Signal Foundry — AI Product Studio Kit

An instant download of the whole kit: docs/RESEARCH.md (20 opportunities scored on a weighted rubric, a 20-source evidence register with dated links, a ten-row pre-mortem, and the reasoning for what got cut and why), docs/PRODUCTS.md (three product briefs with buyers, golden paths, pricing hypotheses, a 30-day validation plan and explicit stop criteria), docs/ARCHITECTURE.md (containers, prompt layers, the D1 schema and its multi-tenant evolution, a threat-model table and the release gates), and the full application source — Next.js 16, React 19, Cloudflare Workers, D1 with Drizzle, the Anthropic provider, and the test suite. npm run eval runs 17 tests; they were green on a clean-room install from this exact zip.

Both, deliberately, and the difference is labelled everywhere. With no API key it runs in bounded demo mode: deterministic keyword rules over three bundled sample documents. That mode is a demo — it recognises the samples and little else, and the UI says so. Set ANTHROPIC_API_KEY and it analyzes documents it has never seen. We verified that before listing: on a security questionnaire using none of the demo's hardcoded controls it answered the two supported questions with verbatim citations and blocked the third for lack of evidence; on an invoice ledger with unfamiliar headers and non-standard IDs it parsed every row, ranked by recoverable cash and routed the disputed one to review. That transcript is reproducible with npm run eval:live.

It's the one rule that makes AI output usable in a room with a lawyer in it: no source, no answer. Every finding marked ready or review carries a quote copied character-for-character from the document. When the source doesn't support a claim, the finding comes back blocked with an empty evidence array and a note saying what's missing — the model is never allowed to fill the gap from background knowledge. The prompt states the rule; the provider then re-checks it after parsing, hands a violating response back to the model once, and fails the run rather than shipping an unsupported claim. Anything contractual or irreversible stays behind a human gate even when the evidence is strong.

It defaults to Claude Opus 5 and you bring your own Anthropic key — we don't resell model access, so tokens are billed to you at Anthropic's rates. Swap the model with one env var and the reasoning depth with another; the per-product policy prompt carries a cache breakpoint, so repeated runs re-read it at cache rates. There's also a gateway provider if you'd rather route through your own endpoint and own model selection, logging and retention yourself.

Two things, both written down in docs/VALIDATION.md rather than buried. First: we hold no Anthropic key on the build machine, so while the request shape, parsing, retry and every failure branch are tested against a mocked transport, and the prompts are tested against a real model, the live HTTP call to api.anthropic.com has never been run as one piece here. Set your key and run one analysis before you trust it in production. Second: because of that, cost-per-run and P95 latency are targets in the architecture doc, not measured results. Everything else — lint, types, 17 tests, the production build — is green from the zip you download.

The research, briefs and application were produced with heavy model assistance and then verified by hand before listing — and unlike our engine-scouted kits, this one didn't come out of a scored tournament, so we won't claim it did. What we did instead is test the claim this listing rests on: that it analyzes real documents rather than only its own samples. It does, and the check is in the box.