3. The borderline band is ±0.025¶
Status: Accepted · Date: 2026-08 · Evidence: docs/m0-findings.md, section 10
Decision¶
A result within ±0.025 of a decision threshold is reported as borderline. The width first proposed was ±0.005.
Context¶
The decision bands are 0.10 wide, and the decision they produce is what a user acts on. ARR is estimated from a sample, so a result near a boundary can land on either side depending on which documents were drawn. The borderline band exists to say so.
±0.005 is a claim about precision. It asserts that the measurement resolves ARR to half a percentage point.
Evidence¶
Bootstrap 95% CI half-width for ARR, measured:
| Tier | Sample | Half-width |
|---|---|---|
| T0 | ~1,000 held-out proxies | ±0.024 |
| T1 | real query logs | ±0.042 |
The proposed band was five to eight times narrower than the uncertainty it was supposed to describe. A result 0.01 from a threshold would have been reported as settled while the interval spanned the boundary.
Consequences¶
- The band matches the measurement instead of flattering it.
- More results are reported as borderline. That is the point: they are borderline, and the previous number was hiding it.
- The report also says when the confidence interval itself straddles a threshold, which is stronger than the band and is what a careful reader wants.
- Narrowing the band later requires narrowing the interval first — more samples, or a real query log — which is the honest order.
Alternatives¶
Keep ±0.005 and report the interval separately. Rejected: two signals that disagree, and the smaller number is the one people read.
Widen to ±0.042, the T1 figure. Rejected as too wide for T0, where the measurement is genuinely tighter. ±0.025 sits at the T0 uncertainty, and the interval check covers T1's wider spread.