Preserve and correct
Five fictional documents are hashed, including an exact copy. The checked answer changes from 3 retries to 2 while the earlier source remains represented.
Consumption Engine: public fixture replay
This package gives a reviewer a small, deterministic replay of the evidence contract: preserve synthetic sources, carry a correction forward, refuse an earlier snapshot, and inspect the preserved history after archive restoration.
Public synthetic fixture. Python standard library. No network after cloning. No model calls
Run it from a fresh clone
The replay uses Python 3.11 or newer and the standard library. It archives the completed synthetic task, moves the original generated state out of the active path and restores the archive into a fresh directory on the same machine.
A guarded subprocess checks the restored history, source hashes and current or stale task state. The Python read and socket guards cover this fixture; they do not provide operating-system isolation or prove recovery on a clean machine.
git clone https://github.com/thisisntjon/ce-stranger-replay-v1.git
cd ce-stranger-replay-v1
git checkout 8374bb6ee93edef37f57f284532d9487d5fd2e2d
python examples/public_demo/run_public_demo.py --output .tmp/public-demo-run
python examples/public_demo/verify_receipt.py .tmp/public-demo-run/RECEIPT.json --output .tmp/public-demo-run/ORACLE-DECISION.json Read the README and runner source on GitHub. The runner reports passed: true. The receipt checker reports accepted_automated_oracle when all six integrity/projection checks pass. Keep the generated RECEIPT.json, REPORT.md and ORACLE-DECISION.json for review. Use a new output directory for another run; the runner refuses to overwrite an existing one. The pinned source is tagged v1.0.0-fixture, and the repository’s CI runs the same replay and receipt check on Linux, Windows and macOS. CI shows the fixture replays across operating systems; it is not an independent reproduction.
What it checks
Five fictional documents are hashed, including an exact copy. The checked answer changes from 3 retries to 2 while the earlier source remains represented.
The lifecycle executes correction and stale-snapshot checks. The broader oracle projection includes manually specified refusal and UNKNOWN states. Matching that projection checks the declared contract; it does not prove a fresh query for every case.
The release README predates this clarification. Read the corrected explanation of executed behavior and declared labels. The commands above still select the same released code.
The archive manifest is verified and the original generated state is renamed. Checks then inspect preserved history, source identity and current or stale task state in the restored copy.
Qualification boundary
The public runner makes the published synthetic evidence contract repeatable. It is not the private Consumption Engine runtime. It does not assess semantic truth, automatic claim extraction, model quality, general retrieval quality, production deployment, clean-machine recovery or customer value.
Current result: the released runner checks the lifecycle and restored artifact integrity. The six receipt checks cover exact input identity and the declared oracle projection.
Next for the wider research: Evaluate a concise briefing of current facts, decisions and next steps against ordinary source reading, including preparation and review effort. This does not change what the fixture demonstrates. Private-runtime distribution and clean-machine recovery remain separate qualification work. Read the current product story.
Historical record
The September 24 simplified replay result records an earlier standalone runner. It remains available with its original contents.
The commands above run the pinned public package and generate a new RESULT.json, REPORT.md and RECEIPT.json. Use those files and the receipt check to inspect your run.
Format, size and a short SHA-256 identifier, measured when this page was built.