One §103, from the action to the draft.

The workbench detects whether the action is final, then pulls each rejection apart: its statutory basis, the claims it touches, the references cited against them, and the examiner's stated rationale.

Nothing here is guessed past what's on the page. Where the action is ambiguous — a reference cited by shorthand, a claim group left implicit — it's flagged for you to confirm, not silently resolved. Every parsed field is editable, and a correction re-runs what depends on it.

An independent claim is only arguable element by element, so the workbench splits it into limitations and preserves the original language exactly. You can re-segment — split a clause, merge two — and the mappings that hang off the old segmentation are dropped and rebuilt, not left stale.

This is the spine the rest of the analysis hangs on. Get it right and everything downstream lines up.

For every limitation, the workbench finds the passage the art relies on and quotes it — then verifies that quote against the reference text. A quote found verbatim is anchored to its exact location. A quote that's only paraphrased is marked partial; one that can't be found is marked unverified and held out of the analysis rather than counted as coverage.

That last rule matters: fabricated evidence can't quietly close a gap. When Smith and Lee both teach a fixed limit and neither teaches a learned threshold, the workbench shows the gap — the opening you'd argue.

Two honest paths usually exist: amend to add the distinguishing limitation, or traverse on the merits. The workbench builds both and attaches what each one costs — argument strength, scope surrendered, new-matter risk, and prosecution-history estoppel exposure — so the comparison is explicit.

It won't pick for you. Allowance is one goal; not giving away valuable scope is another, and the tradeoff is yours to make.

Once you've selected a path, the workbench renders a response skeleton organized by rejection: the amendment in USPTO redline format, remarks built from your selected argument, and the sources behind each section carried along. Open questions become highlighted [ATTORNEY: …] placeholders — a checklist, not filler.

It exports to Word with a page-broken “Attorney Work Product — remove before filing” appendix listing every open issue, source, and low-confidence flag. You finish it. It never files.