Skip to content

Replay of a real run — captured 2 Jul 2026

Calibrate

Scenario-Tuning Agent

Every step below is a real trace event from the signed run ledger.

sc-calibrate

Transcript

Calibratesonnet
Steps14
Tools5
Tokens66k
Elapsed0:53
Briefed on the case

I'll work through this methodically — history first, then hypothesis, backtest, gate, and finally commit. Let me start by pulling the labeled alert history.

Excellent — the history tells a clear story. Let me break it down before forming a hypothesis:


Current cashDepositThreshold: 8,000 | Total alerts: 8

AlertfiredValueLabel
H-STR-048,200fail FP
H-STR-068,400fail FP
H-STR-058,600fail FP
H-STR-078,800fail FP
H-STR-039,100pass TP
H-STR-029,400pass TP
H-STR-019,500pass TP
A19,700pass TP

Current precision: 4/8 = 50% | Current recall: 4/4 = 100%

Key observation: There is a clean gap between the highest false positive (8,800) and the lowest true positive (9,100). All 4 false positives cluster between 8,200–8,800; all 4 true positives sit at 9,100 and above. Raising the threshold to 9,000 should eliminate all FPs while preserving every TP — a precision improvement with zero recall cost.

Let me record this hypothesis and run the backtest simultaneously:

Backtest confirms the hypothesis exactly. Now I'll run the binding eval gate:

Backtest results (proposed = 9,000):

  • Precision: 1.00 (+0.50 vs current 0.50) Major improvement
  • Recall: 1.00 (±0.00 vs current 1.00) No coverage loss
  • Alert volume: 4 (−4 vs current 8) — volume halved, driven entirely by FP removal

The gate passed cleanly — not marginal. All three conditions satisfied:

Gate conditionResult
Precision strictly improvespass 0.50 → 1.00 (+0.50)
Recall ≥ floor (0.90)pass 1.00 ≥ 0.90 (held perfectly)
Volume not catastrophically collapsedpass 4 alerts (halved, not zeroed)

gatePassed: true, marginal: false, impliedRecommendation: adopt — the gate certifies this change as safe. Committing now:

The decision brief

Recommendation
Adopt

Tuning Decision — structuring scenario

Recommendation: ADOPT

Current (8,000)Proposed (9,000)Δ
Precision0.501.00+0.50
Recall1.001.000.00
Alert volume84−4

Reasoning

The alert history showed a structurally clean false-positive band: every FP fired between 8,200 and 8,800, and every TP fired between 9,100 and 9,700 — no overlap, no ambiguity. The 300-unit gap between the highest FP (8,800) and the lowest TP (9,100) allowed a threshold of 9,000 to sit squarely between them.

  • Volume reduction (−4 alerts, −50%) is entirely attributable to FP suppression — no true positives were lost at this threshold.
  • Recall held at 1.00, comfortably above the 0.90 recall floor. No financial crime coverage was sacrificed.
  • Precision doubled (0.50 → 1.00), cutting analyst workload in half.
  • The eval gate returned gatePassed: true, marginal: false, and impliedRecommendation: adopt — a clean, non-borderline certification. Change is safe to promote to production.
Cost$0.0702
Turns6
Time53.2s
Recorded to the audit ledger · tamper-evident