ECR-000003 Run Instructions

Execution Mode

Manual execution only.

Fresh conversation for every packet.

No previous outputs visible.

No model sees another model's response.

Recommended Run Order

Phase 1

Run EXP-001 only.

Run GPT, Claude, and Gemini across the four topology variants.

After each model output, operator should record recognition_value using metrics/recognition-scale.md before reading detailed structure.

Use P001D packet version 1.1.

Any prior P001D responses from version 1.0 are excluded from the corrected primary comparison.

P001D must be rerun for GPT, Claude, and Gemini after the fix.

P001A, P001B, and P001C do not need rerunning unless their responses are missing or invalid.

Stop and review.

Phase 2

Run EXP-002 only if EXP-001 results justify continuing.

Phase 3

Run EXP-003 only after EXP-002 review.

Filename Convention

responses/<provider>/<packet_id>-<provider>.json

Example:

experiments/EXP-001-topology-perturbation/responses/gpt/ECR-000003-EXP001-P001A-gpt.json

Comparison Step

After all corrected EXP-001 responses are present:

1. Change into experiments/EXP-001-topology-perturbation/.
2. Run npm run compare.
3. Review comparison/generated-v3.1/data-quality-report.md first.
4. Review comparison/generated-v3.1/recognition-persistence-report.md.
5. Review comparison/generated-v3.1/structural-stability-report.md.
6. Treat Comparator v3.1 as frozen for ECR-000003 after limited approval.
7. Use research/evidence-runs/ECR-000003-representation-sensitivity/edr/EDR-ECR-000003-EXP001.md as the working EDR, with comparison/generated-v3.1/edr-draft.md as source material only.
8. Complete and accept EDR-ECR-000003-EXP001 before activating any later experiment.
9. Use the accepted EDR to decide whether EXP-002 should proceed.

EXP-002 may not begin until the accepted EDR selects it or another experiment as the highest-information next action.