Production AI
Watchman Shepherd RAI Framework AI Innovation Model
Services About Research
Live Demo

Watch a real audit.

Every line below is from an actual Watchman run. The reports are the real output, unedited except for the removal of environment-specific file paths. No marketing mock-ups. This is the tool.

Overview Compliance Live Demo Talk to Us

An audit, start to finish

Auditing a real instruction-tuned model against the base it derives from. Replayed at the audit's true timing: start to verdict in under six seconds for a 1.5-billion-parameter model on a single workstation.

watchman: model provenance audit
0.00s

Three real reports

Exit 2 · Detected

Modification detected

A full-precision model audited against its base. Watchman flags the modification, classifies it as instruction tuning, and shows the change concentrated in the attention value and key projections at mid-depth.

Open the report
Exit 0 · Clean

Clean quantized release

A compressed model audited against its full-precision base using a matched-compression control. Watchman confirms the release is the base plus declared quantization and nothing more.

Open the report
Exit 1 · Fails safe

Wrong claimed base

A candidate audited against a base it does not derive from. Rather than guess, Watchman returns indeterminate, never a false "clean". The verdict you act on is one the tool can defend.

Open the report

The AI-BOM attestation

Alongside every report, Watchman emits a CycloneDX-style attestation fragment: the model's identity, its per-file weights identifier, the verified lineage to its base, and the verdict. Merge it into your AI bill of materials; pin it in your model registry. This is the real output of the run above.

attestation.cdx.json Download 
loading…

Run it in your own pipeline

Watchman deploys as a container inside your environment. Submit an audit over REST, poll for the verdict, pull the report and attestation. Exit codes make it a one-line CI gate. (The hostname below is illustrative; your deployment is private to you.)

# submit an audit: is this candidate what it claims to be?
curl -X POST https://watchman.internal.example/v1/audits \
  -H 'Content-Type: application/json' \
  -d '{"base":"org/base-model","candidate":"org/candidate-model"}'
# → {"id":"aj-1a2b3c4d5e","status":"queued"}

# poll for the verdict
curl https://watchman.internal.example/v1/jobs/aj-1a2b3c4d5e
# → {"status":"done","exit_code":2,
#    "result":{"modification_detected":true,"classification":"instruction_tuning"}}

# pull the evidence
curl -O https://watchman.internal.example/v1/reports/aj-1a2b3c4d5e.html
curl -O https://watchman.internal.example/v1/reports/aj-1a2b3c4d5e.cdx.json

# CI gate: exit 0 clean · 2 modification detected · 1 indeterminate

Run it on your own models.

Book a live session and we'll audit a model of your choosing, walk through the evidence, and scope a deployment inside your environment.

Book a Live Demo