Purpose

Turn saved experiment artifacts into compact evidence. Do not treat ordinary successful runs as requiring a long report, and do not choose the winning method.

Inputs

• run_summary.json
• method card and probe summary
• decision ledger
• saved tables, metrics, and figures
• session rigor_profile

Stop if the run summary claims outputs that do not exist or if main and baseline are not comparable.

Modes

Ordinary lean round

• Validate the run summary and referenced artifacts.
• Return a compact evidence digest in the conversation.
• Do not save a Markdown report unless:
  • a fallback trigger fired;
  • a material anomaly or contradiction exists;
  • the human must make a proceed/adjust/fallback decision.

Decision-point round

Save:

results/Qx/experiments/roundN/qx_decision_report.md

Include only:

• main vs baseline metrics;
• output-degeneracy/concentration evidence;
• assumption or feasibility warnings;
• robustness evidence already available;
• fallback trigger state;
• unresolved trade-offs.

Then invoke decision-prompt-builder. After the human answers, route the answer to modeler-decision-logger.

Final/submission mode

Save:

results/Qx/reports/qx_final_result_analysis.md

Include:

• final main/baseline comparison;
• uncertainty and error;
• concentration/degeneracy interpretation;
• robustness links;
• limitations and applicable scope;
• exact source paths for numerical claims.

Rejection and Fallback

• Archive a method only after a human result_verdict or fallback_activation decision.
• Move rejected code and outputs to workspace/archived/<Qx>/<method>_REJECTED_roundN/.
• Add one compact history line to qx_method_card.md; do not create a separate iteration log.
• Do not archive from an AI suggestion alone.

Rules

• Do not fabricate metrics, comparisons, or interpretations.
• Separate facts from human verdicts.
• Do not create result-report-generator_modeler_decision.md.
• Do not repeat the full run summary; cite it and extract only decision-relevant evidence.
• Do not call a diagnostic reference a usable baseline.
• Do not generate paper prose.

Verification

• Every reported number resolves to a saved artifact.
• Main/baseline comparison uses the same split, unit, and metric definition.
• Output concentration and fallback trigger are addressed.
• Reports are generated only at decision points or final mode.
• Human verdicts are read from or appended to the canonical JSONL ledger.