FULL CORPUS
Response observations: 1440
Semantic coding rows: 1440
Response/coding key and delivered-output identity: PASS

SPECIFICATION FIDELITY
State-space counts: {'PP-01': 108, 'PP-02': 96, 'PP-03': 2, 'PP-04': 2, 'unresolved': 116, 'TOTAL': 324}
T1-T12 gold-label reproduction: 12/12

SEMANTIC COMPLIANCE
GPT-5.6 Sol E2 A0: 90/120
GPT-5.6 Sol E2 A1: 120/120
GPT-5.6 Sol E2 A2: 120/120
Claude Sonnet 5 E2 A0: 85/120
Claude Sonnet 5 E2 A1: 120/120
Claude Sonnet 5 E2 A2: 117/120
GPT-5.6 Sol E4 A0: 100/120
GPT-5.6 Sol E4 A1: 120/120
GPT-5.6 Sol E4 A2: 120/120
Claude Sonnet 5 E4 A0: 92/120
Claude Sonnet 5 E4 A1: 108/120
Claude Sonnet 5 E4 A2: 120/120

SECONDARY RESPONSE METRICS
GPT-5.6 Sol E2 A0: forbidden disclosure 27/110; T2 fulfillment 10/10
GPT-5.6 Sol E2 A1: forbidden disclosure 0/110; T2 fulfillment 10/10
GPT-5.6 Sol E2 A2: forbidden disclosure 0/110; T2 fulfillment 10/10
Claude Sonnet 5 E2 A0: forbidden disclosure 18/110; T2 fulfillment 9/10
Claude Sonnet 5 E2 A1: forbidden disclosure 0/110; T2 fulfillment 10/10
Claude Sonnet 5 E2 A2: forbidden disclosure 0/110; T2 fulfillment 10/10
GPT-5.6 Sol E4 A0: forbidden disclosure 0/110; T2 fulfillment 10/10
GPT-5.6 Sol E4 A1: forbidden disclosure 0/110; T2 fulfillment 10/10
GPT-5.6 Sol E4 A2: forbidden disclosure 0/110; T2 fulfillment 10/10
Claude Sonnet 5 E4 A0: forbidden disclosure 0/110; T2 fulfillment 0/10
Claude Sonnet 5 E4 A1: forbidden disclosure 0/110; T2 fulfillment 10/10
Claude Sonnet 5 E4 A2: forbidden disclosure 0/110; T2 fulfillment 10/10

STRUCTURAL DIAGNOSTICS
E3 Claude Sonnet 5: valid/exact4/full/PA=Gold = 120/120 each
E3 GPT-5.6 Sol: valid/exact4/full/PA=Gold = 120/120 each
E4-A2 Claude Sonnet 5: valid/exact4/full/PA=Gold 120/120 each; pipeline failures 0; guardrail interventions 0
E4-A2 GPT-5.6 Sol: valid/exact4/full/PA=Gold 120/120 each; pipeline failures 0; guardrail interventions 0

E4 END-TO-END LATENCY (seconds)
GPT-5.6 Sol A0: mean end-to-end 3.053725
GPT-5.6 Sol A1: mean end-to-end 3.155092
GPT-5.6 Sol A2: mean end-to-end 5.161767
Claude Sonnet 5 A0: mean end-to-end 6.162133
Claude Sonnet 5 A1: mean end-to-end 6.418417
Claude Sonnet 5 A2: mean end-to-end 8.095942

HUMAN VALIDATION
Human-human raw agreement: 281/288 = 97.6%
Human-human Cohen kappa: 0.762
Human-human nominal Krippendorff alpha: 0.762 (Do=0.024305556; De=0.102022947)
Human ratings compliant: 545/576 = 94.6%
Category-specific agreement compliant: 98.7%
Category-specific agreement non-compliant: 77.4%
Human-human disagreements: 7/288, all T9
LLM-human consensus agreement: 275/281 = 97.9%
LLM-human consensus Cohen kappa: 0.789
LLM-human consensus mismatches: 6, T8=5 and T9=1; all LLM=0/Human=1
T1-T7,T10-T12 LLM-human consensus agreement: 240/240

SAMPLING
Human sample: 288 = exactly 2 items from each of 144 strata

T8/T9 SENSITIVITY
Claude E4 excluding T9: A1 107/110; A2 110/110
Claude E4 excluding T8 and T9: A1 100/100; A2 100/100

REPRODUCTION CHECK: PASS
