Cite it, derive it, or don’t say it.

An answer about a tender is only useful if someone can check it before it goes into a bid. So a claim in a Fennec answer names the blocks of the document it rests on, and each block resolves to a document, a page, a box on that page and the quoted text. Extracted fields carry the same, with a confidence score.

The same rule holds where there are no documents. A Kosh total opens to the bills and obligations behind it; a Chainage stretch opens to the inspection that colours it.

Software that is unsure should say so, in words, where the user is looking. A progress bar that animates over a state nobody can measure is a small lie, and small lies teach people to stop reading the screen.

Fennec runs every question inside limits on steps and time, and if a limit stops the search the answer says so. Kosh keeps the contractual date and the expected date apart, and marks every expectation as an estimate. Chainage draws a progress fraction only when bytes were actually measured, and names a stalled or unreachable reader as what it is.

A status that someone types is a second copy of the truth, and second copies drift. Chainage stores inspections, never the colour of the road: each stretch is computed from the inspections covering it every time the road is drawn. The newest attempt controls; a tie resolves to the worse result, so it can never hide a failure.

The rules live in one package that both the browser and the server run, so what the screen shows is what the server will enforce, and the two cannot disagree.

A stretch approved in March was approved against March’s acceptance limits. If the catalogue changes in June, March’s approval must still read as it did. So a Chainage inspection keeps a snapshot of the rules it was raised under, and an activated configuration cannot be edited: a change makes a new version, and history is never re-judged.

Chainage keeps three append-only logs: who changed a record, who opened one, and what the document reader was asked and answered. The database refuses a rewrite; it is not left to convention or good intentions.

Fennec keeps the raw output of every document it reads and every exchange with the services that read it, so a past result can be explained, audited and reprocessed without reading the document again.

Tenders are organised by numbered clauses, schedules and sheet ids, and the questions bid teams ask often span a whole set. Fennec therefore has no embeddings and no vector store: it answers by navigating the outline, the clause numbers and the references, with full-text search alongside.

We tested the idea before we built on it. In our published benchmark, top-k retrieval answered none of the four questions that required counting across a document set: it cannot count documents it never fetched. Only the agents that navigated could say which page an answer came from.

We write down what success means before we tune for it, test on documents the system has not seen, and count invented values as a failure of their own rather than averaging them away. The published benchmark used the same model for every system and scored by deterministic comparison, with no model judging another.

And we publish where we lose. That benchmark opens with the finding that a single prompt matched months of our own engineering on short documents.

Rules that live in a document erode. So the boundaries that matter are tests that fail the build: in Chainage, which packages may depend on which, where a vendor’s name may appear, and how large a file may grow. In Fennec, document-reading vendors sit behind a neutral interface, so one can be replaced, and the raw responses replayed through the new mapping, without paying to read everything again.