Skip to content

Latest commit

 

History

History
40 lines (30 loc) · 1.61 KB

File metadata and controls

40 lines (30 loc) · 1.61 KB

Score robustness

The harness is deterministic (proven by the temp-0 answer/judge probe in docs/audits/eval-remediation-2026-07-11.md), so re-running gives the same score. These figures say how precise that score is.

Overall: 192/201 = 95.5% (Wilson 95% CI: 91.7%–97.6%).

Per-suite pass rate with 95% CI

Suite Passed Total Rate 95% CI
conversation 8 10 80.0% 49.0%–94.3%
cross_agency 3 3 100.0% 43.9%–100.0%
edge_cases 46 48 95.8% 86.0%–98.8%
freshness 9 10 90.0% 59.6%–98.2%
groundedness 27 29 93.1% 78.0%–98.1%
multilingual 22 22 100.0% 85.1%–100.0%
refusal 34 34 100.0% 89.8%–100.0%
sensitivity 28 30 93.3% 78.7%–98.2%
stretch_tagalog 15 15 100.0% 79.6%–100.0%

Leave-one-suite-out (jackknife)

Change in the overall rate, in points, when each suite is dropped:

Suite dropped Δ overall (points)
refusal -0.91
conversation +0.81
multilingual -0.55
groundedness +0.41
sensitivity +0.38
stretch_tagalog -0.36
freshness +0.29
edge_cases -0.10
cross_agency -0.07

Next: paraphrase sensitivity

Determinism covers identical inputs. The open question is whether the score moves when a question is reworded to the same meaning. The planned experiment: hand-author meaning-preserving paraphrases for a stratified sample, run both versions live, and report the pass/fail flip rate. It needs live generation and a labeled sample, so it is specified here rather than computed from a finished run.