Reproducing results

Reproducing results

Choose the claim you want to check

Claim	Check
Evaluator behaviour	Run the regression suite and synthetic demonstration
Archived structural counts	Replay the frozen checker over retained responses
Generated artifact consistency	Run the public-artifact drift check
Semantic correctness	Requires independent source-based annotation, not just code execution

Follow Quick start to prepare an isolated environment and pin a release. The released regression record has 42 tests.

Offline historical replay

From the repository root, with the environment active:

python -m pip install -r evidence/development-2026-11-10/requirements.txt
python evidence/development-2026-11-10/replay.py

The retained release result has recorded_attempts: 26, replayed_attempts: 26, differences: [] and counts_match: true. This makes no model calls. Do not run the archived runner to repeat live inference; it is guarded and retained for inspection.

The archive contains parsed responses, requests, configurations and a retained checker snapshot. These are not original HTTP wire bytes or proof of a bit-identical original runtime. Read the archive disclosure.

Public and wiki artifacts

On the current main documentation revision:

python tools/build_public.py --check
python tools/check_wiki.py

The first checks generated copies and packages. The second checks wiki navigation and Markdown source links. A local check does not prove deployment: HTTPS readback and GitHub publication are separate evidence.

For an audit, state the commit, environment, command, expected result, actual result and smallest public or synthetic reproducer. Contributing | Release evidence