Every outcome teaches the next decision.
Most automation freezes on day one. Corin's loop closes: what actually happened — accepted quotes, late payments, disputed invoices — becomes the starting state for the next simulation.
94.5–98% extraction accuracy, backtested in production.
The email-AI pipeline Corin grew from runs in production today, extracting RFQs and POs from real inboxes. Every model change is backtested against held-out history before it ships, and the range is the honest spread across document types — not the best number we ever saw.
extraction accuracy across production backtests
Production email-AI pipeline backtests (aviation MRO deployment)What feeds back
Decision outcomes
Approved, edited, or rejected — reviewer corrections are training signal with a name on it.
Commercial outcomes
Won, lost, discounted, disputed — the quote's real fate, joined back to the simulation that produced it.
Financial outcomes
Paid on time, late, partial — terms recommendations learn from actual cash behavior.
Versioned, reviewable, reversible.
Every decision records the model version that made it. Updates roll out staged, are measurable against the previous version, and roll back in one step. The ledger never lies about which brain made which call.