Skip to content
ELAGrades 6-8

OpenSciEd 6-8 Science

OpenSciEd

Judge anthropic/claude-opus-4-7 · Signal Studio judge (claude-opus-4-7) · 2026-05-28

90

% indicator exact match

Gateways 3/3 Meets

Agreement vs gold

Indicator level — gold and the judge as two raters over the rubric criteria.

Exact match

90.2%

41 indicators vs gold

Weighted κ

0.90

ordinal-weighted agreement

MAE

0.12

lower is better

Signed bias

+0.12

judge over-scores

Attribution

Where the judge diverges, and whether it's signal: per-gateway bias localizes the gap, the confusion matrix shows over- vs under-rating, and the noise floor is the judge's self-agreement across passes (a gap below it isn't trustworthy).

gateway-onen=977.8%bias +0.33
gateway-threen=887.5%bias +0.13
gateway-twon=2495.8%bias +0.04

Indicator confusion · rows = gold, cols = judge

MeetsPartialDNM
Meets3700
Partial400
DNM000

Diagonal = agreement. Amber = judge rated higher than gold (over-rating).

Gateway rollup

Indicator scores rolled up to EdReports gateway ratings (sequential gating + no-0s cap), gold vs judge.

G1 · Designed for NGSS

26/26 pts

GoldMeets
JudgeMeets
agree

G2 · Coherence & Scope

56/56 pts

GoldMeets
JudgeMeets
agree

G3 · Usability

18/18 pts

GoldMeets
JudgeMeets
agree

Gateway-level agreement: exact 100% · κ 1.00 (3 gateways)

Divergences (4)

Indicators where the judge's score differed from gold. Amber = judge under-scored, blue = over-scored.

1c
24
gold Partially Meets Expectations · judge 4 pts
3b
12
gold Partially Meets Expectations · judge 2 pts
1h
12
gold Partially Meets Expectations · judge 2 pts
2e-v
12
gold Partially Meets Expectations · judge 2 pts
Live read from the Signal Studio spine (core-data).