---
title: Gray throughput
family: heuristic-search
summary: Clear numbered discs and open covered ones fast enough to keep the board from filling.
status: rejected
evidence: task-record only
reads: public
---

Stop trying to score and start trying to keep up — clear numbered discs and
open covered ones fast enough that the board never fills.

<EvidenceLabel status="rejected" evidence="task-record only" reads="public" />

## The intuition

There is a piece of arithmetic in this game that has nothing to do with
strategy. A board has 49 cells. Every five drops, five numbered discs land on
it *and* a [rise](/learn/glossary) pushes in a row of seven covered discs.
Twelve discs arrive per cycle, so a board that is not slowly filling up must
remove twelve per cycle too:

<Stat label="Numbered discs that must clear" value="2.4 per move" hint="12 arrivals ÷ 5 drops" />

and every one of those seven covered arrivals has to be opened before it can
be cleared at all:

<Stat label="Covered discs that must be revealed" value="1.4 per move" hint="7 covers ÷ 5 drops" />

This policy is that arithmetic turned into an objective. It does not ask "how
many points is this move worth"; it asks "does this move keep me at, or above,
the replacement rate", and it pays a **surplus** bonus for clearing or
revealing more than the requirement, not merely for clearing or revealing.

## How it works, step by step

1. **Read the position** — board, next disc, rise clock.
2. **Simulate each legal column** over a few stratified guesses about what the
   hidden gray numbers would show, then continue for a short horizon.
3. **Score the resulting board** with thirteen state features: occupancy,
   total covers, solid and cracked counts, how much cover load sits high, load
   above the safe band, peak and mean height excess, pressure from the coming
   rise, low-number caps and adjacent low caps, and two topology terms for
   whether covers are exposed and numbers are still live.
4. **Score the move** with seven transition features, dominated by the two
   surplus terms — clears above 2.4 per move and reveals above 1.4 per move.
5. **Blend the immediate value with the continuation value** and play the best
   column.

The companion `tune.ts` fits the twenty weights with a cross-entropy method on
complete games from a training seed range, then replays the frozen champion on
a separate calibration range.

## What happened

Pushing harder on gray discs made the board look better and did not make the
policy live longer. The retained sentence is that stronger gray weights
improved *exposure* (how much of the covered population was attackable) but
not the sustainable rate of cracks and reveals, and not the score.

That is a more interesting failure than it first appears. It says the deficit
is not caused by the policy failing to *value* reveals highly enough. Wanting
reveals more does not produce more of them.

<Callout title="This policy reads two things a deployable policy may not" tone="warn">
The chance-sample salt in `gray-throughput-policy.ts` hashes the game's
**level** and **moves played** along with the board and rise clock. The
research contract excludes both from the deployable public interface, and the
benchmark registry flags this policy accordingly
(`publicInformation: false` in `src/bench/policies.ts`) so that it is shown
separately on the [leaderboard](/leaderboard). Neither field can leak a hidden
disc value or a future disc (they only vary the deterministic sampling) but
the flag is correct and the policy as written could not be submitted for
qualification without changing that salt.
</Callout>

<TechnicalDetails title="The technical record">

The [experiment index](/docs/research/experiment-index) records this as
**rejected, task-record only**: *"stronger gray weights improved exposure but
not sustainable cracks/reveals or score."* Task-record only means the result
was reported in a research conversation and never promoted into the
[experiment history](/docs/research/history). **No cohort size, score, or
measured clear/reveal rate for this policy is retained in this repository**, so
none appears on this page.

The 2.4 and 1.4 targets are arithmetic, and they are in the source:
`REQUIRED_CLEAR_THROUGHPUT = 1 + BOARD_SIZE / 5` and
`REQUIRED_REVEAL_THROUGHPUT = BOARD_SIZE / 5` in
`src/core/typescript/gray-throughput-policy.ts`, with the same two constants
repeated in `phase-horizon-evaluator.ts`. The independent measurement of how
far the reference policy falls short of them —  1.973 clears and 1.090 reveals
per move over 64 games — is exploratory, development tier, in
[finding 01](/docs/exploratory/finding-01-score-is-survival), and the
[research status](/docs/research/status) page cautions that these rates are
*"diagnostic targets from limited runs, not proven universal thresholds."*

Repository-verified defaults: thirteen state weights ranging from +520
(exposed cover topology) to −2,400 (peak height excess), seven transition
weights led by +650 (clear surplus) and +620 (revealed covers), and a
continuation blend of 0.72. `tune.ts` runs a cross-entropy method over all
twenty parameters on the `0x1d70…` training range with a separate `0x5d70…`
calibration range and a bootstrap over game results; `benchmark.ts` compares
the gray, rollout, combined and tunneling profiles on the same games.

</TechnicalDetails>

## What this taught us, and what is still open

- **Naming the right objective is not the same as reaching it.** This lab
  encoded the correct conservation law and still could not satisfy it. The
  later exploratory work reached the same wall from the other side: no legal
  policy tested has achieved flow balance
  ([finding 06](/docs/exploratory/finding-06-flow-ceiling), which carries two
  published correction notices and should be read with them).
- **Exposure is a decoupled measurement.** Making covers attackable and
  actually converting them into reveals came apart here. Any successor should
  report both, because improving the first is easy.
- **What was ruled out** is this weight family, tuned this way. The
  conservation objective itself is now the organizing idea of a separate
  research line, [lifetime objective](/approaches/lifetime-objective).