Provenance and Evidence Boundaries

Sources used to reconstruct the project

This repository was organized from four evidence sources:

1. the copied TypeScript, C++, Python, model, manifest, and protocol files;
2. the long native experiment notebook, now retained as docs/research/history.md;
3. the referenced research task, read as a live task record; and
4. deterministic tests and builds run in this checkout.

The current directory did not contain Git metadata when the audit began, so there is no local commit history establishing when each copied experiment was created. The research task is not vendored here. Claims found only there are explicitly labeled task-record only.

Engine provenance

The copied TypeScript engine bytes initially matched the historical typescriptSha256 value in the frozen million-point protocol exactly. The native and TypeScript implementations also expose deterministic trace parity checks.

The reorganization changed paths, include statements, and purpose comments. These changes alter byte hashes even where executable behavior is unchanged. Current source files must therefore be tested directly and must not be presented as byte-identical to the historical snapshot.

Frozen protocols

Files under artifacts/protocols/ are archival records. Their serialized format identifiers retain the original drop7-...-v1 names because those are compatibility identifiers, not repository filenames.

The protocols contain source and manifest SHA-256 values from the earlier tree. Those hashes were not silently replaced during relocation. In particular, the optimistic phase n-tuple executable verifies transitive source bytes, and its historical lock is expected to reject the reorganized files. A resumed study needs a new protocol version and new current hashes; editing the v1 artifact would erase the distinction between the historical run and a new one.

Result artifacts

Many recorded experiments wrote checkpoints, corpora, and JSON summaries to temporary storage. Most of those artifacts are absent from this checkout. The detailed ledger records their paths and historical hashes, which helps identify the experiment, but a hash without the corresponding bytes is not independently verifiable here.

For that reason:

• local tests and parity are described as reproduced;
• expensive metrics in the ledger are ledger-recorded;
• immutable state in a retained protocol is protocol-recorded; and
• a statement present only in the task is task-record only.

Third-party prior work

Five copied page images were identified as rasterized pages from Erez Klein and Ben Friedmann's Drop7 Q-learning final report. The images had no bundled license or provenance notice and exposed personal contact details, so they were removed rather than redistributed. The original report remains cited as prior work in the strategy landscape.

David Walton's Sequence-mode solver write-up and source repository are also relevant prior work, but their known-future-disc setting is not comparable to this project's public-information Hardcore target.

Using this repository in research

When citing a result, include its evidence label, rules/scoring variant, cohort size, seed role, move cap, and whether the run was censored. Cite the source and protocol version separately. For a new experiment, retain its complete result artifact and transitive-source manifest inside a durable artifact store rather than relying only on a temporary path.