FE-012C Results

Purpose

Report the current FE-012C manual multi-model replication outputs without expanding the theory.

Research Question

Do GPT, Claude, and Gemini converge on primitive grammar extraction when they receive the same packet instrument and primitive vocabulary?

Data Set

• Expected packets: 15
• Expected model responses per packet: 3
• Expected total JSON files: 45
• Valid JSON files: 45
• Missing files: 0
• Malformed files: 0

Model Participants

• GPT
• Claude
• Gemini

Procedure

Responses were read from provider-specific folders, grouped by packet_id, validated, and compared on entry primitive, exit primitive, dominant primitive, reasoning shape, transition structure, artifact recognition, and candidate missing primitives.

JSON Validity

• Valid JSON responses: 45
• Missing responses: 0
• Malformed responses: 0

Agreement Summary

• Entry primitive full agreement: 15/15
• Exit primitive full agreement: 11/15
• Dominant primitive full agreement: 12/15
• Reasoning shape full agreement: 9/15

Transition Agreement Summary

• Strong transition agreement: 14/15
• Moderate transition agreement: 1/15
• Weak transition agreement: 0/15
• Unclear transition agreement: 0/15

Recognition Summary

• Packets with at least one recognized artifact: 15/15

Candidate Missing Primitive Summary

• Packets with at least one candidate missing primitive: 0/15

Key Factual Findings

• All counts in this report were computed from current response files.
• Agreement was strongest on entry primitive and reasoning shape if the current response set remains complete.
• Exit primitive and dominant primitive disagreements persisted in a minority of packets.
• Artifact recognition occurred often enough to weaken blinding.
• Candidate missing primitive absence, when present, should be interpreted cautiously because the provided vocabulary may constrain responses.