For Canadian startups claiming SR&ED
Connect your repos once. On the 1st of every month, T661.dev extracts your commit evidence, clusters it into falsifiable hypotheses, and drafts a Technical Activity Log that answers CRA's five eligibility questions — while the work is still fresh.
01
SR&ED refunds up to 35% of eligible R&D salaries federally, plus provincial credits. For a small dev team, that's routinely a 5–6 figure cheque.
02
CRA's first ask in a review: contemporaneous records of your systematic investigation. “We'll reconstruct it from memory at year end” is how claims get cut.
03
Your hypotheses, failed approaches, reverts, benchmarks, and breakthroughs are sitting in your commit history — the dated narrative just needs to be assembled from it, monthly.
Install the GitHub App (read-only) and pick the repos that hold R&D work. Each becomes an SR&ED project under your claiming corporation, with its own project code.
Every branch is scanned. Commits are attributed to the month they were authored, deduplicated, and classified — experimental work, benchmarks, reverts, and abandoned approaches are surfaced; routine maintenance is set aside.
Claude clusters the month's experimental commits into falsifiable hypotheses (method → result → learning, with SHAs) and drafts the monthly review. Anything it can't confirm from the evidence gets a ⚠ VERIFY marker — logs can't be finalized until you resolve every one and confirm the hours.
One zip per corporation: curated monthly logs plus the raw extraction data backing them, organized the way your SR&ED advisor wants it. Every log includes a reproduction command a reviewer can run against your repo.
## 6. Work Performed — Iterations (March 2026) | Hypothesis / Approach | Method | Results & Metrics | |---|---|---| | **H1** — Streaming TTS chunking at clause boundaries will cut perceived latency below 800ms without prosody artifacts | Implemented clause-boundary splitter (`a3f81c2e`, `9d04bb71`); benchmarked against sentence- level baseline (`c2e94d10`) | p95 first-audio 1.9s → 740ms. Prosody regression at clause joins — reverted joiner (`f81d3a55`) |
Real structure, sample content — hypothesis-driven iterations with commit SHAs, exactly what a reviewer wants to see.
Every log answers them in named sections, so a reviewer never has to hunt — and neither does your SR&ED consultant.
Q1
Named constraints, with the standard approaches that were ruled out — and why.
Q2
H1, H2, … — falsifiable statements, not a feature list.
Q3
Method → result → learning per hypothesis, with commit SHAs.
Q4
Monthly: what is now known or possible that wasn't before.
Q5
Generated monthly from dated commits, with a reproduction command.
A rounding error against the claim it protects.
$49/month
$99/month
CRA expects contemporaneous evidence that your work was a systematic investigation: what hypothesis you tested, what you tried, what happened, and what you learned — with dates. Reconstructing this at filing time, months later, is the most common reason claims get reduced under review. T661.dev generates the record monthly, while the work is fresh and the git history proves the dates.
The GitHub App requests read access to repository contents, but the pipeline only consumes commit metadata: messages, authorship, dates, file paths, and line counts. Diffs and file contents are never stored. Your evidence bundle contains commit messages and statistics — not code.
It's instructed not to, and it's structurally prevented from sneaking guesses past you: any result the model can't confirm from commit evidence is tagged with a ⚠ VERIFY marker, and a log cannot be finalized until every marker is resolved by a human. Hours are never auto-accepted either — you confirm or override the suggested figure before anything is final.
No and no. T661.dev produces technical documentation from your repository history. Eligibility determinations, expenditure calculations, and filing are between you, your SR&ED advisor, and CRA. Good contemporaneous records make that conversation dramatically easier — that's the product.
GitHub is supported today (GitHub App, read-only). A CLI for self-hosted and GitLab/Bitbucket repos — which also keeps all metadata extraction on your machine — is on the roadmap. Tell us if you need it and we'll prioritize.
Yes. Generation works for any past month your git history covers. Backfilled logs are honest about being derived from repository history; going forward, your logs are generated on the 1st of each month automatically.