Info Desk

Evidence review · recorded case and local live desk

Incoming information.
Traceable answers.

Compare the claims in dated OFAC, White House and AP excerpts. Open the source quotations and export your review.

The public case uses stored captures. The separate live workspace collects official updates and generates cited drafts through the local Python backend.

  1. Choose the sources. Keep all three selected or choose an example below, then compare.
  2. Select a claim. Open a table row to inspect the exact quotation, source and capture date.
  3. Review and export. Read the findings, add your notes and download a cited memo. Nothing is published.

Connected sources

Energy & sanctions desk

Connecting to the backend…

Source inbox

0 documents

No documents selected: search across selected sources.

Connect to the backend to collect current source records.

“New to desk” means first seen by this installation, not newly published. Changes compare stored versions, not a guessed historical baseline.

Ask the evidence

Answers use the collected source scope: RSS summaries, notice abstracts and license titles. They are not a substitute for full articles or legal review.

Recent jobs — persisted across restarts

Read-only research. Exact citation matching checks where a quote came from; it does not prove that the model’s interpretation is correct. Source failures, stale inputs and insufficient evidence remain visible.

Under the hood

From a source update to a reviewable answer.

The live pipeline collects official records, keeps document versions and retrieves passages with SQLite FTS5 / BM25. A local model writes a short answer with exact source quotations. Failed checks stay failed; no recording is substituted for a live result.

ComponentDepends onProduces / enforces
Live connectorsEIA RSS, Federal Register API, OFAC HTMLConditional HTTP requests, provider cooldowns, response limits and visible source errors
Version storeValidated source documentsAppend-only captures, fingerprints and diffs; publication and fetch dates stay separate
RetrievalLatest stored versions + user-selected scopeBM25-ranked passages with document/version IDs; no-match abstention
AI briefQuestion + retrieved passages + structured-output modelQuote and numeric validation, visible failures, reviewable interpretation — no automatic publication
Job workerLocal API + SQLite checkpointsResponsive progress, one-job backpressure, cancellation and restart recovery to an interrupted state

The separate recorded case remains a reproducible regression lab. Live coverage is publisher summaries, abstracts and license titles, not every full document. Lexical retrieval can miss relevant passages, and citation matching cannot prove semantic correctness. Public multi-user hosting requires authentication and a deployed backend.