FE-013 Procedural Language Hypothesis

Purpose

FE-013 tests whether procedural frameworks can be analyzed like natural-language procedural programs by extracting compiler-like structural properties from their blinded text.

The experiment focuses on:

• entry conditions
• exit conditions
• preconditions
• postconditions
• branches
• loops
• invariants
• termination conditions
• procedural ASTs

Why This Follows FE-012C

FE-012C showed that models often converged on reasoning shape, transition structure, and dominant procedural elements even when literal primitive sequences diverged.

FE-013 follows that result by testing whether a programming-language-style structural representation is a more informative comparison layer than primitive sequence comparison alone.

How This Differs From Primitive Extraction

FE-013 is not a primitive extraction study.

It does not use the FE-012C primitive vocabulary.

It asks for compiler-like procedural properties and AST structure rather than primitive sequences.

Warning

This experiment does not validate Framework Engineering.

It tests a narrow working hypothesis about whether some procedural frameworks exhibit language-like structural properties recoverable from natural-language descriptions.

Directory Structure

• hypothesis.md
• protocol.md
• procedural-ast-schema.md
• constraint-extraction-guide.md
• packet-template.md
• artifact-set.md
• pre-registration.md
• analysis-plan.md
• threats-to-validity.md
• interpretation-rules.md
• packets/
• responses/gpt/
• responses/claude/
• responses/gemini/
• comparison/

Manual Workflow

1. Open a packet from packets/.
2. Paste the full packet into a fresh GPT, Claude, or Gemini conversation.
3. Save the raw JSON response into the matching provider folder under responses/.
4. Repeat until all provider outputs are collected.
5. Compare AST structure, constraints, and disagreement cases in comparison/.

No API Requirement

This experiment is designed to be run manually.

No API, SDK, or automation is required.