DSI · decision-space integrity

The concept, defined

What is Decision‑Space Integrity?

Decision-Space Integrity (DSI) is a method and self-hosted software for measuring which configured expected paths are visible, missing, or recovered in an AI model's output. It audits a response the model already produced, against a defined set of expected points, and records reproducible omission evidence.

The problem it addresses

What is decision-space collapse?

Decision-space collapse is the tendency of advisory AI — assistants that help someone decide something — to narrow a genuinely multi-path decision down to a single path, leaving other reasonable options unstated. The answer can read well, sound safe, and be factually fine, and still leave reasonable options out of view. Aggregate, answer-level scores judge the answer; they do not surface which reasonable options were kept in view.

The advisor optimises for an answer; DSI audits whether configured expected paths remain visible.

How it differs

Why this is not hallucination, bias, or ordinary evaluation.

DSI measures a different property from the tools teams already run. It is complementary to them, not a replacement.

Not hallucination detection

Hallucination detection asks whether what was said is true. DSI asks whether what should have been in view is present. A response can be entirely accurate and still omit reasonable options.

Not bias measurement

Bias work asks whether outputs differ unfairly across groups. DSI asks, for a single answer, whether the configured expected paths were surfaced — a question about coverage, not fairness.

Not benchmark scoring

Benchmarks give an aggregate quality score over many items. DSI produces a per-response report of surfaced vs omitted expected points, with the supporting evidence attached.

Not a guardrail

Guardrails block or rewrite unsafe outputs. DSI never writes the answer and blocks nothing; it audits a response it did not produce and reports what was left out.

Read the fuller comparison: how DSI differs from existing AI evaluation →

The boundary, stated plainly

What DSI measures — and what it does not.

It measures

  • Which configured expected paths are surfaced in a response.
  • Which configured expected paths are omitted, with the source evidence.
  • Whether a follow-up regeneration recovers missing paths.
  • A reproducible fingerprint so a result can be re-derived.

It does not measure

  • Advice quality, or whether the advice is correct or optimal.
  • Factual correctness, safety, or user outcomes.
  • Regulatory compliance — DSI does not certify compliance or governance.
  • Any property it was not configured, by its expected set, to look for.

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.

How it is used

Audit → recommend → re-audit → evidence.

01 · Audit

Visibility

Audits a supplied response against a configured map of the reasonable option paths, reporting expected-path visibility and which expected paths are missing.

02 · Recommend

An instruction

Recommends which missing paths to surface, as an instruction your system uses to regenerate. DSI never writes the answer.

03 · Re-audit

The change

Re-audits your regenerated response and reports the visibility change.

04 · Evidence

A bundle

Packages everything into a portable, reproducible evidence bundle with a trust chain.

DSI runs entirely on your own infrastructure. It does not call or replace your model, it makes no hosted calls, and it stores nothing. See getting started →

Official research record

Grounded in original research.

DSI originated from a preprint introducing decision-space collapse and the DSI measurement framework, with public replication materials.

Decision-Space Collapse in Advisory Language Models
Measuring Trajectory Omission, Framing Sensitivity, and Recovery through Decision-Space Integrity.

Cite as

Cousins, A. J. (2026). Decision-Space Collapse in Advisory Language Models: Measuring Trajectory Omission, Framing Sensitivity, and Recovery through Decision-Space Integrity. Preprint. https://decisionspaceintegrity.com/paper.html  doi:10.17605/OSF.IO/KW25A

See the full research & evidence status →

Common questions

Frequently asked.

Is decision-space collapse the same as hallucination?

No. Hallucination is about saying something untrue. Decision-space collapse is about leaving reasonable options unstated. A response can be fully accurate and still collapse the decision space.

Does DSI decide what the right answer is?

No. DSI never generates the advice and does not judge whether advice is correct or optimal. It reports which configured expected paths were surfaced or omitted, with evidence.

Where does the "expected set" come from?

It is configured for a domain — a defined reference of the reasonable option paths for that kind of decision. Coverage is measured against that configured set, not against every conceivable point.

Does DSI send my data anywhere?

No. DSI is self-hosted and stateless. It runs on your own infrastructure, makes no hosted calls, and stores nothing.

Can DSI be used for AI-governance documentation?

It can support AI-governance documentation by producing reproducible omission evidence. It does not certify governance, safety, or compliance.

Next

See a worked example, or read the paper.