How it works
SR&ED's real problem is not the writing — it is that the claim must survive an auditor. A write-up that cites a PR that does not exist, inflates routine work into "uncertainty", or silently drops half the year's evidence fails audits. This app parses what your paste actually contains, has the model group and judge only that, then re-checks the output in public: links, numbers, item coverage, incident exclusion.
The scan parses, free
PR lists, tickets, doc titles and notes become work items — kind, URL, date, incident markers — with counts, a claim-window check (Feb 1 to Jan 31) and a per-item table you can export. No sign-in, no model call, no charge.
The model groups, judges, writes
Items are grouped into projects; each project faces the CRA two-part test — was there a genuine technological unknown, and was it attacked by systematic experiment? Eligible projects get the submission write-up: description and goals at 100 words max, uncertainties each with experiments, results and evidence links. Routine work is left out honestly, with reasons.
The checks keep it audit-safe
Every URL and number in the write-up is searched for in your paste. Every parsed item must land in exactly one project or be explicitly unassigned. Excluded incident work must stay out of the claim. What the evidence never establishes is listed in a gaps section, not papered over.
A derived work of the @getsentry/sred-work-summary and @getsentry/sred-project-organizer skills (getsentry/skills, Apache-2.0). The project grouping, the SREDable test and the submission template are theirs; the paste lane, the work scan, the grounding checks and the account history are this app's. This tool drafts documentation - it is not tax advice; review everything with your SR&ED advisor.