outputs

Outputs Directory

This directory holds the raw and structured experiment artifacts that support the repository’s claims. Treat it as a read-mostly evidence store.

Top-level layout

Path pattern	Purpose
runs/	Original smoke and base-experiment outputs
final_gemini_experiment_*	Frozen-vs-adaptive final-experiment outputs
matched_budget_pilot_*	Frozen prompt-family pilot outputs
matched_budget_revision_*	Cross-only prompt-family revision outputs
3d_ethics_stability_*	3D Ethics prompt-rewriting runs, diagnostics, preflights, and audits
cache/teacher/	Cached teacher responses keyed by request hash

Most important run roots

Run root	Why it matters
runs/smoke_seed_101/	Fastest original sanity check
final_gemini_experiment_qwen_0p5b_seed17_checkpoint320/	Strongest completed checkpoint case study
matched_budget_revision_qwen_0p5b_smoke/	Best completed prompt-family follow-up
3d_ethics_stability_qwen_3b_scaffold_family_prospective_seed2801/	First clean held-out 3D support-state win
3d_ethics_stability_qwen_3b_scaffold_family_tournament_v2_3s_wvs_guarded_seed4523/	Second clean held-out 3D WVS-guarded named-criterion win
3d_ethics_stability_qwen_3b_scaffold_family_tournament_v2_4b_semantic_gate_seed4909/	Negative prospective selector-gap / fragility-boundary run
3d_ethics_stability_qwen_3b_scaffold_family_tournament_v2_5o_targeted_named_support_seed6803_dev/	Current strongest dev-only named-support hard-gate pass
3d_ethics_stability_qwen_3b_scaffold_family_tournament_v2_5p_targeted_named_support_seed6907/	Prospective no-launch that kept final_test locked after the seed-6803 pass
3d_ethics_stability_qwen_3b_scaffold_family_tournament_v2_10f_minimal_endpoint_exact_count_prospective_seed8563/	Later exact-count prospective held-out boundary: salience/fragility gains but sensitivity collapse
3d_ethics_stability_qwen_3b_scaffold_family_tournament_v2_10l_operation_artifact_salience_lift_fresh_seed8629_dev/	Latest operation-artifact fresh-dev no-launch; all v2.10i-l final_test access remained zero

How to read a corrected 3D run

Corrected 3D runs write under:

• stability_prompt_rewrite_runs/seed_<seed>/

Most useful files inside a corrected run root:

File	Purpose
stability_summary.json	Run-level metrics, gates, and interpretation limits
prompt_rewriting/rounds.json or prompt_rewriting/prompt_rewrite_progress.json	Prompt evolution across rounds
prompt_rewriting/final_prompt_selection.json	Selected prompt and selector rationale
data/access_log.json	Locked-split access log
diagnostics/selector_validity_diagnostic.*	Whether selector-dev ranking tracked held-out behavior
audit/audit_manifest.json	Blinded audit packet summary

3D run roots worth knowing

Early scaling and selector diagnostics

• 3d_ethics_stability_qwen_0p5b_expanded_smoke/: exposed the 0.5B format and sensitivity bottleneck
• 3d_ethics_stability_qwen_1p5b_expanded_smoke/: showed that 1.5B fixes the invalid-format sensitivity collapse
• 3d_ethics_stability_qwen_1p5b_selector_calibrated_expanded_smoke/: first selector-calibrated prospective 1.5B diagnostic

3B scaffold-family frontier

• 3d_ethics_stability_qwen_3b_scaffold_family_replay_seed211_dev/: first fresh dev replay where a teacher-generated family cleared all dev gates
• 3d_ethics_stability_qwen_3b_scaffold_family_prospective_seed307/: clean held-out no-launch under a frozen protocol
• 3d_ethics_stability_qwen_3b_scaffold_family_prospective_seed2801/: first clean held-out win by context_preserving_support_state_scaffold
• 3d_ethics_stability_qwen_3b_scaffold_family_tournament_v2_3s_wvs_guarded_seed4523/: second clean held-out win by named_criterion_no_import_update_scaffold
• 3d_ethics_stability_qwen_3b_scaffold_family_tournament_v2_4b_semantic_gate_seed4909/: negative prospective run showing the selector/fragility frontier is still real
• 3d_ethics_stability_qwen_3b_scaffold_family_tournament_v2_5o_targeted_named_support_seed6803_dev/: strongest current dev-only gate-clean named-support hybrid
• 3d_ethics_stability_qwen_3b_scaffold_family_tournament_v2_5s_localized_wvs_direction_seed7207_dev/: latest localized WVS-direction boundary; final_test locked
• 3d_ethics_stability_qwen_3b_scaffold_family_tournament_v2_10f_minimal_endpoint_exact_count_prospective_seed8563/: exact-count transfer held-out boundary with salience/fragility gains but sensitivity collapse
• 3d_ethics_stability_qwen_3b_scaffold_family_tournament_v2_10l_operation_artifact_salience_lift_fresh_seed8629_dev/: latest fresh-dev operation-artifact route boundary; no held-out launch

Archive policy

This directory is intentionally large because it stores claim-relevant raw predictions, access logs, split manifests, and config snapshots. For broad public distribution, keep reports/artifact_index.md and this README in git even if raw run roots are mirrored to Git LFS, a release asset, or another artifact store.

Reading pattern

1. Start with a top-level summary markdown or JSON file if one exists.
2. Inspect config_snapshot.json or the run report to confirm what was frozen.
3. Use prompt-selection artifacts before interpreting top-line metrics.
4. Check data/access_log.json before treating any result as claim-relevant.
5. Use ../docs/current_status.md and ../reports/artifact_index.md as the release-facing navigation layer.