Prosecution Workbench Request access →

It accelerates the analysis. It never makes the judgment.

A tool that drafts patent responses earns trust by what it refuses to do. These limits aren't settings — they're the design. The strategic call is always the attorney's.

The hard guardrails — never violated

Four things the workbench will not do.

— does not file

It never files with the USPTO. The deliverable is a reviewable skeleton. Submission is a deliberate act you take, outside the tool.

— does not decide

No final patentability, validity, infringement, freedom-to-operate, or inventorship calls. It will not present a generated argument as a substitute for your judgment.

— does not promise

No guarantee of allowance, validity, or non-infringement. It frames options and their costs; outcomes belong to the record and the examiner.

— does not hide

Low-confidence and unsupported analysis is marked, not quietly folded into a clean-looking conclusion.

Verified provenance

Nothing is asserted that isn't in the record

Every generated mapping quotes a real passage, and the quote is checked against the source text before you ever see it. The status of that check is a field on the analysis, not a vibe:

Verified — found verbatim and anchored to its exact location (paragraph, figure, claim, or column and lines).
Partial — the substance is present but the words aren't exact; shown, and flagged as such.
Unverified — not found in the source. It is surfaced as caution and held out of the coverage analysis — so fabricated evidence can't silently close a gap or invent one.

The same discipline runs through the draft: each section carries its sources, and the export lists them.

provenance element 2 verified

Supporting quote — checked against Smith

“…a temperature sensor 102 configured to generate temperature readings…”

Smith [0012] offset 1184–1247 found verbatim

If it weren't there, you'd see that too

Uncertainty & confidentiality

It tells you what it doesn't know.

uncertainty

Assumptions stay visible

Low-confidence output is marked. Missing documents are named — a drawing that would supply §112 support but was never provided is flagged, not assumed. Unverified mappings are held out, not blended in.

your record

You bring the documents

Today the workbench works from what you upload — the action, the claims, the specification, the cited art. It doesn't reach into your matters or act on your behalf. The record it analyzes is the one you give it.

confidentiality

Unpublished stays yours

Any future public-record retrieval will be limited to published applications — which by definition carry nothing confidential. Unpublished, pre-publication matters are never auto-fetched; they're handled only from what you choose to provide.

Auditability

A record you could defend

Generated text logs its provenance. Your accepted and rejected suggestions are recorded. Every parsed field is overridable, and a correction re-runs what depends on it. The matter accrues an append-only history — what was read, what was changed, what was generated, and when.

The point isn't automation you have to trust. It's analysis you can check, on a record you can stand behind.

✓ mapping accepted — Smith [0012] → el.2

✎ reference label corrected — “Lee”

↻ claim 1 re-segmented — mappings rebuilt

◆ strategy selected — traverse, no-motivation

… draft generated — 3 sources · 1 flag

append-only matter history

Workbench, not autopilot

Leverage without losing control.

The analysis moves fast. The judgment stays exactly where it belongs.

Request access → See how it works