Bringing back the old placement rewards

Before the look-ahead search existed, a simpler one-move policy scored each drop directly: where the disc landed, how close it was to completing something, how many covered discs it exposed, how deep a chain it set off. Those terms were dropped when the search was built. These two experiments put them back — first all of them, then only the two that looked most defensible — and both made the policy worse.

The intuition

There are two different places a policy can express what it wants.

• State terms describe the board after a move: how tall it is, what is ready to fire, how buried the gray discs are. The search's leaf evaluator is entirely made of these.
• Transition terms describe the move itself: this drop landed at height four, this drop revealed two covered discs, this drop set off a three-wave chain.

The one-move policy needed both, because with no look-ahead the only way to say "prefer moves that expose covers" is to pay for exposing covers. When the search arrived, only the state terms were carried over — and that is a real gap in fidelity, worth testing rather than assuming.

The argument for putting them back is that some of these are genuinely new information: whether a drop cracked a cover, or whether it produced a second wave, is an outcome you cannot read off the resulting board. The argument against is double counting: the resulting board already shows how tall the column is and how good the structure looks, so paying separately for the landing geometry charges the same thing twice — and can outvote it.

How each one works, step by step

All the terms. (1) Take the confirmed depth-3 fair search unchanged. (2) At every level of the search, add back the complete frozen block from the one-move policy: 300 points per revealed cover, 120 times the square of the continuation chain depth, and the original landing-height, column-content, build-distance, overshoot, neighbour-height, centre, one-away and imminent-rise terms. (3) Change nothing else. No coefficient is fitted or chosen from gameplay — they are the historical values, copied exactly.

Only the transition terms. (1) Remove every placement prior. (2) Keep only the two signals that describe newly observed outcomes: 300 per revealed cover and 120 times squared chain continuation depth. (3) The separation was argued for in writing before any gameplay: the placement terms duplicate geometry the leaf already sees, whereas cover damage and a multi-wave transition are new observations.

What happened, in plain English

All the terms together were clearly harmful. On the eight fresh screening games the unchanged search averaged about 134,900 points and 89 moves; with the historical block restored, about 94,800 points and 65 moves — roughly 40,000 points and 24 moves worse. The most damaging single game had the reference surviving 175 moves and the candidate 80.

The stripped-down version looked promising and then did not replicate. On a 24-game training-only diagnostic it improved the average score by 22,105 points and survival by 12.4 moves. On the fresh eight-game screen it lost about 17,000 points and 11.6 moves, improving only three of eight games on each measure, and in one game it threw away a trajectory the reference took to 145 moves after just 65.

Both rejections stand under either scoring mode: score and survival moved the same way, so rescoring the historical games to corrected Hardcore rules would widen the gap rather than close it.

The ledger draws a specific conclusion from the first result and a broader one from the second. The first shows that one-move action geometry cannot be transplanted wholesale into a deeper evaluator, because successor-state scoring already represents much of the same height and build information, so charging both can reverse good decisions. The second is described as an explicit false-positive record: rewarding the outcomes associated with a strong planner does not tell a public search which preparations produce them reliably.

The technical record

Sources. full-fair-horizon.cpp (all terms) and transition-reward-horizon.cpp (transition terms only). Experiment index: rejected — ledger-recorded, "neither transfer survived its fresh screen." Both were run under the historical 7,000-point Sequence scoring — the first section says so explicitly; the second carries no label and is classified from the score identity by audit 03.

All terms — ledger "Full historical fair action terms", fresh screen 0x3ea10000...007, eight games:

Arm	Mean score	Mean moves
Fair leaf, depth 3	134,923.875	89.125
Full historical block, depth 3	94,812	65.25

Paired change −40,111.875 points and −23.875 moves; three of eight score pairs and three move pairs won with one tie; neither cohort censored; worst case seed 0x3ea10002, 175 moves versus 80. Confirmation 0x3ea20000...00f not read. Artifact /tmp/drop7-full-fair-horizon.json.

Transition terms only — ledger "Fair transition rewards without placement priors":

Stage	Cohort	Reference	Candidate	Outcome
Training diagnostic	24 games 0x3d9a0000...017	—	—	+22,104.625 points, +12.375 moves
Fresh screen	8 games 0x3ea50000...007	113,772.5 / 78.125	96,759.625 / 66.5	−17,012.875 points, −11.625 moves; 3/8 score and 3/8 move pairs improved

Neither cohort censored; 0x3ea60000...00f remained unread. Artifact /tmp/drop7-transition-reward-horizon.json.

Verification. Optimised and ASan/UBSan strict builds passed deterministic completion, reflection, metadata blindness, legal-action, fixed-seed and public-state checks, plus — for the second experiment — an explicit action-prior exclusion test.

What this taught us, and what is still open

The durable finding is about representation, not coefficients: information the resulting board already contains should not be paid for twice. A deeper search sees the successor state, so restating the landing geometry as a bonus gives that geometry two votes and lets it outvote everything the search learned by looking further ahead.

The stripped-down follow-up is the more interesting failure, because its two surviving terms really are new information. It still failed — and its 24-game training gain, which reversed on eight fresh games, is a compact demonstration of why this repository will not accept a fitting result as evidence.

Still open: the same two signals have never been tried as features of the leaf rather than as payments per move, and the closely related question of paying per numbered clear or per reveal was tested separately, with the same shape of outcome — a strong fit followed by a held-out reversal (transition rewards).

Sources

• full-fair-horizon.cpp — evaluates the fixed action and transition terms inside depth-3 search; action terms charged once per decision, transition terms once per realised sampled move at every level. Gameplay selects no coefficient.
• transition-reward-horizon.cpp — disables placement priors while retaining the dense transition signals for revealed covers and squared continuation chain depth.