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

comp-answer-judge@v1

shadow
owner · payroll-intelligencemodel · gemini-2.5-protemplate · aag/policy-answer
κ 0.71vs ≥ 0.80false-pass 3.4% · 1/29vs < 2%95% CI 0.55–0.82 · 62 calibration examples · since 2026-05-28

The 95% interval spans the 0.80 floor — at this sample size certification is not even decidable; more labeled examples are needed either way.

Shadow mode — scores everything, gates nothing until the agreement bar clears and an SME signs off.

Rubric

v1 · plain language, SME audience
  • R1answers the exact comp question
  • R2every figure traced to the comp statement or plan document
  • R3′figures match the comp system of record for the subject
  • R4governing plan / section cited
  • R5′effective dates on any rate or plan figure
Inherited
  • R1
  • R2
  • R4
Overridden
  • comp-statement grounding (R3′)
  • pay-figure attribution (R5′)

Calibration record

62 examples
The dangerous direction
false-pass 3.4% · 1/29vs < 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 24% · 8/33

judge failed an answer the SME passedburning triage time — a rubric cost, not a gate risk

Agreement trendκ per calibration window · 0.80 bar · SME↔SME ceiling shown
κ 0.80SME↔SME ceiling · κ 0.81 · n 30SME↔SME 0.8106-01 · κ 0.66 · n 1806-01n 1806-15 · κ 0.70 · n 2206-15n 2206-29 · κ 0.74 · n 2206-29n 22

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 0.81 (n 30): partial-period proration genuinely divides experts — the same theme the calibration session surfaced (§103). Read the judge's κ 0.71 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
28
agree · bad answer caught
8
false fail — costs triage time
Judge pass
1
false pass — the dangerous direction
25
agree · good answer passed

n = 62 double-labeled verdicts · the two off-diagonal cells are the disagreements.

1
Raw agreement

53 of 62 verdicts — 85%. Looks strong.

2
Chance agreement

This judge fails 58% of answers; the SME fails 47%. Two stamps falling that way at random would already agree 49% of the time. Raw agreement flatters.

κ
Beyond luck

(observed − chance) / (1 − chance) = 0.71 — the 85% agreement rescaled by the 49% 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 0.71 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
3 within · 2 exceeds
  • verbositysame content, padded to 2× length with restated policy textΔ +3.2 ptsvs ±3 ptsexceeds
  • positioncandidate answer placed first vs last in the judge contextΔ +1.1 ptsvs ±2 ptswithin
  • self-preferencematched-quality answers authored by the judge's own family vs a cross-family modelΔ +2.4 ptsvs ±3 ptswithin
  • formattingsame content rendered as a markdown table vs plain proseΔ +1.9 ptsvs ±3 ptswithin
  • sycophancyutterance embeds a confident wrong figure from the userΔ +4.9 ptsvs ±2 ptsexceeds

    user-asserted figures pull the judge toward pass — the same blind spot as the proration false-passes (§103); rubric v2 backtest re-probes at +1.7

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.

Battery run on the 62-example calibration set — small n, deltas carry wide error bars; re-run scheduled with each calibration batch.

Version history

every rubric or model change forces recalibration (§31)
  1. v1drafted2026-05-28

    drafted from aag/policy-answer with comp-statement grounding (R3′) + pay-figure attribution (R5′) · shadow at κ 0.71 on 62 examples

Recent calibration examples

every disagreement is gold (§34)
  • fp_loa_monththeme · partial-period prorationrecorded false-pass · v1 shadow

    leave-of-absence month priced as a full month

Path to gate-eligible

§32 · both bars, then sign-off
  • κ ≥ 0.80
    0.71
  • false-pass < 2%
    3.4%
  • SME sign-off
    pending — blocked until both bars clear
SME calibration workspace

A queue of live and shadow verdicts is waiting for SME double-labeling. Every disagreement grows the calibration set and can mint a rubric fix.

Open calibration workspace5 verdicts · ~12 min this week

Trust bar

platform floor + suite raises (§33)
Platform floor
κ ≥ 0.80 AND false-pass < 2%
AAG fleet minimum

Consumers & runs

who binds this judge, and where it ran

SME economics

the leverage, per judge
SME time~12 min this week
Keeps trusted214 shadow verdicts scored this month
Human-review equivalent≈ 18 h full human review