research-document

ECR-000002 Pre-Registration

Research Area: General Research
Status: draft
Version: 0.1
Priority: medium
Updated: 2026-07-22
Source: research/evidence-runs/ECR-000002-recognition-sensitivity/pre-registration.md

Purpose

Run a recognition-sensitivity study before making stronger claims about procedural extraction.

Recognition rate should be highest in canonical packets.
Recognition rate should decrease from canonical to paraphrased to structural.
P001D should reduce lexical recognition cues more than P001C.
If topology alone is sufficient for recognition, P001D graph_only may still trigger recognition despite stronger lexical reduction.
If structure extraction is not recognition-dependent, structural and constraint layers should remain relatively stable within each artifact family.
Primitive layer may drift more than structural or constraint layer.
Static or classification packets should remain less procedurally rich than reasoning or execution packets.
Product relevance should remain observational and should not dominate extraction.

Recognition persists and structure remains stable.

Interpretation: Possible structural recognition.

Recognition drops and structure remains stable.

Interpretation: Extraction may be topology-driven rather than recognition-driven.

Recognition drops and structure destabilizes.

Interpretation: Recognition may contribute materially to extraction.

Recognition persists but models introduce forbidden domain terminology.

Interpretation: Prior-knowledge leakage remains a major threat.

Structure collapses when recognition decreases.
Structural packets produce unrelated outputs across models.
Static packets appear as procedurally rich as reasoning packets.
Models fill in canonical details not present in paraphrased or structural packets.
Recognition remains high even for structural packets.
Recognition remains high in P001D, but extracted structure becomes unstable or speculative.