Positioning
Existing AI evaluation measures a different property.
Every widely-used AI evaluation measures a real and useful property. None of them measure whether an answer quietly left reasonable options out of view. That specific property is decision-space collapse — and it is what Decision-Space Integrity was built to make measurable, alongside the tools you already run.
This is not a claim that other tools are wrong or that you should stop using them. They answer different questions. The point is simpler: a response can pass all of them and still narrow a genuinely multi-path decision down to a single path. DSI is complementary — it measures the one property the others do not.
Other tools ask is the answer good? DSI asks what did the answer leave out?
The comparison
Each approach measures a different property.
Read across, not down: each row is a distinct, useful question. Decision-Space Integrity adds one the others do not ask.
| Approach | Primarily measures |
|---|---|
| Hallucination / factuality evaluation | Factual correctness — whether statements are true and grounded. |
| Toxicity evaluation | Harmful or offensive content. |
| Guardrails | Policy compliance — blocking or rewriting outputs that break a rule. |
| RAG evaluation | Retrieval quality and factual grounding in the retrieved context. |
| Benchmark scoring | Aggregate performance across many items, as a single score. |
| Decision-Space Integrity | Visibility of configured decision paths — which are surfaced, omitted, or recovered, with source evidence and a reproducible fingerprint. |
None of this ranks the approaches. They answer different questions, and most teams will want several of them. DSI is designed to sit alongside the tools above and report the one property they do not: whether the reasonable option paths stayed in view.
Why the gap exists
Answer-level scores judge the answer, not the omission.
The common thread is that existing evaluation is answer-centric: it asks how good the produced answer is. Decision-space collapse is about what is absent — the reasonable options that were never surfaced. Absence is invisible to a score of the text that is present, which is why a fluent, accurate, safe answer can still collapse the decision. Measuring absence needs a defined reference of what a reasonable answer would keep in view — a configured expected set — and a per-response comparison against it. That is exactly what DSI does.
DSI does not certify answer quality, safety, compliance, or truth. It measures coverage against a configured expected set and records reproducible omission evidence — a different property from the tools above, meant to sit alongside them.
Next