---
title: Edge priority
family: heuristic-search
summary: Turn two claims from an experienced player into measurable evaluator changes by attacking high covered discs and distrusting edge columns.
status: unknown
evidence: repository-verified
reads: public
---

Test the two things an experienced human player actually says about this game
— attack the high covered discs, and distrust the edge columns, as measurable
changes to one term in the evaluator.

<EvidenceLabel status="unknown" evidence="repository-verified" reads="public" />

<Callout title="No retained result" tone="warn">
This lab is complete, gated and buildable, and **no outcome for it exists
anywhere in this repository**, not in the
[experiment history](/docs/research/history), not in the machine-readable
records, not as a task-record note. The
[experiment index](/docs/research/experiment-index) marks it *unknown*. This
page therefore describes what the program does, from the program. Nothing on
it is a claim about whether the idea works.
</Callout>

## The intuition

Nearly every strategy in this repository was invented by someone reading code.
There is one source of ideas that was not: the strategy notes of an
experienced human player, who recommends prioritising **high covered discs**
and watching the **less-connected edge columns** in this mode. The author
explicitly warns that the advice may be wrong or luck-dependent, and the
strategy catalog labels it *anecdotal hypotheses*, not evidence.

The two claims have a mechanism behind them that a program can express:

- **Altitude.** A [covered disc](/learn/glossary) that is high up is close to
  the ceiling and has few moves left before a [rise](/learn/glossary) pushes it
  higher. The evaluator already charges for that as the square of its height —
  so the question is whether the true penalty is steeper than square.
- **Edges.** A disc in column 1 or column 7 has only one horizontal neighbour,
  so there is one fewer direction from which a clearing disc can reach it. The
  evaluator already applies a 1.3× multiplier there, so the question is
  whether it should be larger.

The lab isolates exactly those two constants. Nothing else about the policy
changes.

## How it works, step by step

1. **Take the reference behaviour policy unchanged**: a bounded expectimax
   with the phase-safety evaluator, the same one used as the behaviour policy
   in other experiments.
2. **Replace one term of its evaluator**, the cover-potential term, with a
   parameterised version: penalty proportional to *height raised to the
   altitude exponent*, times an *edge multiplier* on the outer columns.
3. **Play six profiles on paired games**: the same seeds for all six, so the
   comparison is game by game:

   | Profile | Altitude exponent | Edge multiplier |
   | --- | ---: | ---: |
   | baseline (the unchanged reference term) | 2.0 | 1.3 |
   | edge2 | 2.0 | 2.0 |
   | edge3 | 2.0 | 3.0 |
   | edge4 | 2.0 | 4.0 |
   | altitude3 | 3.0 | 1.3 |
   | altitude3-edge3 | 3.0 | 3.0 |

4. **Screen on eight games, then confirm the winner on eight fresh games**,
   reporting paired score and move differences, clears and reveals per move,
   board clears and search work for each profile.

The design is a clean single-variable experiment: the edge rows isolate the
edge claim, the altitude row isolates the altitude claim, and one joint row
tests them together.

## What is known

That the program exists, that it builds, and that it checks itself. Its
self-test verifies that the baseline profile reproduces the unchanged
reference potential exactly, that its action matches the reference behaviour
policy's action, that decisions are deterministic and legal, and that a
mirrored board produces the mirrored column. It reads only the visible board,
the next disc and the rise clock.

What is not known is anything at all about the six profiles' performance.

<TechnicalDetails title="The technical record">

The [experiment index](/docs/research/experiment-index) records this as
**unknown, repository-verified**: *"no ledger or task-record outcome was
located."* Repository-verified means the source is present and its purpose can
be established from reading it; it is explicitly **not** a claim that a run
occurred.

Repository-verified from the source: `edge-priority-lab.cpp` defines six
`CoverProfile` entries with the comment *"Adam Saltsman's two rules are
isolated by the edge-only rows, the altitude-only row, and one joint row. The
reference profile is the internal control and remains unchanged outside this
executable."* Screen and confirmation cohorts are eight games each, from
`0x3d700300…` and `0x3d700400…` respectively, at a 500-move cap. Profile
selection prefers the larger paired mean score difference and breaks ties on
paired mean move difference. Output is per-profile JSON including every game's
score and move count.

Eight games is a small cohort. Even a completed run of this lab would be a
screen, not a confirmation of anything.

The anecdotal source is described in the
[strategy catalog](/docs/strategies), which links the player's notes and
labels them explicitly as hypotheses.

</TechnicalDetails>

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

- **This is the cheapest unfinished experiment in the family.** Sixteen games
  of a bounded search, two constants, a self-test that already passes. If the
  reader is looking for something small and real to do, this is it.
- **A negative here would be worth as much as a positive.** The strategy
  catalog's sixth open direction is precisely to convert human strategy advice
  into reflection-safe public features and test them, because that is more
  informative than copying a human rule into a policy. Either answer settles
  something.
- **The related terms are scattered.** An edge multiplier of 1.65 appears in
  the `fertile` evaluator inside
  [policy comparison](/approaches/heuristic-search/policy-comparison), and an
  edge-specific cover penalty appears in
  [tunneling](/approaches/heuristic-search/tunneling). Neither has been
  isolated to a gate. This lab is the one that would do it.