Every approval, every override, every skip — in XFCD, in XFBA, in XSIA, in every HIL prompt XF surfaces — becomes training signal. XFPA aggregates decisions per rule with recency decay, computes rolling confidence, and graduates auto-apply when a rule earns it. First-time users get the full guardrails. Power users get the intelligence of having trained the platform to their judgment.
Part of Memory Layer
Connect with GitHub — it's freeXFPA tracks confidence per matched rule (e.g., ‘suppress-stamp-removal’, ‘approve-resolution-xfcd’). A rule you've approved 12 times is graduated; a brand-new rule still prompts. Trust is scoped.
An approval from a year ago counts half as much as one from last month. Stale patterns fade. Current patterns dominate. The platform adapts as your judgment evolves.
Default: 5+ signal decisions at >= 0.8 confidence graduates a rule. Contradictions and novel rules always escalate to HIL. The gate only fires when the rule has earned it.
Decisions from XFCD, XFBA, XSIA, and other HIL modules log to xfcd_decisions.jsonl.
/xfpa ingests the log into the XFMT tree as process-audit leaves.
Rolling confidence computed per rule with recency decay.
Threshold table written to ~/.claude/xfmt/xfpa_thresholds.json.
Eligible rules auto-apply on next firing; ineligible rules prompt HIL as before.
/xfpa status shows current confidence table for review.