How The Standards Desk works
The parts that carry the guarantee are deterministic Python. The agents advise. The gate decides.
A reader question runs a six-stage pipeline, live and offline: a temporal-planner decomposes it into time windows, a hybrid retriever (BM25 plus a local embedding) pulls candidate passages, a coverage floor abstains if nothing clears the score threshold, the answerer writes a strictly grounded answer with inline [S#] citations, a deterministic verbatim-quote check matches every quoted string against its cited source, a provenance auditor grades each claim, and a deterministic policy gate issues one decision. Every stage lights up in a console that streams its own latency, so nothing is a black box.
Two engines, deliberately split
Three agents that advise
The temporal-planner, answerer, and provenance auditor are three Pydantic AI agents with typed per-claim verdicts. They are provider-swappable (default claude-opus-4-8), and the same pipeline runs end to end with no key through a deterministic offline stub or a keyless local Claude bridge, so the gate stays functional either way.
Code that decides
The trust boundary sits outside the agents. The verbatim-quote rule, an exact string match against the cited source text, and the policy gate are plain Python. Publish-safety never depends on an LLM's self-report, because a buyer needs a provable, auditable decision, not a model's say-so.
One of four gate decisions per answer
Every answer resolves to exactly one decision. The two that publish do so only when the answer is fully grounded; the two that do not are honest about why.
| Gate decision | What it means | Decided by |
|---|---|---|
| AUTO-CLEARED FOR PUBLICATION | Every claim grounded, every quote verbatim, coverage above the floor | Deterministic policy gate, no human touch |
| PUBLISHED AFTER PRUNING | An unsupported sentence is dropped and the remaining grounded answer clears | Deterministic, counts as published without a human touch |
| HELD FOR STANDARDS-DESK REVIEW | A fabricated or non-verbatim quote, or an unsupported claim | Deterministic, blocked from readers and routed to the queue |
| OUTSIDE COVERAGE | Retrieval clears nothing above the score floor | Deterministic, honest abstention rather than a guess |
The gate invariant and the audit receipt
The one deterministic, unit-tested guarantee is the gate invariant: no answer with an unverifiable quote or an unsupported claim is ever auto-cleared. On top of that, one click exports a JSON audit receipt per answer, recording the query, the decomposed sub-questions, the retrieved sources with article id, date, and url, the published answer, per-claim verdicts with evidence passages, per-quote verbatim checks, the gate decision and reason, the engine and model, and a UTC timestamp. A kill switch on the dashboard pauses the reader widget while the back end stays live.
