RFP Compliance Matrix & Disqualifier Extractor

What it does (for agents & pipelines): give it a US federal solicitation (RFP/RFQ/IFB) as text or a public URL/PDF, and it returns a structured compliance matrix — every shall/will/must requirement with a verbatim quote, each auto-classified against a disqualifier table (page limit, font, deadline, submission method, SAM registration, reps & certs), plus a deterministic Section L vs Section M crosswalk that surfaces hidden requirements. Machine-readable, pay-per-document, no seat license.

Returns, as structured JSON:

• ✅ Every binding requirement (shall/will/must/required) with its requirement ID, section, and a verbatim source_quote — nothing invented
• ✅ A ready disqualifier list — the procedural/administrative requirements whose violation can eliminate a proposal (page limits, fonts, margins, submission deadline, submission method/# of copies, file format, volume structure, SAM.gov registration, reps & certs), each with a plain-English why and a critical/high/medium severity
• ✅ Section L ↔ Section M crosswalk — the differentiator: it flags evaluation factors scored in Section M that no Section L instruction tells you to submit (a hidden requirement, high-severity) and instructions not mirrored by an evaluation factor (possible wasted effort)
• ✅ Optional proposal coverage — pass your draft and get a per-requirement heuristic of what appears addressed vs. missing

Built for proposal managers, capture teams, GovCon consultants, and RFP-response agents who need the compliance matrix a competitor charges $4,000–$5,600 of labor (10–18 hours) to build by hand — in seconds, per document.

Why the disqualifier flags and crosswalk are trustworthy

The model only extracts requirement text and verbatim quotes from your solicitation. It never decides what is a disqualifier and never compares L to M. Deterministic code does both, from a curated rule matrix (rules/compliance-rules.json): a disqualifier pattern table grounded in the uniform contract format and the late-submission rule (FAR 15.204-1, 15.204-5, 52.212-1(f)), and a standard set of evaluation subjects used to run the L/M crosswalk as pure set logic. Every flag traces back to a verbatim quote you can click to the source.

Input

{
  "documents": [
    { "id": "rfp-1", "text": "SECTION L - INSTRUCTIONS ... The Technical Volume shall not exceed 30 pages ... SECTION M - EVALUATION FACTORS ... Past performance ..." },
    { "id": "rfp-2", "url": "https://example.com/public-solicitation.pdf" },
    { "id": "rfp-3", "text": "SECTION L ... SECTION M ...", "proposal_text": "Our technical approach ..." }
  ]
}

1–20 documents per run. Provide text (paste Sections L and M — fastest and most reliable) or a url to a public solicitation page/PDF (fetched best-effort). Optional proposal_text/proposal_url adds the coverage check. Do not point url at sam.gov — its Terms of Use prohibit automated harvesting; paste the text instead.

Output (per document)

{
  "meta": { "solicitation_no": "47QTCA-26-R-0007", "agency": "GSA", "due_date": "15 September 2026", "vehicle": "RFP" },
  "requirements": [
    { "requirement_id": "L.1", "section": "L", "modal": "shall", "category": "disqualifier",
      "disqualifier_id": "page-limit", "severity": "critical",
      "why": "A stated page limit. Evaluators commonly do not read — and may remove — pages beyond the limit ...",
      "source_quote": "shall not exceed 30 pages" },
    { "requirement_id": "L.4", "section": "L", "modal": "shall", "category": "substantive",
      "subjects": ["technical-approach"], "source_quote": "describe its technical approach and management plan" }
  ],
  "disqualifiers": [ /* the critical/high items, sorted */ ],
  "crosswalk_gaps": [
    { "kind": "M_not_in_L", "subject": "small-business", "severity": "high",
      "detail": "Evaluated under Section M but no Section L instruction ... asks you to submit Small business participation content — likely a hidden requirement." }
  ],
  "counts": { "requirements": 7, "disqualifiers": 5, "crosswalk_gaps": 1 },
  "disclaimer": "..."
}

vs. VisibleThread / GovDash / Loopio / Responsive

These are strong platforms. VisibleThread does deterministic shall extraction; GovDash, Loopio, and Responsive build compliance matrices inside a full proposal-collaboration suite (typically $7k–$28k+/year, sales-led). This Actor is the pay-per-document, API-first, agent-friendly version of just the matrix job: no seat subscription, no sales cycle, structured output you can drop into a pipeline. What it adds beyond raw shall extraction is the deterministic disqualifier taxonomy (with severity + why) and the L/M crosswalk — the two things that actually keep a proposal from being thrown out. If you want an end-to-end content library and team workflow, use Loopio/Responsive. If you want automatable compliance-matrix extraction you pay for by the document, use this.

Limits (read this)

• NOT legal advice and not a guarantee of compliance or award. It is a drafting aid. Always verify every flag against the exact solicitation text, amendments, and Q&A, and confirm the deadline and submission method directly with the contracting officer.
• The disqualifier table and the L/M crosswalk are heuristic pattern/subject matching (v0.1). They catch the common killers and hidden-requirement patterns; they are not a substitute for a human compliance read. The crosswalk needs both requirements and Section M factors present in the document.
• Extraction is text-only and depends on document quality. Garbage in → low quote_hit_ratio → not billed.

Fair charging

Pay per successfully processed document (matrix-built). Failed fetches, non-solicitation text, low-verification extractions, and documents skipped after your run's charge limit is reached are not charged. Every source_quote is re-verified verbatim against the document before the run is billable.

Development

pnpm install
pnpm typecheck   # tsc --noEmit
pnpm test        # vitest: rules table, validate layer, classify + crosswalk
pnpm smoke       # offline end-to-end (no network, no LLM)
pnpm golden      # real-LLM acceptance vs hand-labeled solicitation samples (needs DEEPSEEK_API_KEY)

Architecture (single-direction deps): main → fetch/extract/validate/match → rules → domain. Same three-layer engine as the tender-deep-extractor and the GovCon clause extractor; the disqualifier + evaluation-subject rule matrix is the moat.

Current-source verification

Rule matrix grounded in primary FAR sources retrieved 2026-07-18: uniform contract format Section L/M identity — FAR 15.204-1, FAR 15.204-5; late-submission rule — FAR 52.212-1(f), FAR 15.208.