Summary
The Phase-19 calibration policy defines five monitoring bands by fixed quantile cuts on the test-fold structural-distress score distribution. The bands are operator-facing triage primitives, not automated decisions. The Watch band (80–95th percentile) is the most informative tier; the top-1% Critical band shows a tail dropoff and is noise-prone.
The five predeclared bands
| Band | Percentile | n in band | Pos in band | Event rate | Lift | Recall captured |
|---|---|---|---|---|---|---|
| Normal / background | 0–60 | 11,646 | 49 | 0.42% | 0.35 | 21% |
| Elevated | 60–80 | 3,882 | 76 | 1.96% | 1.65 | 33% |
| Watch | 80–95 | 2,911 | 90 | 3.09% | 2.60 | 39% |
| High structural concern | 95–99 | 745 | 15 | 2.01% | 1.69 | 6.5% |
| Critical / inspect manually | 99–100 | 226 | 1 | 0.44% | 0.37 | 0.4% |
Headline reads
- Watch band (80–95) is the most informative tier. Event rate 3.1%, lift 2.6×, captures 39% of strict positives in 15% of the universe.
- Bottom 60% is "below the screening line", not "safe". 21% of strict positives still live below the 60th percentile.
- High structural concern (95–99) is similar to Elevated in lift but captures only 6.5% of positives.
- Top-1% Critical band has lower event rate than the Watch band. See the tail-dropoff notice below.
Tail-dropoff notice
The Critical / top-1 % band does not have the highest event rate. Its lift is below 1.0 (worse than random for that slice). This is consistent with the Phase-18 pooled top-1% lift of 0.43×, and is honestly flagged here. Operators should treat the top 1% as noise-prone, not as the highest-priority flag. The Watch band is the actionable tier; the Critical band is for human review of extreme outliers (some of which are real distress, some of which are data-quality outliers).
Why these bands and not F1-tuned cutoffs
- Monitoring-inspection capacity scales as a fixed fraction of the universe; quantile cuts match that operating reality.
- F1 collapses at low base rate (here ~ 1.2%); tuning to it would over-fit threshold to a single fold.
- The cuts are predeclared, not chosen to maximize any metric.
What this is and is not
These bands are research-grade triage primitives. They are not automated decisions, not deployment thresholds, and not calibrated probabilities. Each company appearing in any band requires manual inspection. Lift 2.6× on a 1% base rate means roughly 3% precision in the Watch band — useful for triage, not for action.
CORPORATE_PHASE19_CALIBRATION_POLICY.md and
processed/phase19_threshold_policy_table.csv
in the project repository.
Back to Validation · Evidence · Documents index.