Cover-throughput probe

A test of one tempting instinct: gray discs are what eventually kill you, so tell the player to attack them much harder and it should survive longer. Turned up over four increasing settings, it barely changed the outcome.

The intuition

Every five drops a new row of gray discs is pushed in underneath. A gray disc carries a hidden number, contributes nothing until it is opened, and takes up a cell in the meantime. A player that never opens them fills the board and dies; a player that opens them keeps the board turning over.

So: take a hand-written evaluator and multiply the terms that reward attacking covers — points for cracking, points for revealing, points for reducing the cover count, penalties for covers that are high up or hard to reach, rewards for having a trigger ready when the next rise arrives. Then play complete games at four increasing intensities and see whether the score curve rises with the dial.

The instinct is not silly. It is the same intuition the flow arithmetic supports: the board demands roughly 1.4 reveals per move to stay level. The question this probe asks is narrower and worth separating out — does weighting the immediate, local cover reward more heavily produce that flow?

How it works

1. Start from the hand-written weight vector that the evolutionary experiments use as their reference point.
2. Build four variants by scaling about a dozen cover-related weights with a single intensity parameter at 0.25, 0.50, 0.75 and 1.00. Nothing else about the player changes.
3. Play the same 256 complete games with each of the five profiles, from a fixed, previously used exploratory seed range, with a 1,000-move cap and seven imagined chance outcomes per decision.
4. Report mean, median, 10th and 90th percentile, minimum and maximum score, mean moves, and the censored count for each profile.

Everything read is public. The output is explicitly stamped "selectionEligible": false: the file states in its own header that it is not a validation benchmark, and it names the seed range it reuses.

What happened

The dial did almost nothing. In the recorded reading, the best of the five profiles was nearly flat against the others, so larger local cover weights were not promoted into any policy.

The honest reading is narrow. This rejects one mechanism — scaling up immediate, one-move cover rewards inside a shallow hand-written evaluator — and it does not reject the underlying idea that reveal flow governs survival. A policy can want more reveals per move and still be unable to get them by paying more for each individual crack, because the crack has to be available, and making it available is a structural problem several moves earlier. That is the same conclusion the constructive and reservoir family reaches from the other direction.

The technical record

Status: rejected — task-record only. The experiment index records: "the strongest profile was nearly flat, so larger local cover weights were not promoted." No per-profile numbers, protocol, artifact hash or per-game data survive, and audit 04 §B.3 lists this source among the 30 whose rejection "cannot be re-derived". Treat the verdict as provisional in exactly the way the index instructs; this page quotes no scores because none are retained.

Configuration, read from the source: 256 games, 1,000-move cap, 7 chance probes, seed start 0x3d700000 (also catalogued in audit 03); profiles hand, cover-0.25, cover-0.50, cover-0.75, cover-1.00. The weights scaled are immediate clears, immediate reveals, immediate crack progress, occupancy reduction, cover reduction, cover load, cracked load, cover and cracked altitude, cover exposure, accessible cracked, stored high numbers, direct and release readiness, next-drop and rise triggers, projected occupancy debt and phase height risk. It embeds approaches/heuristic-search/evolution/evolution.cpp for the policy and game loop, and prints the engine's level-bonus constant into its own output so the scoring mode of a run is self-documenting.

Because it reruns a previously evaluated exploratory range, a fresh run of it is a re-reading of spent data, not new evidence.

Source: throughput-probe.cpp.

What this taught us, and what is still open

• Paying more for each individual crack or reveal does not buy more of them. Availability, not price, is the binding constraint.
• The rejection is one configuration on one shallow evaluator, recorded only as a task record. It says nothing about cover-focused terms inside a deeper search.
• The same family's trajectory throughput study attacks the same question from the measurement side, by asking what clear and reveal rates a much stronger — and privileged — player actually sustains.