Latch

The one-liner

Latch tells you whether a red test run is infra noise you can ignore or a real failure that blocks the merge.

The problem

• E2E suites go red for reasons that have nothing to do with the code: a down service, DNS, a network timeout.
• Developers learn to ignore red, or spend 20–40 minutes triaging before merging.
• Existing tooling (Datadog CI Visibility, BuildPulse, Trunk, Merge queues) is heavy, enterprise-priced, and gives dashboards, not a blunt decision.

The wedge

A read-only reporter whose last line is a verdict: Gate: PASS or Gate: BLOCK. No agent, no test rewrite, no new platform. It answers one question and stops.

Who it is for (the first ten customers)

• Teams of ~5–50 engineers.
• Running Playwright (or Jest / pytest) on CI.
• On GitHub Actions.
• Whose suite is red several times a week for infra or flake reasons.

Not enterprises (procurement kills a seed-stage sale), not solo devs (no budget).

Why us / defensibility

• Read-only builds trust: it never touches the tests.
• Code owns the decision; the model is a replaceable judge. Cheap, no model lock-in.
• The signature → history ledger (seen, flake %, suppression) compounds per repository.

Honest assessment: defensibility at the feature level is weak. The moat, if there is one, is per-repo history, the trust of a blunt verdict, and the workflow that consumes it.

Business model (hypothesis, to be tested)

• Open-source CLI + reporter: free, for adoption.
• Hosted GitHub App: history, suppression, flake budget, team dashboard, SSO — the paid tier.
• Price per repo or per seat; test willingness to pay before building any of it.

What would kill it

1. Customers do not actually care enough (they tolerate red).
2. The verdict is not trusted — a false PASS on a real bug is catastrophic.
3. One of Datadog / Playwright / GitHub ships the wedge for free.

Biggest risk today

No customers. Everything above is a hypothesis. The next dollar of effort should buy evidence, not code.