E-104
Review Queue & Learning
The loop itself: queue, corrections, rules, and the agreement curve.
Not one decision but the habit around all of them: who reviews what, how a correction becomes a rule, and how you know when the engine can be trusted.
WHO BUYS IT · Any client of the other three patterns · Teams already using AI who want it to stop repeating mistakes · Owners who need to see trust in a number
INPUTS
- A stream of proposals from any engine
- An expert with a few minutes a day
- A definition of 'acceptable without changes'
THE PROPOSAL
- Review queue with accept and correct
- Rule ledger with history
- Agreement rate by job type over time
- A go/no-go for unattended use
WHAT THE EXPERT CORRECTS (CAPTURED AS RULES)
- Which rules conflict
- Rules that were right once and wrong now
- What counts as 'the same kind of job'
WHERE THE LANGUAGE MODEL EARNS ITS PLACE
Turns a free-text correction ('too much excess for cased hole') into a candidate rule with conditions, which the expert confirms before it is applied.