How it works

How a question becomes an answer — or doesn’t

The decision path behind a grounded retrieval system. Three outcomes, and the reason two of them look identical from the outside.

The claim

Most assistants have one outcome: an answer. This one has three, and the order of operations is what makes the difference trustworthy. Permissions are applied to the corpus before anything is searched or ranked — so a document the person asking may not see is never scored, never summarised, and never reaches the model.

That produces two different refusals with deliberately identical wording. If they read differently, the phrasing itself would disclose that a restricted document exists.

CURATED CORPUSMODEL CALLED PAST HEREQuestionFilter by who’s askingbefore anything is scoredRetrieve & rankwhat’s left onlyEnough evidence?measured thresholdCited answera source per claimsurvivorswithheld by permissionbelow the thresholdInsufficient evidencein the knowledge baseOne box, two causes — the asker cannot tell which applied.
Both refusal branches terminate at the same box. That convergence is the security property: a permission-driven refusal is worded identically to an ordinary one, so the refusal itself discloses nothing.

See it happen

Fifty-three seconds of the live system

The diagram shows the mechanism; the footage shows the consequence — a cited answer, a refusal, and the same question answered differently under two roles.

Demonstration system. Synthetic data for a fictional manufacturer.

The three outcomes

The diagram carries the idea. These prove it exists.

Cited answer

Evidence cleared the threshold, so every factual sentence carries the document and section it came from — a reviewer checks the answer against the source instead of trusting it. Here an engineer asks what acrylic coating thickness a humid-service part requires, and the sources panel scores and cites each passage behind the answer.

The Decision Accelerator answering an engineer's question about acrylic coating thickness. Every sentence carries a bracketed citation, and a 'Sources behind this answer' panel lists passages with relevance scores of 0.86, 0.71 and 0.76, each marked cited.
Engineer role · every sentence cited · each source scored and shown.

Insufficient evidence — nothing relevant enough

The corpus doesn’t support an answer, so the system says so instead of assembling a plausible one from adjacent material. The best passage scored 0.45, below the 0.50 relevance floor — the model was never called, and all 110 chunks were visible to this asker. Nothing was hidden; it simply isn’t there.

The same system declining to answer a torque-spec question. A refusal box reads 'Insufficient evidence in the knowledge base — best passage scored 0.45, below the 0.50 relevance floor,' with status badges: model not called, 0.0s, 110 of 110 chunks visible.
model not called · 110/110 chunks visible · best passage 0.45.

Insufficient evidence — not this asker’s to see

Now the same question, asked by two different people. The engineer — cleared for customer-linked failure analysis — gets the full root-cause answer, drawn partly from restricted sources. The external contractor asks the identical question and gets a refusal worded word-for-word like the one above, because 23 restricted chunks were removed before the search ran. A distinct message would itself confirm that something was being withheld.

The engineer role asking 'What was the root cause of the Halverson field returns?' and receiving a detailed, fully cited root-cause answer. The sources panel shows passages marked both restricted and cited.
Engineer Full cited answer, restricted sources included.
The contractor role asking the identical Halverson question and receiving the same insufficient-evidence refusal. Badges read model not called, 87 of 110 chunks visible, 23 withheld by ACL; an operator panel notes 23 chunks were removed before scoring.
Contractor Same question, refused — 23 withheld by ACL.

Same question, same system, different person asking. The contractor’s view here shows what access control withheld — that panel is instrumentation for the demo; a real contractor sees only the refusal.

Deterministic ranking

In failure triage, the model doesn’t decide what’s likely.

When an engineer describes a failure, the system first turns the description into structured fields — product, environment, time in service, onset, symptom pattern — and shows them back for correction before anything is ranked. Candidate causes are then ranked by a scored match against a failure-mode library your senior engineer has reviewed and signed off, mode by mode. That scoring is plain and repeatable: once the fields are confirmed, the same fields produce the same ranked list every time, with no model involved. A mode nobody has approved is never matched.

Only then is the model called — to explain why each candidate matches, cite the precedent and design rules behind it, and name the check that confirms or rules it out. It can’t add a cause, drop one, or reorder the list. If nothing clears the match threshold, or the only matches rest on documents the asker isn’t cleared to see, the answer is the same sentence as every other refusal: Insufficient evidence in the knowledge base. It refuses rather than guesses.

  1. 1

    Describe, then confirm

    The model turns free text into structured fields. The engineer checks them before anything is ranked.

  2. 2

    Filter by who’s asking

    Same rule as above: modes whose evidence the asker can’t see are removed before scoring.

  3. 3

    Score against the signed-off library

    Plain, repeatable scoring against approved modes only. No model call.

  4. 4

    Explain and cite — or refuse

    The model explains each candidate from the evidence it’s handed. It can’t add, drop or reorder one.

The recording above shows the question-answering path. Triage runs on the same permission filter and the same refusal, with the ranking step added.

Practice Mode · add-on

Your golden set, turned into practice for junior engineers.

The 100+ golden Q&A pairs your experts verify during the pilot double as training material. Practice Mode presents them as realistic scenarios — “A customer reports intermittent resets after 14 months on a marine platform. What do you suspect, and what do you check first?” — and the engineer answers in their own words.

Answers are graded only against the points your experts verified for that scenario. The feedback shows what they covered, what they missed, and the source passages behind each point, quoted directly rather than written by the model. Scenarios built on restricted documents never appear for people who can’t see those documents. Juniors learn by doing, instead of by interrupting your seniors.

$12,000 one-time, added after any pilot and built from that pilot’s golden set.

See it on your own documents

Want to watch this run against your corpus?

Bring the decision your team waits on most. I’ll show you the same three outcomes — cited, refused, and refused-by-permission — on a slice of your own material.

Prefer email? Reach me directly at