Skip to content
grid-diamondcutter

Audit walkthrough

This page is for readers verifying the claims, not exploring the methodology. Allow about twenty minutes.

Everything on this site is computed from the repository

This site is a static rendering of files in github.com/eir-inc/grid-diamondcutter. It has no database and no server. Every number you see on this site was computed at the moment the site was built from a specific commit. The commit hash is shown in the header of every page.

To verify the measurement counts

The measurements page shows N measurements with their pass/fail status. To verify:

Terminal window
git clone https://github.com/eir-inc/grid-diamondcutter
cd grid-diamondcutter
find examples/voices phase_2/examples/voices -name '*.sidecar.json' | wc -l

The number should match.

To verify any single measurement

Pick a measurement on the measurements page, follow the link to the source. The Python file is the code that ran. The JSON file alongside it contains:

  • The prediction the measurement made.
  • The kill condition.
  • The exact code path that ran.
  • The result.
  • A SHA-256 hash of the first four items.

To verify the hash:

Terminal window
python -c "
import json, hashlib
data = json.load(open('examples/voices/YOUR_VOICE.sidecar.json'))
unit = {k: data[k] for k in ('voice_name', 'prediction', 'kill_condition', 'run_protocol')}
canonical = json.dumps(unit, sort_keys=True, separators=(',', ':'))
print(hashlib.sha256(canonical.encode('utf-8')).hexdigest())
"

That hash should match the sidecar_sha256_pre_verdict field in the JSON. If it matches, the prediction was not edited after the result came in.

To verify a historical event recovery

The historical events page names three Phase-A events (Texas 2021, EU 2022, Japan 2011) and three Phase-2 events. Each event has a source file listing the public reports the model was calibrated against. The reports are real and the citations are verifiable. The model’s recovery of the event is computed by the code; you can re-run it.

To verify the boundaries are not being crossed

The boundaries page lists thirteen claims the project has committed not to make. Each boundary has at least one checker — a small Python file in examples/voices/ whose name starts with bound_defender_. Each checker scans the repository for the forbidden claim. If it finds one, the checker fails.

To run them all:

Terminal window
python -m pytest tests/test_voice_registry_contract.py

What this is not

This is not third-party certification. The project audits itself according to commitments it has written down. A reader who wants independent certification needs to bring their own.

The project’s full list of things it commits in writing not to claim is on the boundaries page.