Trust From Regret

We frame “trust” as calibrated expectation of low future regret enabled by near‑zero‑latency internal feedback and bounded updates.

Definition (Operational)

  • Trust rises when recent decisions show: (i) low regret at fixed λ, (ii) good deferral quality on ambiguous cases, (iii) calibrated uncertainty, (iv) stable corrections after contradiction signals, and (v) constraint compliance.

  • We measure these via the Control‑of‑Error Index (CEI) and Control KPIs.

Instantaneous Feedback → Faster Correction

  • Compitum emits endogenous, judge‑free signals at decision time: feasibility, boundary ambiguity (gap/entropy/uncertainty), and drift/trust‑radius state.

  • A Lyapunov‑inspired trust‑region caps update step sizes, turning contradiction signals into gentle, stable adjustments of the learned SPD metric.

  • This reduces correction latency and noise vs. delayed judge schemes, supporting lower future regret.

Evidence (What to Report)

  • Fixed‑WTP performance: regret and win‑rate vs. best baseline at λ ∈ {0.1, 1.0}; bootstrap CIs.

  • CEI components (from existing CSVs/certificates):

    • Deferral quality (boundary vs. high‑regret): AP, AUROC.

    • Calibration: Spearman ρ(uncertainty, |regret|) and reliability curve.

    • Stability: Spearman ρ(shrink in trust‑radius vs. future regret decrease).

    • Compliance: feasible rate ≈ 1.

  • Control KPIs: trust‑radius event counts (shrink/expand/steady), r summary stats, shrink→improve correlation.

Helper Commands

  • CEI report

python tools\analysis\cei_report.py ^
  --input data\rb_clean\eval_results\<latest-compitum-csv>.csv ^
  --out-json reports\cei_report.json ^
  --out-md reports\cei_report.md

  • Reliability curve

python tools\analysis\reliability_curve.py ^
  --input data\rb_clean\eval_results\<latest-compitum-csv>.csv ^
  --bins 10 ^
  --out-csv reports\reliability_curve.csv ^
  --out-md reports\reliability_curve.md ^
  --out-png reports\reliability_curve.png

  • Control KPIs

python tools\analysis\control_kpis.py ^
  --certs reports\certificates.jsonl ^
  --eval data\rb_clean\eval_results\<latest-compitum-csv>.csv ^
  --out-json reports\control_kpis.json ^
  --out-md reports\control_kpis.md

Notes

  • “Lyapunov‑inspired” emphasizes bounded, stabilizing updates without claiming a formal proof.

  • Approximate shadow prices are report‑only diagnostics; selection is feasibility‑first argmax U.

  • Coherence prior is bounded (clipping, small β_s); conclusions robust to modest β_s changes.