#243 Friction telemetry: measure interventions per session, propose 2FA-gated tuning

[Anonymous]

Originally created by: fu351

Approval fatigue is the churn risk of any human-in-the-loop system. The decision log already records auth_required, auth_result, and the final verdict per decision, keyed by session — the data to measure fatigue exists; nothing reads it yet.

Build

• A doberman tune report: interventions per session (the headline number), top reason codes by AUTH volume, approval rate per reason-code and target class, trend over time; --json for tooling.
• Where the approval rate is ~100% over at least N occurrences with a narrow blast radius, emit a proposal: a scoped standing elevation or a policy adjustment, shown with exactly what would loosen and why.
• Proposals are surfaced only. Applying one runs through the existing possession-factor-gated policy-change path — approve the pattern once instead of the event forever. The tool never applies anything itself.

Tests must prove: aggregation correctness on fixture logs; a proposal never auto-applies; accepting one triggers the weaken gate; the report is stable under --json.

The goal state is a hands-off runtime: the guarantees stay fail-closed underneath while measured, human-ratified tuning drives interruptions per session toward zero.

[Anonymous]

Originally posted by: fu351

Council review of RAND's AI Security Guide, paired with [#242]: if the velocity/volume signal there ever gets a tunable threshold, it earns it here — 2FA-gated and driven by measured friction, never a hand-set constant. A hijacked agent can learn a fixed number and fragment its work to stay under it; a threshold that only moves through evidence and a human-approved path can't be gamed the same way. C7 is where a rate clause gets calibrated, not where it's guessed.

[Anonymous]

Ticket changed by: fu351

• status: open --> closed