# Blinded Inter-AI Agreement — v1.0

**Experiment:** UNI-EXP-001A-R2

**Stage:** Exploratory AI Annotation

**Preregistered Human Annotation:** Pending

**Join:** 1,020/1,020 items matched one-to-one by blind identifier

**Blinding:** Restricted analysis key not accessed; no M/control identities or outcome rates used.

## Agreement by frozen field

| Field | Paired n | Raw agreement | Cohen's κ | AI-A positives | AI-B positives | Disagreements |
|---|---:|---:|---:|---:|---:|---:|
| HUMAN | 1,020 | 99.61% | 0.9832 | 136 | 140 | 4 |
| PERSON | 1,020 | 94.71% | 0.7004 | 72 | 126 | 54 |
| PEOPLE | 1,020 | 94.31% | 0.2393 | 64 | 14 | 58 |
| IDENTITY | 1,020 | 96.27% | 0.8249 | 104 | 142 | 38 |
| HUMAN-ATTRIBUTE | 1,020 | 88.92% | 0.6536 | 246 | 155 | 113 |
| UNCERTAIN | 1,020 | 99.90% | 0.0000 | 0 | 1 | 1 |

## 2×2 cross-tabulations

| Field | A0/B0 | A0/B1 | A1/B0 | A1/B1 |
|---|---:|---:|---:|---:|
| HUMAN | 880 | 4 | 0 | 136 |
| PERSON | 894 | 54 | 0 | 72 |
| PEOPLE | 952 | 4 | 54 | 10 |
| IDENTITY | 878 | 38 | 0 | 104 |
| HUMAN-ATTRIBUTE | 763 | 11 | 102 | 144 |
| UNCERTAIN | 1,019 | 1 | 0 | 0 |

## Exact six-label vector

- Exact vector agreement: **828/1,020 (81.18%)**
- Exact vector disagreements: **192**

## Interpretation notes

The largest disagreement concentration was in HUMAN-ATTRIBUTE, followed by PERSON, PEOPLE, and IDENTITY category boundaries. The low κ for PEOPLE and zero κ for UNCERTAIN should be read alongside their highly imbalanced positive-label frequencies. Raw agreement, positive counts, and cross-tabulations are therefore retained with κ.

## Stop status

Diagnostics remain blinded and contain no experimental/control assignment. No M versus Control rates were calculated. No disagreements were adjudicated. The restricted analysis key remained unopened for this stage. Work stopped pending a separate decision to unblind exploratory AI results or wait for preregistered human annotation.