rust-engine-parity-throughput-4036a91f (completed)result: pass