Governance That Compounds With Every Outcome
Confirmed fraud, a chargeback, a false positive - each past outcome adjusts future trust, weighted and decayed over time. The loop can escalate a verdict but never de-escalates, so it fails safe.
Illustrative product view
How the loop learns
score_history(history, as_of) walks an actor’s past outcomes, and each contributes weight × decay(age). Decay is exponential - 0.5^(age/90) - a 90-day half-life, so recent outcomes matter more than old ones. The result is a trust_adjustment clamped to a floor of −40 and a ceiling of +20, an autonomy_confidence from 0 to 1, cited reasons, and an explanation. It is golden-vector tested, so the math is pinned.
Outcome decay
contribution = weight × 0.5^(age_days / 90)
Each past outcome’s influence halves every 90 days, so the loop weights recent behavior more heavily while never fully forgetting older events.
Outcome weights
| Outcome | Weight | Direction |
|---|---|---|
| confirmed_fraud | −25 | Lowers trust most |
| chargeback | −18 | Lowers trust |
| disputed | −8 | Lowers trust |
| confirmed_ok | +4 | Raises trust |
| false_positive | +6 | Raises trust most |
A false positive raises trust more than a confirmed-ok because it is a stronger signal that the system was being too cautious about that actor - the loop learns to stop over-blocking, as well as to block real bad actors.
Escalates but never de-escalates
apply_learned_escalation() can walk a recommendation up the ladder - allow → challenge → require_approval → block - but it never walks it down. Learning can make governance stricter on its own; loosening always requires an explicit human decision. That asymmetry is the fail-safe: the automatic direction is always toward more caution.
How it connects
- A learned escalation is one of adaptive trust’s three triggers
- The trust adjustment feeds the enriched risk behind the Action Trust Score
- Outcomes are recorded via record_outcome and replayed by load_history
- The loop lives behind the /learn router
Frequently asked questions
What is the compounding learn loop?
It is the mechanism by which past outcomes make future decisions smarter. score_history weights each past outcome and decays it with a 90-day half-life, producing a clamped trust adjustment and an autonomy-confidence score. Confirmed fraud and chargebacks lower trust; false positives and confirmed-ok outcomes raise it.
Why does the loop only escalate, never de-escalate?
For safety. apply_learned_escalation can move a recommendation up the ladder - allow toward block - automatically, but it never moves it down. Loosening a control always requires an explicit human decision, so the only automatic direction is toward more caution.
How does time decay work?
Each past outcome contributes weight times an exponential decay of 0.5^(age in days / 90), a 90-day half-life. Recent outcomes therefore weigh more than old ones, so the loop adapts to current behavior while still remembering older events with diminishing influence.
Why does a false positive raise trust the most?
Because a false positive is strong evidence the system was over-cautious with that actor. Weighting it at +6, above confirmed_ok at +4, teaches the loop to stop over-blocking good actors, not just to block bad ones - improving precision in both directions.
Stay in the loop
One practical finance briefing a week - new guides, checklists, and benchmarks.
Make every outcome count
Let confirmed outcomes compound into smarter, fail-safe decisions.
Talk to us