Research roadmap toward a million-point policy

This roadmap turns the opportunities in strategies.md into an ordered, falsifiable program. It is not a claim that the proposed model will work. It prioritizes the failure most often seen in the existing research: models that predict visited states but cannot reliably rank all legal sibling actions over several row-rise cycles.

Research question

Can a policy using only public state rank every legal action by its long-horizon stochastic outcome, across disjoint whole-game origins, well enough to improve complete games over unchanged fair D4?

Fair D4 is the strongest retained reference, with a ledger-recorded broader mean near 308,296, but it is far below the target. Privileged planners and long rollouts show that valuable multi-cycle structure exists; previous students, partial sibling panels, and expensive search variants did not transfer or estimate it reliably. See status.md for the evidence boundary.

Safe bootstrap work package

In a checkout with no target Linux machine profile and no authorized retained dataset manifest, the first autonomous task is a CHECK-tier AFBR-40 closure preflight—not gameplay or model training. Register a no-gameplay diagnostic that:

• freezes the public state/afterstate key and required sibling/label/provenance fields;
• inventories repository-local filenames and manifests without reading seed contents;
• advances only if the coordinator names authorized, previously evaluated development artifacts; and
• otherwise records that closure is unestablished.

A suitable initial bound is 15 minutes, one CPU, 2 GiB of host memory, and an immediate stop on seed-content access or ambiguous provenance. This preflight can falsify data availability, but it cannot establish policy strength or claim that AFBR-40 has been implemented.

Priority map

1. Re-establish the reference on the target machine

Before changing strategy, create a current source/binary/machine manifest and measure fair D4 under corrected scoring. Preserve fixed algorithmic work and selected actions. Benchmark raw transitions, complete games, D1/D2/D4 work, thread scaling, cache scaling, RSS, and thermal stability using permitted development diagnostics.

This step is engineering evidence. It does not re-open old confirmation data or upgrade ledger-recorded score means.

2. Accelerate exact search without changing it

Test incremental board/features, compact canonical keys, sharded transposition tables, parallel complete games, cache-size sweeps, and profile-guided changes. A candidate speedup must match transition traces, chance identities, root values within the frozen numeric contract, work counts, and chosen actions. Otherwise it becomes a new algorithmic experiment.

The 128 GB machine makes large tables possible, but table size is an independent factor. Measure hit rate and end-to-end speed rather than treating memory use as progress.

3. Resolve AFBR-40 before implementation

AFBR-40 is only a proposed action-free public afterstate model with an intended multi-cycle target. There is no implementation, protocol, checkpoint, or result. Its first task is a read-only closure audit:

1. define the exact public state and afterstate keys;
2. enumerate every legal action at each candidate root;
3. enumerate required successors and long-outcome labels;
4. determine what existing retained artifacts can reconstruct;
5. report completeness by origin, phase, height, and legal-action count; and
6. list missing data without filling it from new seeds.

If closure is incomplete, close the audit as a valid negative result and design a new corpus. Do not call a sparse or played-action panel successor-closed.

4. Build action-complete multi-cycle data

Start with about 10,000 public roots, not a full production corpus. For each root:

• evaluate every legal sibling;
• use common event-keyed random scenarios across siblings;
• attach H40, H100, and H200 outcome distributions where budgets permit;
• record score, survival by rise, terminal hazard, clear/reveal flow, chain statistics, and logical work;
• store natural-game roots separately from curriculum/restart roots; and
• split training, calibration, and held-out data by whole origin game.

Measure label stability between independent scenario halves. Scale toward 100,000+ roots only if completeness, stability, throughput, and held-out ranking justify the storage and compute.

5. Train a distributional afterstate ranker

Use one shared evaluator on each legal afterstate so action identity cannot act as a shortcut. Start simpler than a transformer:

• categorical 7×7 public board encoding;
• next visible disc, rise phase, and legal mask;
• reflection-equivariant residual CNN or compact axial model;
• return quantiles or distribution bins;
• survival hazard by future rise;
• clear/reveal flow and terminal-probability auxiliary heads; and
• within-root pairwise/listwise loss in addition to scalar prediction.

Train on ROCm in large batches, but validate inference at every legal sibling. The first offline gate should compare top-1, top-2, pairwise accuracy, normalized regret, calibration, origin non-regression, and scenario-half action stability against fair D4. Freeze exact thresholds in the experiment record. The existing 477-root panel is reusable diagnosis, not fresh model-selection evidence.

6. Add the model above fair D4 conservatively

First compare only fair D4's top two near-tied actions. Override when a calibrated lower-confidence advantage exceeds a frozen threshold. Record coverage as well as accuracy; a policy that almost never acts has not solved the problem.

After a successful offline gate:

1. test the confidence-gated override on paired development games;
2. batch leaves across roots/games for GPU utilization;
3. test full-width D4 with the multi-cycle leaf; and
4. only then consider PUCT or progressive widening.

Do not begin with learned action pruning. Every legal root action remains eligible until the model has independently demonstrated safe sibling ranking.

7. Turn human strategy into testable structure

Humans describe long-term play in patterns: reservoirs, ladders, keeping a high covered disc reachable, protecting overlapping triggers, and releasing a chain at the right rise phase. Encode these as reflection-safe public features or five-move options:

• build a reachable reservoir;
• expose or crack covered discs;
• preserve trigger overlap;
• release stored chain energy; and
• stabilize before a row rise.

Use auxiliary labels for trigger reachability, occupancy debt, cover access, terminal hazard, and build/stabilize/release phase. Keep a feature or option only when a preregistered ablation improves disjoint-origin action ranking. Saliency or a visually convincing board is not evidence.

Compare an eight-cycle option search against a primitive H40 search at equal logical work. This directly tests whether human-like temporal abstraction makes long-horizon planning easier.

8. Revisit depth and stochastic tree search as a factorial — RUN, 2026-08-21

This item has been executed at development tier and is closed. The factorial varied depth (2, 3, 4, 5) against chance strata (5, 7), plus a factored ladder over reveal-sample count, on 64 paired games per cell using the bit-exact accelerated engine.

The measured findings, in three figures. Depth and chance resolution interact with a sign — the fourth ply pays only under the exact chance model; the fifth ply is not measurable by this design, which is a different claim from being worthless; and the two axes substitute rather than compound, leaving a budget frontier with a flat top at the fair-D4 operating point, reachable from either axis. The arms, contrasts and floors are recorded in RS-20260821T205102Z-d89df4b5 and RS-20260821T192140Z-189fe392 and analysed in finding-15 and finding-16.

Consequences for the rest of this roadmap. Priority 2's acceleration goal is met and no longer unlocks strength — it unlocks data rate, which is still valuable. Any future search change must be argued on a mechanism other than "more plies" or "more samples", because both axes are measured and flat. The remaining untested factors from the original list are the learned leaf inside the search (tried, negative) and transposition-table size (untested, and now low priority given that the table is 1.01x of a decision).

Deeper search is not merely "not monotonically stronger" — past ply four its effect is below what this program can measure. State it that way and not more strongly.

The binding methodological constraint, discovered here. Paired whole-game deltas have a standard deviation of 228,827 to 371,351 depending on the contrast, so a 64-game paired cohort has a one-sided 95% detection floor of roughly 47,000 to 108,000 points. Every significant result in this factorial is above its floor; every null result is below it and is therefore a non-measurement. Giving the observed depth-5 estimate a bound would take about 684 paired games — about 13 wall-days at that arm's measured throughput of 1,647 s per game on 14 threads (estimated range 8-18 days, since the two chunks differed 2.3x under other agents' load), and finishing the planned 64 would not have come close.

This changes how to choose experiments. Preregister the effect size the mechanism predicts and compare it to the floor before committing the machine. An experiment whose plausible effect is 20,000 points is not a cheap experiment; it is an unaffordable one. The alternatives are to pursue mechanisms with large predicted effects, or to build a lower-variance estimator than the complete game.

What a flat top implies for priority. When both estimate-quality axes are exhausted, the binding constraint has moved from the estimate to the objective being estimated. The two named candidates are in finding-15 §5: the terminal utility supplies no death-depth shaping, and the leaf is an uncalibrated potential. Those are one constant and one function. A terminal-utility sweep at depth 5 with seven strata separates them, and is the recommended next consumer of serious machine time — not another point on the depth or chance-resolution axis.

9. Iterate with every sibling, not only the played action

If the neural-guided policy passes:

1. play on new training-only origins;
2. evaluate every legal sibling at encountered roots;
3. add mature public restart states with independent future randomness;
4. freeze a new data manifest;
5. retrain for a preregistered number of rounds; and
6. keep origin-level validation folds unchanged across rounds.

This is action-complete dataset aggregation. It is intended to prevent the distribution shift that hurt behavior cloning and played-state value learning.

10. Qualification sequence

Use benchmarks.md: CHECK, bounded PILOT, paired SCREEN, reusable STANDARD, and fresh 256-game QUALIFY. Only a qualifying candidate is frozen with current transitive source, binary, model, corpus, compiler, command, and work hashes.

Then open the next permitted protected block and the one-shot final cohort only as the versioned protocol allows. Report the complete distribution and failures, not only the mean. The same unchanged policy must pass every gate.

Parallel tracks that do not consume strategy data

While the main candidate is training or running, agents can independently:

• audit data closure and manifests;
• add seed-free policy-boundary and reflection tests;
• improve result schemas, checksums, and resumption verification;
• profile compiler flags and simulator hot paths on synthetic fixtures;
• reproduce external algorithmic ideas in toy/no-seed domains; and
• review experiment arithmetic and claim scope.

These tracks increase confidence without competing for the same cohort. Their owners still write separate contribution records.

Decision rule

Prefer the next experiment that most reduces uncertainty about long-horizon sibling ranking per expected CPU/GPU hour. A larger model, deeper search, wider tree, or bigger table is justified only after its cheaper predecessor shows the specific signal it is meant to scale.