Adding a rise-cycle residual to the evaluator

Bolt a small extra term onto the board evaluator — one that cares about where you are in the five-drop cycle before the next row rise — and see whether the policy plays better. The lasting product of this bench is not the extra term but the five frozen evaluator weights it left behind, which the reference search still uses.

The intuition

Drop7's clock is the row rise: every fifth drop, the whole board shifts up and a new row of gray discs appears underneath. That means a board is not simply "good" or "bad" — it is good or bad relative to where you are in the cycle. Four drops before a rise, a tall board is survivable and a slow-building plan is affordable. One drop before a rise, the same board is an emergency and the same plan is worthless.

The evaluator described on the fair leaf page has no such clock inside it. This bench adds a residual — a separate phase-aware score, added on top of the fixed fair evaluator rather than replacing it, so the base behaviour is preserved and only the correction is being tested. The residual used here doubles two phase-throughput terms — how ready the board is to trigger, and how ready it is to release — so the combined policy pays extra attention to keeping things moving as a rise approaches.

How it works, step by step

1. Compute the fixed fair evaluator's value for the position. Nothing about it changes.
2. Compute the phase-aware residual from the same public features plus the number of drops remaining before the rise.
3. Add the residual at one of three fixed scales — 0.25, 0.5, or 1 — chosen from a menu frozen before any game was played.
4. Use the combined value as the leaf of a sparse look-ahead search with a one-million-unit work limit, a 40,000-entry cache, a 500-move cap, and a flat −1,000,000 for a modelled game over.
5. Select the scale on two development games, then lock it and play a sixteen-game confirmation. Calibration only begins if the selected policy clears both preregistered gates: a 400,000-point score bar and a 5% improvement over the paired baseline.

Everything the policy reads is public: the visible board, the visible next disc, and the rise clock.

What happened, in plain English

Two things, and only one of them is a result.

The residual did not help. The ledger's account of the recovered evaluator is blunt: adding any tested phase residual broke the two long historical trajectories the fair evaluator had produced, cutting them from 155 and 160 moves down to somewhere between 55 and 110. That is why the native reference search deliberately excludes the residual and keeps only the base leaf. Note the size of that evidence: two games, on already-used training seeds, from the 7,000-point scoring era. It is a mechanism observation, not a measured cohort, and no held-out cohort for the residual is retained anywhere in the repository.

The bench's weights survived. In the course of this work the fair evaluator was pinned to five explicit overrides, exported from this directory as a named constant. The native depth-3 reference imports exactly those five values, and everything downstream of it — the four-move reference, and every experiment on the neighbouring pages — inherits them. The experiment index records this approach as completed — ledger-recorded, "it is part of the recovered D3 baseline lineage" — that is, the line of work that produced the three-move reference search — and that lineage is its real contribution.

The technical record

Source. approaches/fair-expectimax/phase-fair-combination/main.ts. Status: completed — ledger-recorded.

The exported constant. FAIR_PHASE_BASELINE_WEIGHTS overrides the laboratory's initial fair-policy weights with: direct potential 1,600, latent chain potential 700, height load −20, roughness 0, revealed-cover value 300. approaches/fair-expectimax/reference/fair-only-horizon.cpp recovers the fair-only leaf from fair-policy/tune.ts with these five overrides and "intentionally excludes the later phase residual" (ledger, "Historical fair-only horizon evaluator"). The 300-point revealed-cover term is a transition feature and is inert when the model is used purely as a leaf.

The residual. RELEASE_DOUBLE_PHASE_WEIGHTS doubles triggerReadiness and releaseReadiness on the default phase-horizon weight vector from src/core/typescript/phase-horizon-evaluator.ts. Residual scales tested: 0.25, 0.5, 1. Declared seed roles: training 0x1d70_0000, calibration 0x5d70_0000, validation 0x7d70_0000, reserved final 0xd700_0000. Defaults: 2 pilot games, 16 confirmation games, 16 calibration games, 500-move cap, 1,000,000 work, 40,000 cache entries, −1,000,000 terminal utility. Gates: REQUIRED_TRAINING_MEAN 400,000 and MATERIAL_IMPROVEMENT 1.05.

What is not retained. No cohort table, paired comparison, artifact hash or confidence bound for the residual itself appears in the ledger. The only recorded outcome is the sentence quoted above — that every tested residual shortened the two historical trajectories from 155/160 moves to 55–110 — inside the section describing the native recovery of the base evaluator, under historical 7,000-point scoring.

What this taught us, and what is still open

The negative half is a useful piece of the family's story: the rise clock is obviously important to a human player, and the most direct way of writing it into the evaluator — a bolted-on phase term — made the policy die sooner. One plausible reason is that the searches on these pages already see the rise: the look-ahead simulates the rise happening, so its consequences are already priced into the boards at the bottom of the tree, and an extra phase bonus double-counts them in the same way the restored placement rewards did. That is a hypothesis this repository has not tested directly.

The positive half is a caution about provenance rather than a scientific result. The five weights that the entire family depends on were pinned here, and their fitting protocol is not in the ledger; they should be treated as a well-tested starting point, not as a derived optimum.

Still open: whether phase information helps as an input feature the evaluator can weigh, rather than as a residual added to a finished score, and whether it matters more for a search too shallow to see the next rise than for one that can. The related rise-boundary depth experiment on the looking-five-moves-ahead page attacked the same intuition from the search side and was also stopped.

Sources

• main.ts — adds only a phase-throughput residual to the fixed fair-policy leaf; candidate selection uses two development seeds, then locks the exact scale for a sixteen-game confirmation, and calibration starts only after the selected policy clears both the 400k score and 5% paired-baseline gates.