73
Contract checks
Reproducible technical evidence
Safety and quality are tested as executable contracts.
A local, deterministic evaluation suite measures CasePilot's structured analysis, refusal behavior, privacy masking, authorization gates, bilingual consistency, and verified-fact grounding — with no API key or network request.
All current checks passedcasepilot-deterministic-contract-v1DeterministicDemoProvider
73
Passed
14
Analysis fixtures
0
Network calls
Evaluation suite
Measured reliability dimensions
100%
Schema-valid structured output14/14
Expected missing fields and risks14/14
Sensitive-request refusal14/14
Deterministic replay14/14
English / zh-TW behavioral parity3/3
PII masking4/4
Authorization and dispatch gates6/6
Grounded, human-reviewed Growth Pack4/4
This is a reproducible contract benchmark, not a claim of live-model or real-world accuracy.
How the evidence works
1. Synthetic golden set
English and Traditional Chinese intake samples, ambiguous requests, six sensitive-request categories, and four PII patterns define expected behavior without customer data.
2. Deterministic graders
Exact schema, classification, replay, state-transition, privacy-boundary, bilingual, grounding, and publication checks produce the score shown above.
3. Safe expansion path
Owner-approved work material can contribute generalized facts to new synthetic fixtures. Raw customer media and identifiers stay outside this suite.
What this evidence does not claim
The public judge experience runs DeterministicDemoProvider. These results do not claim live GPT-5.6 runtime accuracy, production conversion lift, real-customer validation, or permission to publish generated content.