Patent Audit

Pre-filing patent application audit

Audit a draft utility patent application against 316 checklist items and return a ranked defect list with coverage accounting.

Not legal advice. The checklist is grounded in 35 U.S.C., 37 CFR, the MPEP, and current USPTO materials. Fee amounts, form names and filing mechanics must be re-verified against current PTO practice, and §101 case law moves. Say so in the report; the template already does.

What ships with this skill

Routing is precomputed. Every item carries:

• inputs — what evidence it consumes: draft_text (210 items), intake_facts (99), forms_packet (68), drawings (47), search_report (40), prior_art_refs (20), ppa_text (20), parent_app (13).
• automation — mechanical (28), assisted (169), judgment (119).
• owner — which agent runs it.
• blocking — 105 items halt the audit on failure.

26 items are answerable only from intake facts. No document review reaches them: A19 A20 A22–A25 A36 C05 C08–C12 C21 D19–D26 D28 D29 I38 I42. If intake is skipped, those are cannot_assess — never pass.

Core principles

The draft fits in one context. Parallelize for attention, not capacity. Every agent gets the whole application. Never chunk the document across agents — enablement, antecedent basis and numeral consistency are whole-document properties.

Never let a model do arithmetic or exhaustive cross-referencing. Dates, claim counts, fee tiers, dependency graphs and numeral reconciliation run in code. Models find 19 of 20 numerals and report "all consistent."

Silence is not a pass. Every item ends as pass, fail, cannot_assess, not_applicable, or not_reached. An agent that omits an item has not passed it.

cannot_assess is a correct answer. Some items — the duty of candour above all — cannot be established from any document. Guessing is worse than admitting.

Phase 0 — Intake (blocking)

Run prepass.py with no --intake; it writes the questionnaire into the working directory. Have the user fill it. Parts A, C and D are unanswerable without it.

Do not proceed past the gates on assumed facts. If the user wants to start anyway, run Phases 1 and 3 only, and mark every intake-dependent item cannot_assess.

The most consequential fields are the disclosure and bar dates, and disclosure_made_by — post-AIA, only an inventor-derived disclosure gets the one-year grace period. A third party's independent disclosure before filing means the U.S. right is already gone, and no amount of drafting fixes it.

Phase 1 — Deterministic pre-pass

python scripts/prepass.py --draft DRAFT --drawings DRAWINGS \
    --intake intake.yaml --workdir audit_run

Exit 1 means intake is missing; exit 2 means the draft did not parse and the run must stop. Individual steps if you need them:

python scripts/parse_application.py DRAFT --drawings D -o parsed.json
python scripts/mechanical_checks.py parsed.json --intake intake.yaml -o findings/findings_mechanical.json
python scripts/deadlines.py intake.yaml -o deadlines.json [--asof YYYY-MM-DD]

parse_application.py accepts .txt, .md, .pdf, .docx. prepass.py refuses to continue when no claims parsed, fewer than four sections were recognised, or the detailed description yielded no reference numerals — all signs the draft uses headings the parser did not expect. Fix the parse before dispatching agents; otherwise they confidently audit a document that isn't there.

The mechanical checks credit the checklist items they cover, so those items are not reported as unaudited. A mechanical pass never suppresses a later agent fail on the same item — the synthesizer keeps the worst verdict.

Hand parsed.json to every agent. It carries the claim dependency graph, element decomposition, antecedent-basis candidates, reference-numeral tables and figure inventory, so no agent has to recompute them.

Phase 2 — Gates (sequential, fail fast)

Run in order. Each gets intake.yaml, parsed.json, deadlines.json, the draft, and only its own part file.

If any gate returns a fail on an item tagged blocking: true, stop. Report the blocker and do not spend tokens auditing the prose of an application that is statutorily barred, names the wrong inventors, or has already lost its priority date. Tell the user what would have run.

gate_priority must not compute dates. deadlines.json already has them, with rollover applied. Its job is to interpret them and catch the conflicts the script flags in warnings.