Price list lookup by email: ask, get the row and its date

A lookup answer is a value, its source and a date, with no arithmetic

A quote needs arithmetic: quantity times price, a discount band, a setup fee. A lookup needs none. Someone asks what the after-hours rate is, how fast you respond to an outage, or how long backups are kept. The answer is a value that already sits in a note. The work is finding the right note and not inventing anything when there is no note.

That makes lookup a useful first workflow to test. It is the easiest to grade (the answer either contains the value or it does not) and it shows quickly whether your documents are written clearly enough for anyone, human or machine, to answer from them.

What a good lookup answer contains

• The value, stated plainly: 15 minutes, 24/7, for a P1 (Fieldstone IT's policy; Fieldstone is fictional).
• The source: the note the answer was read from.
• The date of that note, taken from the updated field in its header, so the person reading knows how old the claim is.
• Nothing else. A lookup answer that adds advice or rounds a number is a defect, because the reader will trust it as if it were the note.

When Staffbox drafts an email reply to a question rather than a quote, the footer lists the notes the answer read and their updated dates. The draft replies page shows the exact footer.

A lookup has to find the right row when the email carries other numbers. In this test request the signature has a suite number and a phone number; the draft used neither.

Your notes, not the model, decide whether the answer is found

What the lookup tests look like

The Fieldstone IT demo brain (a fictional IT provider) has 40 tests: 16 quote, 10 lookup, 8 triage and 6 unknown. These are four of the ten lookup tests, taken from the test file in the open repository. A lookup test passes when the answer contains every expected phrase.

The wording of a real question will not match the note's heading. Nobody writes "P1 first response". The note carries aliases (the Fieldstone response-times policy lists SLA, response time and priority) so that the everyday phrase finds it. Writing aliases is the main skill in preparing notes for lookup.

Check what the worker will read before it answers

The command line has a command for exactly this check. It prints the notes the worker would read for a question, without asking a model anything:

staffbox context VAULT "A whole office is down. How fast do we have to respond?"

If the right note is not in that list, the answer will be wrong however good the model is. The fix is in the note (add an alias, split a long note), not in the model. In our own scoring we fix in that order: the brain first, then a tool, then a bigger model. On 30 September 2026 the same 40 Fieldstone tests scored 1 of 40 for Qwen3 8B with no notes and 2 of 40 for Qwen3 27B. With the notes alone they scored 33 and 39; with notes plus the quote action, 40 of 40 each.

Run staffbox check VAULT after editing notes. It lints headers, links and dates and, when the vault is in git, checks that the activity log has only been appended to.

What has been measured

How we measured: Fieldstone IT scored 40 of 40 on 2 October and again on 6 October 2026 (16 GB Mac mini M4, Qwen3 8B, notes plus the quote action), which includes its ten lookup and six unknown tests. The score is not broken out per task type on the public pages yet, so we do not claim a separate lookup figure. All of this is fictional-company data; no customer has run it. The full method is on the scorecard page.

When no note covers the question, the right answer is NOT IN VAULT

A lookup tool is only trustworthy if it says so when it does not know. Six of the 40 Fieldstone tests are unknown-answer tests, and the expected answer is a refusal. Examples from the file:

• What is the price of a 34in ultrawide monitor? (not stocked)
• What is the warranty on the LT-14 laptop? (the notes never say)
• Do we support Linux desktops for clients? (no note)
• Price 20 iPads with setup. (not stocked)

The correct output is the words NOT IN VAULT plus a reason, and in a draft reply a note to the reviewer. A refusal on a question that does have an answer counts as a miss in our scoring, so a worker cannot score well by refusing everything. See misses and refusals for the counts.

A dated note keeps the answer honest after a price changes

Every Fieldstone note starts with a header: type, owner, updated date, source and tags. The owner field names a role, such as the service desk lead, not a person. That is deliberate: when the answer is wrong, you want to know whose job it is to fix the note.

A lookup tool cannot know that you raised prices on Monday. Update the note, change the updated date, and the next answer cites the new date. If the date on a citation looks old, that is the warning. We recommend a standing monthly review of the notes that get asked about most (our view).

Sources

• Fieldstone IT demo brain and test file (tests.jsonl): github.com/Staffbox-ai/staffbox, read 13 Oct 2026.
• Scores with and without the notes, 30 Sep and 6 Oct 2026: github.com/Staffbox-ai/staffbox/tree/main/evals/results.