Every decision, replayable.
What actually happened — won, paid, late — feeds the next decision in SAP or Infor.
extraction accuracy across production backtests
Production email-AI pipeline backtests (aviation MRO deployment)Observed points are an illustrative series; the band endpoints are the measured production backtest band.
AI agent memory starts with a backtest.
Corin grew out of a pipeline that reads production inboxes today. Every model change is backtested against held-out history — a release ships only inside the measured band, and the headline is the latest first-pass result.
Forecast vs fact, measured.
Corin keeps the decision, not just the record that something happened — what it saw, what it chose, what it rejected, and what came next.
Commercial outcomes
Won, lost, discounted, disputed — the quote's real fate, joined back to the dry run that priced it.
Decision outcomes
Approved, edited, or rejected — a reviewer's correction is a signal with a name on it.
Financial outcomes
Paid on time, late, partial — terms recommendations learn from actual cash behavior.
Corrections become versioned Procedure updates.
Never invisible model changes: a fix ships as a new Procedure version — reviewed, testable, reversible in one step.
01A reviewer edits a draft
The edit is captured with the case it happened on — who changed what, and why it was wrong.
02The fix becomes a version
Corin proposes the Procedure update as a diff. The owner reviews and releases it — or doesn't.
03History re-runs
The new version replays against past cases before it ships. A fix that breaks history goes back to the owner.
Every decision can be replayed step by step.
Run it again next quarter and get the same answer — with proof it matched.
Checkpointed as it runs
Every step writes a checkpoint sealed with a fingerprint — the model's answer, the tool's response, the data it read.
Matched or mismatch
A replay compares event by event against the original and returns a verdict: matched or mismatch.
Same inputs, same decision
The same inputs produce the same decision, every time — audit season stops being archaeology.
Three kinds of memory — each with an off switch.
Memory is a control surface, not a side effect.
- Decision memory
- Every decision with its context, options, confidence, and what actually happened next. Deletable by contract, not by support ticket.
- Runtime memory
- Working state, scoped to your tenant with hard isolation — one customer's data never warms up another customer's answers.
- The work graph
- The work graph — one live view of every job — is the long-lived business memory: cases, customers, documents, built on the decision record underneath.
Scoped to your tenant. Expirable. Auditable. Off until you turn it on.
Versioned, reviewable, reversible.
Every decision records the model version that made it.
on regression → rolled back in one step