Evidence Decision Record

EDR ID: EDR-ECR-000003-EXP001

ECR / Experiment: ECR-000003 EXP-001 Topology Perturbation

Date: 2026-07-14

Reviewer:

Status: draft / reviewed / accepted

Question

What question were we trying to answer about recognition persistence under graph perturbation?

Competing Explanations Before Evidence

• recognition depends primarily on topology
• recognition depends primarily on lexical or semantic cues
• recognition behavior is provider-specific
• graph perturbations alter perceived structure enough to change extraction

Observation

• gpt recognition pattern classified as provider_specific_recognition.
• claude recognition pattern classified as gradual_decay.
• gemini recognition pattern classified as provider_specific_recognition.
• Structural stability profile is mostly_stable dimensionally and mostly_stable literally.
• Structural backbone profile is mixed; wording-sensitive literal profile is mostly_stable.
• 3 primary responses introduced watched domain terminology.
• 5 data-quality issues were recorded and require review before interpretation.

Evidence Inputs

• responses/gpt/ECR-000003-EXP001-P001A-gpt.json
• responses/claude/ECR-000003-EXP001-P001A-claude.json
• responses/gemini/ECR-000003-EXP001-P001A-gemini.json
• responses/gpt/ECR-000003-EXP001-P001B-gpt.json
• responses/claude/ECR-000003-EXP001-P001B-claude.json
• responses/gemini/ECR-000003-EXP001-P001B-gemini.json
• responses/gpt/ECR-000003-EXP001-P001C-gpt.json
• responses/claude/ECR-000003-EXP001-P001C-claude.json
• responses/gemini/ECR-000003-EXP001-P001C-gemini.json
• responses/gpt/ECR-000003-EXP001-P001D-gpt.json
• responses/claude/ECR-000003-EXP001-P001D-claude.json
• responses/gemini/ECR-000003-EXP001-P001D-gemini.json

Interpretation

Explanations Weakened

Explanations Surviving

Remaining Uncertainty

• ECR-000003-EXP001-P001D gpt: tolerant_parse (Smart-quote / tolerant JSON sanitization was required before parsing.)
• dataset : warning_1 (Tolerant parsing used for ECR-000003-EXP001-P001A gpt: responses/gpt/ECR-000003-EXP001-P001A-gpt.json)
• dataset : warning_2 (Tolerant parsing used for ECR-000003-EXP001-P001B gpt: responses/gpt/ECR-000003-EXP001-P001B-gpt.json)
• dataset : warning_3 (Tolerant parsing used for ECR-000003-EXP001-P001C gpt: responses/gpt/ECR-000003-EXP001-P001C-gpt.json)
• dataset : warning_4 (Tolerant parsing used for ECR-000003-EXP001-P001D gpt: responses/gpt/ECR-000003-EXP001-P001D-gpt.json)

Explainability Review

Highest-Priority Differences

Low-Significance Differences

Material Backbone Differences

Unresolved Classifications

Compression / Elaboration Findings

Recognition / Structure Relationship

Domain Leakage Impact

Surprise Index

low / medium / high

Why?

Efficiency Notes

• Primary records included: 12/12
• Data-quality issues recorded: 5

Decision

• hypothesis review only
• queue follow-up experiment
• rerun with instrumentation changes
• collect replication
• pause and defer

Next Highest-Information Experiment

Name:

Why this experiment next:

Queue Entries Created

Confidence Update Recommendation

Do not update confidence here.