Search cases, suites, agents…⌘K
DEMO · synthetic data
Judges/registry

policy-answer-judge@v4

gate-eligible
owner · aag-qualitymodel · gemini-2.5-protemplate · aag/policy-answer
κ 0.91vs ≥ 0.80false-pass 0.8% · 1/121vs < 2%95% CI 0.86–0.95 · 214 calibration examples · since 2025-09

Rubric

v4 · plain language, SME audience
  • R1answer addresses the exact policy question asked
  • R2every factual claim grounded in the retrieved policy context
  • R3jurisdiction-specific policy cited when the subject's state has one
  • R4citations name the governing document section — paraphrase without a citation fails
  • R5effective dates stated when the policy changed within 12 months
Inherited
  • grounding-required
  • must-cite check
  • completeness
Overridden
  • jurisdiction-specificity clause (R3)
  • effective-date requirement (R5)

Calibration record

214 examples

How the set grew · 118 seeded from golden double-labels · 61 promoted disagreements across v1–v4 · 35 ongoing drift samples

The dangerous direction
false-pass 0.8% · 1/121vs < 2%

judge passed an answer the SME failedthe dangerous direction — a false pass ships a bad answer

The cost direction · no gate bar
false-fail 8.6% · 8/93

judge failed an answer the SME passedcosts triage time, not customer trust — no gate bar, tracked for cost

Agreement trendκ per calibration window · 0.80 bar · SME↔SME ceiling shown
κ 0.80SME↔SME ceiling · κ 0.92 · n 40SME↔SME 0.92Apr · κ 0.90 · n 26Aprn 26May · κ 0.92 · n 24Mayn 24Jun · κ 0.91 · n 25Junn 25

The ceiling — a judge is calibrated against human labels, so SME↔SME agreement on the same items is the most it can honestly earn. Here that ceiling is κ 0.92 (n 40): two SMEs double-labeled 40 items — policy answers have crisp ground truth. Read the judge's κ 0.91 against that ceiling, not against 1.00 — the same scale the label pool holds humans to (0.60 floor, 0.80 strong).

What κ measures — and what it hides

Cohen's kappa, computed from this judge's real calibration set
The 2×2 — judge stamp × SME stamp
SME fail
SME pass
Judge fail
120
agree · bad answer caught
8
false fail — costs triage time
Judge pass
1
false pass — the dangerous direction
85
agree · good answer passed

n = 214 double-labeled verdicts · the two off-diagonal cells are the disagreements.

1
Raw agreement

205 of 214 verdicts — 96%. Looks strong.

2
Chance agreement

This judge fails 60% of answers; the SME fails 57%. Two stamps falling that way at random would already agree 51% of the time. Raw agreement flatters.

κ
Beyond luck

(observed − chance) / (1 − chance) = 0.91 — the 96% agreement rescaled by the 51% it would have hit at random. Agreement beyond luck, on a scale where 0 is a coin-weighted judge and 1 is perfect.

κ is symmetric — it cannot tell a judge that ships bad answers from one that wastes triage time. Two judges can share κ 0.91 and fail in opposite directions. Shipping a bad answer is not the same mistake as flagging a good one — which is why false-pass carries its own bar.

The humans are measured with the same instrument — the labeling pool's batch κ ladder lives in the label console.

Bias battery

five standard LLM-judge probes, run against the calibration set
5 within
  • verbositysame content, padded to 2× length with restated policy textΔ +1.8 ptsvs ±3 ptswithin

    v3 on 2.5-flash probed +6.8 — the fluency bias behind the 2026-03 auto-recall; v4 re-probed clean

  • positioncandidate answer placed first vs last in the judge contextΔ +0.7 ptsvs ±2 ptswithin
  • self-preferencematched-quality answers authored by the judge's own family vs a cross-family modelΔ +2.1 ptsvs ±3 ptswithin
  • formattingsame content rendered as a markdown table vs plain proseΔ +1.1 ptsvs ±3 ptswithin
  • sycophancyutterance embeds a confident wrong figure from the userΔ +1.3 ptsvs ±2 ptswithin
Same family as the fleet — All four judges run on Gemini (2.5-pro / 2.5-flash) — the same model family serving the agents under test, so self-preference is probed directly rather than assumed away. The platform floor adds a quarterly cross-family spot-check: a Claude-family shadow judge re-scores a 10% sample; disagreement beyond tolerance triggers recalibration.

Version history

every rubric or model change forces recalibration (§31)
  1. v1drafted2025-09-18

    drafted from aag/policy-answer template · shadow at κ 0.77

  2. v2recalibrated2025-11-04

    rubric R4 tightened after false-fails on paraphrased citations → recalibrated, certified 2025-12-02

  3. v3certified2026-01-20

    judge model gemini-2.0-pro → 2.5-flash for cost · certified 2026-02-10

  4. v3recalled2026-03-06

    weekly drift sample fell to κ 0.74; 2.5-flash over-trusted fluent-but-uncited answers. Auto-demoted to shadow, both consumer suites alerted, certificate event-invalidated

  5. v4certified2026-03-28

    model 2.5-pro, R4 hardened · recalibrated on 214 examples · certified 2026-05-14

Calibration certificate

the sign-off a gate trusts
Signed sme: hermione.g
2026-05-14 → expires 2026-11-14

Time-boxed and event-invalidated — expires after two quarters, and earlier on drift or any model / rubric change. Nothing is trusted forever.

Trust bar

platform floor + suite raises (§33)
Platform floor
κ ≥ 0.80 AND false-pass < 2%
AAG fleet minimum
Suite raise
κ ≥ 0.85 · hr-policy-release-gate
a high-stakes gate can demand stricter

Drift sampling

continuous double-labeling (§36)
6 live verdicts double-labeled weekly (sme: hermione.g) · next due Mon 2026-07-13

Consumers & runs

who binds this judge, and where it ran

SME economics

the leverage, per judge
SME time45 SME-min per quarter
Keeps trusted~1,240 gate verdicts / quarter
Human-review equivalent≈ 103 h full human review