Reliability

Green means the tools moved. Nothing else counts.

Latch is checked like a payments flow: expected outcomes, forced failures, and no success without proof IDs.

100%

Suite match rate

Golden fixtures

Proof IDs per play

Silent customer sends

Job	Fixture	Asserts	Expected	Result
Cancel email → full latch	gmail/cancel_champion_gone.eml	5/5	GREEN	match
Stripe past_due → full latch	stripe/invoice_payment_failed.json	5/5	GREEN	match
Slack token revoked → UNLATCHED	inject/slack_invalid_auth.json	2/5	RED	match
Sheets 429 → retry then pass	inject/sheets_rate_limit.json	5/5	GREEN	match
Calendar conflict → hold rescheduled	calendar/conflict_window.json	5/5	GREEN	match
Ambiguous mail → no side effects	gmail/pricing_question.eml	3/3	GREEN	match
Draft-only guard (never sends)	guard/outbound_send_attempt.json	4/4	GREEN	match
Deletion test → account purged	privacy/delete_account.json	4/4	GREEN	match

Golden jobs

Each fixture says what should happen — green or red. A run passes only when the board matches that expectation.

We break Slack (and other paths) on purpose. If a broken run still shows green, the suite fails.

Steps retry with backoff. When retries run out, Latch stops and records compensating notes instead of silent partial state.

A fixture asks to delete account data, then checks that risk rows and stored IDs are gone. We do not claim “unhackable.”