Positioning
Applications
DSI audits a supplied AI answer against a defined set of expected points and reports what was omitted, with source evidence and a reproducible fingerprint.
Example application
AI summary assurance.
Compare a generated summary against a defined set of expected points and report the omissions — each with source evidence and a reproducible fingerprint. Under the same configured map, pack, and instrument settings, the same input produces the same report.
Example application
Example policy assurance report.
Comparing an AI-generated policy summary against expected policy obligations.
Expected policy obligations
- Surfaced: Password rotation
- Surfaced: MFA requirement
- Surfaced: Logging requirements
- Surfaced: Account lockout
- Surfaced: Privileged access
- Omitted: Third-party access reviewrequired
Evidence
- Third-party access review
- “Access granted to external contractors must be reviewed every 90 days.”
Result
1 required policy obligation was omitted. Evidence attached.
Example report for illustration. Coverage measures expected-point visibility only. It does not certify answer quality, safety, compliance, or correctness.
Also supported
Advisory AI.
Audit an advisory model answer against a configured set of expected points and report which were surfaced, which were omitted, and the source evidence. Example domains include career, finance, and relationship decision-support.
Current
Also supported
Regression comparison.
Compare two sets of paired outputs — for example a baseline and a candidate over the same prompt set — and report which expected points changed in visibility between them, under one fixed instrument. Useful when a model version, a prompt, or a configuration changes and you want reproducible evidence of what moved.
The comparison is instrument-relative: it reports a change in expected-path visibility, not that one output set is better, safer, or of higher quality.
Current
How it's used
Offline, shadow, supervised.
DSI is run offline, or as a shadow alongside an existing system — it observes and records; it does not sit in the response path, and it does not block, route, or enforce. Flagged results are for human review.
Important scope note
The current implementation audits supplied outputs against configured expected points — for advisory answers, and for expected-point documents such as policy summaries — and compares paired outputs under one fixed instrument (Regression Audit). The classifier is challenge-tested; human review is recommended.
This work measures visibility of configured expected paths in model outputs. It does not measure advice quality, factual correctness, user outcomes, or regulatory compliance.