# Independent Replay Standard

This release distinguishes three operations that must not be conflated.

## 1. Derivation replay

The evaluator starts with an empty source cache. The replay fetches 36 exact public arXiv source archives, rejects any source whose SHA-256 differs from the lock, imports 49 portable research kernels, and generates fresh symbolic or numerical result payloads on the evaluator's machine. The compute path does not import, read, or accept the stored reference assertions.

## 2. Falsifier replay

The falsifier corrupts a source archive and requires hash verification to reject it. It also mutates a freshly generated result and requires post-computation comparison to fail. A source-level isolation check confirms that the compute function has no reference/expected-answer dependency.

## 3. Receipt verification

Receipt verification checks file identities and provenance hashes. It is useful integrity evidence, but it is **not independent scientific replay**.

Expected values are permitted only after computation as assertions. They are never inputs to a derivation. Public constants copied from cited equations are legitimate scientific inputs; stored final answers are not.

## Commands

```bash
python -m venv .venv
.venv/bin/python -m pip install -r requirements.txt
export OUROBOROS_SOURCE_CACHE="$PWD/clean-cache"
.venv/bin/python -m ouroboros_replay fetch
.venv/bin/python -m pytest -q tests
.venv/bin/python -m ouroboros_replay compute --output fresh-results
.venv/bin/python -m ouroboros_replay compare --actual fresh-results
.venv/bin/python -m ouroboros_replay falsify --actual fresh-results --temporary falsifier-work
```

On Windows PowerShell, set `$env:OUROBOROS_SOURCE_CACHE = "$PWD\clean-cache"` and use `.venv\Scripts\python` in place of `.venv/bin/python`.

Acceptance requires 36/36 source hashes, 49/49 derivation tests, 49/49 freshly generated payload matches, and every falsifier to pass. Network retrieval failures are reported separately from scientific failures.