Claims Knowledge does not support

The gate runs before drafting#

The order is the point. The check happens on the incoming message, against the scopes that suit its intent, and only a pass leads to a draft being generated at all. There is no stage at which a model writes an answer and something afterwards tries to decide whether it was justified.

1. The message is classified and the scopes for that intent are chosen.
2. Knowledge is scored, and the result is logged as covered or as a miss either way — with the score, the titles matched, which of them carried authority, any critical topics detected and the citations.
3. On a miss, guidance is requested from the workspace's owner with the matched titles attached, so the request says what was nearly relevant.
4. If that request cannot be delivered, an empty draft is held with the reason on it rather than the thread going silent.
5. The thread outcome reads *Knowledge missing — asked owner before replying*, or *reply held*.

Topics where the bar is higher#

Some questions cannot be answered from a topical match however good it looks, because getting them wrong creates an obligation. These are detected in the incoming text and require explicit owner authority:

Pricing has one deliberate exception. Where the workspace has a pricing policy on file, or an owner instruction covering the situation, Connect may acknowledge the question and collect what it needs — scope, quantities, timing — without stating a number. That is a useful reply rather than a refusal, and it is the only critical topic with a safe partial behaviour.

What the record says afterwards#

The score

How well the best match scored, so a near miss is distinguishable from nothing at all.

Matched titles

What was retrieved. Often the fastest signal: the right document, wrong authority.

Authoritative titles

Which of those could have proved something. An empty list beside a full matched list is the diagnosis.

Critical topics

Which commitment detectors fired, if any.

Authority reason

A short machine-readable verdict, such as an owner instruction being required for a critical topic, or no fact authority being present at all.

Citations

The passages that would have been used, in KB: form, for tracing.

Those six together turn 'the agent said it did not know' into a specific, fixable statement: this question needed a warranty commitment, three reference documents matched, none carried fact authority. The repair follows directly.

Why silence beats a hedge#

A hedged answer is worse than no answer in a business context, and it is worse in a specific way. A customer reading *typically around three months, though this may vary* takes three months away with them. The hedge protects the writer and not the reader, and it is indistinguishable from a commitment once it is quoted back.

So the failure mode chosen here is a delay with a named owner rather than a fast reply with an unnamed liability. The engineering note why uncertainty is a valid answer sets out the reasoning; Connect said it does not know covers what to do when it happens more often than it should.

Questions#

Related#

• Retrieving Knowledge for an answer
• Knowledge and Fact are different things
• Connect said it does not know
• Why uncertainty is a valid answer from a business agent
• What Connect may do
• Why a sales agent should refuse to answer