ProofAtlas research case study

Starting frontierBerlekamp's conjectured temperature ceiling 2AI-developed unverified manuscriptLean-checked rectangle-reachable witness at 33/16
Evidence and status →

Research paper

An Exact Counterexample to Berlekamp's Temperature-2 Conjecture for Domineering

Lech Mazur's manuscript gives a rectangle-reachable 28-cell Domineering position with exact game {17/8 | -2+*} and temperature 33/16 > 2. The attached Lean development kernel-checks the board–target equality, explicit target thermograph, 30-move replay, and resulting existential statement through its narrow HasValueTemperature interface; unrelated external replication and specialist review remain open.

The manuscript is only partially formalized. Some parts have been formalized, but this page does not present the whole manuscript as a Lean-checked result. Only separately identified formal results carry formal evidence. Hosting and editorial review do not establish the manuscript's mathematical claims as accepted ProofAtlas results.

The counterexample

Domineering is a partizan placement game: Left places vertical dominoes and Right places horizontal dominoes on a finite region. A long-standing conjecture attributed to Berlekamp asserted that every finite Domineering position has temperature at most 2.

Lech Mazur's manuscript gives an exact counterexample. Lean checks the board–target game equivalence, the explicit target's thermograph, the complete 30-move rectangle replay, and the resulting existential theorem through the development's narrow value-and-temperature witness. The stronger witness is a 28-cell position B' reached by an explicit alternating sequence of 30 legal moves from an 11 × 8 rectangle. Its exact game is

G(B') = { 17/8 | -2 + * },

so, in the nonnegative-temperature convention used in the paper,

t(G(B')) = 33/16 = 2.0625 > 2, with mean 1/16.

The Berlekamp collaboration page retains the broader research frontier and open follow-up tasks.

What the infographic shows

The infographic displays the actual 11 × 8 replay. Odd-numbered cyan dominoes are Left's 15 vertical moves; even-numbered coral dominoes are Right's 15 horizontal moves. The white cells and two gold cells are exactly the 28 cells left after move 30. Thus the graphic shows the moves as well as the final position; the coordinate table in the paper and Lean theorem are the certificates.

The 26-cell core and the repair

The connected 26-cell core B has the same exact value, but it has checkerboard imbalance 14 − 12 = 2. A rectangular board has imbalance at most one, while every played domino removes one cell of each color. Therefore B cannot itself occur as the remainder of play from any rectangle.

The repaired position translates B two cells to the right and adds the single cells (0,1) and (1,4). Both are isolated from every other remaining cell, so neither player can cover either one with a domino. Each is a zero component, and consequently

G(B') = G(B) + 0 + 0 = G(B).

The explicit 30-move replay proves the stronger rectangle-reachable statement.

Lean-checked scope

The frozen Lean source is identified by commit 9edd1d307a62bcd1b59c4450c8675fcdbb5a3ebe, Lean v4.30.0-rc2, and Mathlib commit 5450b53e5ddc75d46418fabb605edbf36bd0beb6. The final declaration is Domineering.berlekamp_conjecture_false in Domineering/ReachableCounterexampleValue.lean.

The machine-readable ProofAtlas evidence record is artifact.berlekamp-domineering-temperature-counterexample.reachable.v001. It is also included in the public packet as formalization/CHECKED_ARTIFACT.json.

Lean checks:

The development contains no sorry, custom mathematical axiom, or native-decision shortcut. Python output and file hashes are not theorem evidence.

The exact theorem uses the deliberately narrow predicate HasValueTemperature: it exhibits a thermographed representative game-equivalent to the board. The development proves that equality and the representative's thermograph, but does not yet include a general theorem that thermographs are invariant under every game-equivalence presentation. It also does not formalize the one-cell deletion audit or separately formalize the checkerboard obstruction for raw B.

Download and reproduce

The August 14 public artifact bundle contains the matching paper PDF and LaTeX source, the complete Lean source and generated certificate modules, Python verifiers and saved outputs, exact evidence tables, the final SVG and PNG infographic, toolchain pins, and a whole-bundle MANIFEST.sha256.

From the bundle root:

sha256sum -c MANIFEST.sha256
./RUN_ALL.sh
cd external_audit_2026-08-12
python3 independent_cgt_audit.py
python3 minimax_cross_check.py
python3 rectangle_reachability.py

To check the final Lean module from source, use the pinned toolchain and run lake env lean Domineering/ReachableCounterexampleValue.lean inside formalization/. The full generated certificate is included, but compilation is a substantial resource job; build products are deliberately omitted.

Additional exact evidence

The non-Lean routes separately rederive the canonical game and thermograph, check the raw-option inequalities and their unique equality cases, replay the rectangle sequence, and prove the value equality again using ordinary win/loss minimax on a 1,039,152-state difference game.

For every one-cell deletion from the 26-cell core, the exact computed temperature is below 2; the maximum is 15/8, attained at two cells. This is deletion-local evidence only. It does not prove global size minimality, uniqueness, or any minimality statement for the 28-cell reachable witness.

Review and priority boundary

The supplied evidence reports successful replays of the bundled Python routes. The checked-artifact record retains the successful cached Lean build transcript from the clean frozen source host, together with the no-sorry result, axiom profile, source commit, toolchain pin, and dependency-lock digest. ProofAtlas has matched the selected public files to the supplied manifest, but has not rerun the substantial Lean build for this page update. These checks were coordinated in the same author-directed workflow; they are not unrelated external replication or independent specialist peer review.

ProofAtlas does not award historical priority here. The manuscript reports that the literature checked through August 12, 2026 contains no earlier published Domineering position above temperature 2. A dedicated prior-art review and a reproduction in an unrelated system remain worthwhile. It also remains open how high finite Domineering temperatures can be and whether there is an infinite family above 2.

Publication here records the exact supplied paper, visual, public bundle, and scoped Lean evidence. It does not substitute for unrelated reproduction, specialist peer review, or an independent historical-priority assessment.

Documents and resources

Scope

What this page does not establish

Credit for this release

Contributors and roles

Authorship and other contributions are recorded separately for this version.

Authorship

Paper author
Lech Mazur

Contributions

  • Direct contributionOpenAI GPT-5.6 Pro

    OpenAI GPT-5.6 Pro has source-reported roles in mathematical exploration, exact computation, adversarial auditing, and exposition.

    • Computation
    • Exposition
    • Gap or error discovery
    • Proof strategy
  • Direct contributionLech Mazur

    Lech Mazur is the manuscript author and accountable editor and directed the rectangle-reachable witness, Lean formalization, audit, and exposition workflow.

    • Exposition
    • Proof strategy
    • Research direction

Selected reviews

Review record

Each entry names the review question, exact subject coverage, and reviewer provenance recorded for this release. A selected review concerns only its named question and is not by itself proof or acceptance of a broader claim. These bindings alone do not establish that proof or computation checks were rerun.

Download exact review bindings

Selected review

Public-wording review

Coverage
5 exact subject bindings
Reviewer
OpenAI Codex
Independence
Independence not asserted

Citation

Cite this release

Use the version-specific citation so the authorship, scope, and public record remain attached.

Release history

Versions and public record

Release
Paper release
Version
public version of 14 August 2026 with scoped Lean artifact
Published
Canonical page
Open canonical page
Earlier release
credit.release.berlekamp-domineering-temperature-counterexample.v001

Release chronology

This version is part of the release history. No independent external timestamp is claimed.

Scope of this release

  • No independent external priority anchor or literature-completeness review is recorded.
  • ProofAtlas has not rerun the substantial Lean build for this page update or reproduced the result in an unrelated implementation.
  • The Lean theorem does not formalize the deletion-local audit or separately formalize the checkerboard unreachability of the raw 26-cell core.
  • The Lean theorem uses the narrow HasValueTemperature witness interface and does not establish a general thermograph-invariance theorem for arbitrary game-equivalent presentations.
  • The scoped Lean theorem is not an accepted ProofAtlas result and does not substitute for specialist peer review.

Expanded visual

Open original image