Partizan reproducibility materials

This repository contains the public experiment code, preregistered protocols, schemas, and compact verification authorities for the Partizan fixed-value representation study.

The study asks what varies among combinatorial-game realizations after exact game value is held fixed. Its main experiments compare equality-only, novelty-guided, and structural acquisition in bounded Digraph Placement, then test the same fixed-value search principle in Domineering. A matched five-arm ablation separates learned novelty from encoder initialization and canonical graph distance.

Repository boundary

This Git repository contains code and compact authorities. It deliberately excludes the manuscript, rendered paper, private submission notes, model weights, and large event or certificate ledgers. The complete immutable primary-policy, historical-transition, and Domineering-scope evidence is available from the archival deposit record, with deposit details in docs/research/FIXED_VALUE_ARCHIVAL_DEPOSIT.md. The Partizan development record contains the frozen five-arm ablation protocol, aggregate metrics, compact authorities, independent verification, and the complete matched-ablation ledger: all 368,640 exact-verifier calls across 180 streams with proposal decisions, event chains, certificate sidecars, and stream metrics.

The reusable game engine and public visualization live in the main Partizan repository.

Interactive atlas

The fixed-value atlas shows all 21,697 quotient-distinct graph forms retained by the historical order-7 study. It lets a reader group the same observed sample by complete-game identity and exact value, inspect a 193-form equivalence class, and review a performed edit study. In that study, the author toggled every one of the 42 loop-free directed arcs from a mechanically selected form; exact replay retained 33 value-1/2 variants with 33 distinct graph quotients. The deployment is built from a pinned Partizan commit, and its evidence files are hash-bound to the frozen protocol and independent verification.

Layout

• scripts/research/: generators, exact verifiers, replay tools, corruption tests, scope evaluation, and archive builders.
• docs/research/: preregistrations, protocol amendments, result summaries, schemas, the evidence index, and compute accounting.
• output/release/: compact archive and replay authorities.
• output/research/: compact terminal authorities, including the public atlas evidence and matched novelty ablation. Large ledgers are supplied by Zenodo.

Environment

Python 3.11 or newer is recommended.

python -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt

PyTorch is required only for neural-model reconstruction. Exact replay and the lightweight test suite use NumPy and the Python standard library.

Lightweight checks

PYTHONPATH=scripts/research python -m unittest \
  scripts/research/test_validate_digraph_order7_diversity_policy_protocol_v3.py \
  scripts/research/test_fixed_value_descriptor_atlas_v1.py \
  scripts/research/test_stage_and_verify_digraph_order7_diversity_policy_v3.py

The fixed-value transition replay test requires content-addressed Stage-0 sidecars from the full evidence archive. Run it as part of the full replay after extracting that archive.

Publication and atlas bindings can be checked without downloading the full archive:

python scripts/research/test_publication_artifacts_v1.py

The release-builder and Figure 4 provenance checks require Pillow and ReportLab. Rebuilding the paper rendering also requires Poppler's pdftoppm executable (poppler-utils on Debian/Ubuntu). Install the Python dependencies from requirements.txt, then run:

PYTHONPATH=scripts/research python -m unittest \
  scripts/research/test_build_fixed_value_submission_bundle_v1.py \
  scripts/research/test_fixed_value_atlas_figure_authority_v1.py \
  scripts/research/test_build_fixed_value_paper_repertoire_figure_v2.py

Full replay

1. Download both files attached to the Zenodo record.
2. Verify their SHA-256 values against the Zenodo description and output/release/FULL_EVIDENCE_ARCHIVE_AUTHORITY.json.
3. Extract the full evidence archive at the repository root.
4. Run the commands in docs/research/EVIDENCE_INDEX.md and docs/research/SUBMISSION_EVIDENCE_AUDIT_V1.md.

The scientific audit passed without scientific failures. The publication receipt records the completed archival action and changes the public release state from blocked to ready. The current manuscript is deliberately excluded from this repository, so its final PDF hash is audited in the private submission workspace.

The original compact and full archives are preserved at Zenodo. The atlas-inclusive release is attached to fixed-value-evidence-v1.1.0. The fixed-value-evidence-v1.2.0 release adds the five-arm ablation protocol, analysis lock, result, replay code, and compact authorities. The fixed-value-evidence-v1.2.1 release adds the atlas and paper-figure provenance, the complete 42-event authorial demonstration, and a self-checking deterministic archive. Its archive has 290 members, is 89,508,431 bytes, and has SHA-256 52e07f2a0f6cd3d62a97b939bb084d242f8019cd323a94c213ec8709e8afe527. These GitHub releases complement rather than replace the Zenodo records.

Matched-ablation ledger

Development-record version 10.5281/zenodo.21881253 adds the complete 951,988-file matched-ablation ledger. Verify the downloaded Zstandard archive against its .sha256 file and NOVELTY_ABLATION_PUBLIC_ARCHIVE_AUTHORITY_V1.json; the separately deposited manifest binds every source path, byte count, and SHA-256 digest. The archive verification authority records a full member-by-member read-back with zero forbidden-content hits and no V5 test material.

After downloading the five matched-ablation files from Zenodo, verify and extract the ledger at the repository root:

sha256sum -c partizan-novelty-ablation-ledger-v1.tar.zst.sha256
zstd -dc partizan-novelty-ablation-ledger-v1.tar.zst \
  | tar -xf - --strip-components=1

The extraction restores the original output/research/ tree. With the neural dependencies installed, replay every proposal decision, exact outcome, memory update, stream metric, paired contrast, and corruption check with:

PYTHONPATH=scripts/research python \
  scripts/research/verify_digraph_order7_novelty_ablation_v1.py \
  --phase test --repo-root .

The expected terminal verification artifact is 69dc34ec8757fac0e7480edd821dbdf69c1963d4d3ab41a4db3cac4537bd2063.

The complete 42-edit authorial demonstration is also small enough to live in this Git repository. Its protocol, event ledger, certificate sidecars, replay code, and terminal authorities are under output/research/digraph-order7-composer-one-arc-v1-c0d8f47996a7 and scripts/research/.

Citation and license

Citation metadata is in CITATION.cff. The versioned evidence DOI is 10.5281/zenodo.21833142. Code, generated data, documentation, and newly rendered figures are licensed under GPL-3.0-or-later. See LICENSE.