[shrsv]

Adds a read-only existence check that avoids copying the stored value.

Simulated for this demo: same diff and same risk score as threshold-p80-lenient-approves-borderline-score, but this org has calibrated risk_threshold_percentile: 10 (strict) against the same recent-score distribution. Side-by-side, these two PRs show the dial moving the decision, not the diff.

[livereviewbot]

LiveReview automated review

Scanned 1 changed file (+8/-0). New read-locked accessor, mirrors the existing Get pattern. No findings; moderate risk score (~0.55-0.6), identical to the companion PR for this same diff under a different org calibration.

[livereviewbot]

Auto-approve policy evaluation

Stage	Result
Eligibility	✅ pass
LLM classification (refactor-only, confidence 9)	✅ pass
Org risk calibration	risk_threshold_percentile: 10, same recent-score distribution: [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.65, 0.7, 0.8, 0.9] → P10 cutoff ≈ 0.15
Risk score vs. cutoff	❌ ~0.57 > 0.15

Verdict: BLOCKED at the risk stage. Same review, same underlying score as threshold-p80-lenient-approves-borderline-score -- the only thing that changed is this org's calibration dial. A strict org auto-approves almost nothing above its own recent 10th percentile; a lenient one auto-approves almost everything below its 80th. Neither is "wrong" -- the dial is the policy.

[livereviewbot]

Auto-approve policy verdict: BLOCKED (risk stage). Same risk score (~0.57) as the companion PR, but this org's strict P10 calibration cutoff sits below it.