A bounded demonstration
Built by Dreamers for FLRA RFQ1827298.
We feel so strongly aligned with your modernization objectives, that we have exposed a fully functional demonstration around the solicitation’s test files.
This is not a mock data placeholder, and reviewers are welcome to try it with real documents.
Why? Because OCR and IDP are solved problems, and we don’t believe in wasting Government time or resources on solved problems.
House rules
Every value carries a receipt
A review state, and a link to the page of record that supports it.
Quotes are verified, not trusted
A model-cited quote counts only after it is found in the page text itself.
Blanks stay blank
A field the source does not state stays empty. Hallucination is guardrailed out at the schema: the model is never permitted to guess.
Approval is the gate
Nothing publishes to the searchable record until a reviewer accepts it.
Data handling
- Uploads
- Processed transiently, then deleted. No CUI or sensitive records.
- Limits
- Live calls are rate- and budget-capped per visitor and globally.
- Training
- No FLRA data trains or improves any model.
- Indexing
- noindex, nofollow throughout. Service pages are never indexed.
- Analytics
- Basic, IP-anonymized page counts on the hosted demo only — never locally, never over document content.
The boundary
The proposed solution runs inside FLRA's Microsoft 365 GCC; this enclave exists so evaluators can touch the workflow with no connection to FLRA systems or data. Not an integration, an authorization, or a corpus-scale benchmark.
The corpus
The two certification PDFs posted with the solicitation (DE-RP-25-0028, WA-RP-25-0075), unmodified, beside fifteen real FLRA certifications recovered from public OPM and union CBA appendices — always labeled public analogues. Certificate-of-service pages list individual contacts, so they are unindexed and hidden by default.