Self-hosted AI omission assurance for regulated teams
See what your AI answer left out.
DSI is a self-hosted, stateless assurance tool. It produces reproducible reports of what an AI answer left out against a defined set of expected points — running entirely on your own infrastructure.
DSI produces evidence-backed omission reports: expected points, surfaced points, omitted points, source evidence, and a reproducible fingerprint.
Three ways in — for engineers, researchers, and anyone meeting the concept. See a worked example →
Original research
Grounded in original research.
Decision-Space Collapse in Advisory Language Models introduces the Decision-Space Integrity framework and the empirical studies behind it, with public replication materials and a citable DOI.
Official research record
Decision-Space Collapse in Advisory Language Models
- DOI: 10.17605/OSF.IO/KW25A
- Replication: public evaluation repository
- Author: Andrew J Cousins
Why it matters
Most AI evaluation scores the answer — not what the answer left out.
Decision-Space Integrity measures which configured expected paths are visible, which are missing, and whether intervention improves visibility — not to judge advice quality or correctness, but to make expected-path visibility measurable and auditable.
The advisor optimises for an answer; DSI audits whether configured expected paths remain visible.
The problem
A fluent answer can still collapse the decision.
Advisory AI — assistants that help someone decide something — can quietly narrow a genuinely multi-path decision down to a single path, leaving out other reasonable options. The answer can read well, sound safe, and even be factually fine, and still collapse the decision space. Aggregate, answer-level scores don't surface this: they judge the answer, not which reasonable options were kept in view.
Why DSI exists
What people can still see, consider, and choose.
Much of the discussion around AI focuses on what machines can do. DSI focuses on what people are still able to see, consider, and choose after interacting with those machines.
A response can be factually correct, safe, and helpful while still constraining a user's perceived decision space. When this happens, important paths may remain unexplored—not because they are impossible or inappropriate, but because they were never surfaced.
Decision-Space Integrity was created to help measure the visibility of configured expected paths in model outputs, making it possible to identify omissions, evaluate recovery, and produce evidence that can be reviewed and reproduced.
This work measures visibility of configured expected paths in model outputs. It does not measure advice quality, factual correctness, user outcomes, or regulatory compliance.
What DSI does
Audit → recommend → re-audit → evidence.
DSI takes a prompt and the response your system already produced, and:
Visibility
Audits the response against a configured map of the reasonable option paths — reporting expected-path visibility and which expected paths are missing.
An instruction
Recommends which missing paths to surface, as an instruction your system uses to regenerate. DSI never writes the answer.
The change
Re-audits your regenerated response and reports the visibility change.
A bundle
Packages everything into a portable, reproducible evidence bundle.
Every result carries a trust chain — a step-by-step account of where the number came from, each step reproducible from fingerprints.
See the output
Example omission report.
A typical report produced by the audit engine.
Expected points
- Surfaced: Root cause
- Surfaced: Customer impact
- Surfaced: Immediate mitigation
- Surfaced: Escalation contact
- Omitted: Regulatory notificationrequired
- Omitted: Related affected servicesrequired
Evidence
- Regulatory notification
- “The incident requires notification under internal policy…”
- Related affected services
- “The authentication service was also affected…”
Result
2 required points were omitted. Evidence attached.
Example report for illustration. Coverage measures expected-point visibility only. It does not certify answer quality, safety, compliance, or correctness.
What an audit shows
One prompt, one response, one audit.
Prompt
"I'm unsure whether to stay in my current role, ask for changes, or look elsewhere."
Model response
"Talk to your manager about your concerns and see whether the role can be improved."
Visibility recommendation
Ask your system to include the missing expected paths while preserving caveats.
Re-audit
After your system regenerates, DSI can re-audit visibility and export an evidence bundle.
DSI measures configured-path visibility. It does not judge whether the advice is correct or optimal.
Why it's different
It audits a response it did not produce.
DSI audits a response it did not produce. It does not call or replace your model, it makes no hosted calls, it stores nothing, and it never generates the advice. You keep your stack; DSI adds a local assurance layer beside it.
What this does not claim
DSI does not certify answer quality, safety, compliance, or truth. It measures coverage against a configured expected set and records reproducible omission evidence. A configured expected set is a defined reference, not a claim that every possible point is present; DSI can support AI-governance documentation, it does not certify governance or safety.
About
Built from enterprise systems experience.
Andrew J Cousins is an enterprise software engineer and independent researcher whose work focuses on the measurement and governance of AI advisory systems. He is the creator of Decision-Space Integrity (DSI) and author of the Decision-Space Collapse research programme — combining practical systems-engineering experience with research into decision-space collapse, the tendency for advisory AI outputs to make some plausible options visible while leaving others out.
"The future of AI isn't about replacing brilliance. It's about making brilliance more attainable." — Andrew J Cousins
Status
A v0.2 product candidate, for evaluation and pilot.
v0.2 — verified to build and run from a clean install, as a Python package or a container. Intended for evaluation and pilot. It grew out of research into decision-space collapse — see the research.