Brain — how the whole product learns
Every decision the product makes is a "brain". You correct it where you work (a card, an ad); your corrections become the frozen ground-truth eval set each brain is scored and tuned against — like teaching a person from their mistakes. This page maps every brain, how well it agrees with you, and what it got wrong.
What the brain corrected in itself
Each run mines your corrections, tightens the rule, and keeps the change only if agreement improves. This is the brain rewriting itself.
The goal: it starts recommending on its own.
Once a brain clears ~85% agreement on 50+ labels, it's trusted — the product promotes its calls to recommendations ("proven winner", "this pattern converts") instead of guesses. WinScore gets there fastest: it's self-verifying, so it can hit that bar with no labeling at all.
How updates happen (same loop for every brain)
| Verifier | your labels — frozen ground truth the brain can't edit |
| Metric | agreement % between the brain's guess and your labels |
| Update | loop tweaks a rule/weight re-scores against your labels keeps only if agreement rises, else rolls back |
| Applied now | your correction is stored as truth immediately and shown back |
| Self-verifying brains | Winner/WinScore needs no human — reality checks it (a "winner" that dies fast was wrong). It can improve fully automatically. |
| Automatic? | yes — nightly, no human in the loop, because the labels are the gate. Kept manual until you flip it on per brain. |