Preconditions

• G2.5 human method choice is recorded in methods/Qx/qx_decisions.jsonl.
• code/Qx/qx_code_plan.md exists.
• Required cleaned data and profile exist.
• The plan targets Python.

Legacy method pools and code/model-code-analyzer.md may be read during migration, but they do not override the human choice.

Workflow

1. Read the code plan, decision ledger, method card, probe conditions, and data profile.
2. Confirm scope:
   • one approved main method;
   • one usable baseline;
   • fallback only when an activation decision or evidenced trigger exists.
3. Generate clear runnable .py files under code/Qx/.
4. Use project-root-safe paths, fixed seeds, explicit inputs, and minimal justified dependencies.
5. Save:
   • tables to results/Qx/experiments/roundN/tables/;
   • metrics to .../metrics/;
   • useful diagnostic/comparison figures to .../figures/;
   • canonical run_summary.json.
6. Evaluate and record output-degeneracy and fallback-trigger metrics required by the plan.
7. Persist full logs only on failure or when a warning needs reproduction.
8. Run the code. Do not claim success from code generation alone.
9. Hand off to code-reviewer.

Script Layout

Prefer the smallest clear layout:

code/Qx/
├── qx_code_plan.md
├── qx_baseline.py
├── qx_main.py
└── run_all.py        # only when coordination is useful

Do not create one script per unapproved candidate. Do not create a README that duplicates the code plan.

Run Summary

Follow the schema in model-code-analyzer. Include:

• approved decision ID;
• method IDs and roles;
• inputs and outputs;
• seed and environment;
• execution status and timing;
• compact metric summaries;
• output-degeneracy evidence;
• warnings/errors;
• fallback-trigger state.

Rules

• Do not change the approved model or baseline.
• Do not read or overwrite raw data.
• Do not hide assumptions in code.
• Do not emit placeholder metrics, figures, or successful statuses.
• Prefer portable .py scripts over notebook-only workflows.
• Keep intermediate files only when needed for explanation, review, robustness, or debugging.
• Use Type 1 diagnostic figures internally; do not present them as paper figures.

Verification

• Main and baseline both ran and are directly comparable.
• Fallback code is absent unless activated.
• Formal outputs and run summary exist.
• Seed, inputs, versions, warnings, and errors are recorded.
• Required concentration/degeneracy checks are saved.
• Next handoff is code-reviewer.