The workflow

1. Upload — pick the request folder (or a zip). Every file is hashed on arrival for chain of custody.
2. Configure — tell the tool what the request asked for (Config button). This is what turns structural checks into responsiveness checks.
3. Rescan — runs all checks and attaches AI-suggested verdicts to what it finds.
4. Review — confirm or dismiss findings. Confirmed findings are your deficiency evidence; dismissals never resurface.
5. Repeat per production — when an interim arrives, upload it into the same request and rescan: the report opens with exactly what's new.

Preparing a production (input quality = finding quality)

• Folder layout: Correspondence/ (request letter, acknowledgments, response letters) + Records/<YYYY-MM-DD - label>/, one folder per release. The date prefix feeds the timeline check.
• Capture deliveries immediately. Agency download links expire (often 14 days). Download, then upload here the same day — a "record-less production" finding usually means a link lapsed.
• Prefer native email files (.eml/.msg) over printed-to-PDF emails: printouts lose addresses and attachments, which weakens the party-scope and attachment checks.
• Keep response letters in the production — the tool reads their claims ("nothing redacted", "N items withheld") and checks them against reality.
• Any file type is accepted; known formats are parsed, everything else is catalogued and hashed.

Configuration is the request letter, restated

• Counterparty is the strongest check — who the communications must be with. Plain domains (nea.org) plus names and known affiliates. If a review shows an in-scope party you didn't list, add it and rescan: the config is meant to improve iteratively.
• Keywords: ALL-CAPS terms match as exact words (AFT will not match "after"); everything else matches inside words (vaccin covers vaccination). Include bill numbers and synonyms.
• Elements: one per bullet of the request letter — this catches an agency answering one bullet and ignoring the rest.
• Record forms: tick what the letter enumerates; the tool flags forms with zero recognizable items so you can ask whether phones/chat/calendars were searched.

Reading the results

• conflict = the agency's letter contradicts its own production; flag = likely deficiency; note = informational / review candidate.
• The AI suggestion beside an open finding is a recommendation with its reasoning — you decide. "review" means it genuinely needs your eyes.
• Confirm = cite it to the agency. Dismiss = not a problem (say why — reasons are your team's record). Verdicts survive every rescan.
• A missing thread root, a named-but-unproduced attachment, or a record produced by a counterpart agency but not this one are your strongest items — the evidence is a specific file you already hold.

What the tool cannot verify

• Whether the agency's search itself was adequate — the findings generate the questions (custodians? systems? terms? date bounds?); only the agency can answer them.
• Whether withheld material was properly exempt — legal judgment.
• Records that were never produced and never referenced — pair requests to counterpart agencies and let the cross-request check compare both sides.