FE-012C Repeatability Protocol

Purpose

Provide a controlled rerun of FE-012C using the same packet instrument, fresh model conversations, and locked prior results.

Core Rule

The repeatability run must test the same instrument rather than a revised one.

Instrument Lock

• Packets were copied from research/experiments/FE-012C-manual-replication/packets/.
• No packet content changes were introduced during setup.
• Do not edit packet text unless a broken placeholder or obvious transcription defect is found.
• If a packet must be corrected, record the exact difference before running any model.

Blinding Rules

• Do not show prior outputs to any model.
• Do not show one model another model's output.
• Use fresh GPT-5, Claude, and Gemini conversations.
• Do not append commentary about prior results.

Prompt Consistency

• Every provider must receive the full packet contents.
• Every provider must receive the same standard execution prompt.
• Do not provide clarifications.
• Do not ask follow-up questions on behalf of the models.

Response Handling

• Save raw GPT JSON to responses/gpt/<packet_id>-gpt.json.
• Save raw Claude JSON to responses/claude/<packet_id>-claude.json.
• Save raw Gemini JSON to responses/gemini/<packet_id>-gemini.json.
• Lock outputs before comparison.

Comparison Timing

Do not compare repeatability outputs against prior FE-012C results until all current outputs are collected and locked.

Interpretation Constraint

Repeatability evidence may strengthen or weaken confidence in the measurement instrument, but it does not by itself prove the underlying theory.