{ "schema_version": 2, "run_id": "paper-final", "config": { "behavior_languages": [ "EN", "DE", "FR", "ES", "IT", "PT", "ID", "ZH", "JA", "KO", "AR", "HI", "BN", "SW", "YO" ], "behavior_n": 500, "high_power_items": 300, "knowledge_margin": 0.0, "knowledge_mode": "intersection", "languages": [ "EN", "DE", "FR", "ES", "IT", "PT", "ID", "ZH", "JA", "KO", "AR", "HI", "BN", "SW", "YO" ], "measure_n": 300, "minimum_primary_items": 100, "models": [ "q35_08b:Qwen/Qwen3.5-0.8B", "q35_2b:Qwen/Qwen3.5-2B", "q35_4b:Qwen/Qwen3.5-4B", "q35_9b:Qwen/Qwen3.5-9B", "q35_27b:Qwen/Qwen3.5-27B", "q25_3b:Qwen/Qwen2.5-3B-Instruct", "q25_7b:Qwen/Qwen2.5-7B-Instruct", "llama8b:meta-llama/Llama-3.1-8B-Instruct", "mistral7b:mistralai/Mistral-7B-Instruct-v0.3", "gemma9b:google/gemma-2-9b-it", "gemma4_12b:google/gemma-4-12b-it" ], "n": 1000, "patch_languages": [ "EN", "ZH", "AR" ], "patch_models": [ "q35_4b:Qwen/Qwen3.5-4B", "q25_3b:Qwen/Qwen2.5-3B-Instruct", "q25_7b:Qwen/Qwen2.5-7B-Instruct", "llama8b:meta-llama/Llama-3.1-8B-Instruct", "mistral7b:mistralai/Mistral-7B-Instruct-v0.3", "gemma9b:google/gemma-2-9b-it", "gemma4_12b:google/gemma-4-12b-it" ], "patch_n": 64, "pool_limit": 0, "random_sets": 49, "top_k": 20, "prompt_judges": [ "google/gemma-4-12b-it", "Qwen/Qwen3.5-27B" ], "behavior_judges": [ "google/gemma-4-31B-it", "Qwen/Qwen3.5-27B" ] }, "config_hash": "9e92809b6a23bb9f", "git": { "commit": null, "dirty": null, "worktree_hash": null }, "code_hash": "65d5ac597322ce2a", "runtime": { "python": "3.11.14", "platform": "Linux-7.0.0-28-generic-x86_64-with-glibc2.35", "packages": { "torch": "2.9.0+cu130", "transformers": "5.17.0", "datasets": "5.0.1", "numpy": "2.3.3", "scikit-learn": "1.9.0", "matplotlib": "3.11.2" }, "torch_cuda": "13.0", "cuda_available": true, "gpu": "NVIDIA H200 NVL" }, "created_utc": "2026-09-20T14:56:09.448745+00:00", "paths": { "project_root": "/work", "run_root": "/work/results/paper-final", "raw_results": "/work/results/paper-final/raw", "derived_results": "/work/results/paper-final/derived", "paper_root": "/work/paper", "figures": "/work/paper/figures", "tables": "/work/paper/tables", "logs": "/work/logs/paper-final" }, "compatible_code_migrations": [ { "from": "def431fe2b9a7185", "to": "e1becbd15b1b59fa", "reason": "memory-only AtP fix: capture final-token gradient slice and disable KV cache; no prompt, item, metric, or estimator change" }, { "from": "e1becbd15b1b59fa", "to": "726a4f16231ee76b", "reason": "OOM-control-flow fix: unwind failed CUDA frame before iterative batch splitting; no prompt, item, metric, or estimator change" }, { "from": "726a4f16231ee76b", "to": "594aaa59e36b25a0", "reason": "Gemma-4 geometry fix: support per-layer head dimensions for alternating local/global attention; no prompt, item, metric, or estimator change" }, { "from": "594aaa59e36b25a0", "to": "af704ce446ac1564", "reason": "correct per-layer head-slice closure variable in exact patching and interventions; no prompt, item, metric, or estimator change" }, { "from": "af704ce446ac1564", "to": "40b2d13e1e30fb61", "reason": "use the largest valid positive-head set when fewer than requested satisfy matched-control constraints, and record actual set size; no prompt, item, or metric change" }, { "from": "40b2d13e1e30fb61", "to": "6498407ce56c1457", "reason": "split prompt/behavior judge roles; pin Gemma-4-31B; score every subject with both behavior judges and retain cross-family sensitivity" }, { "from": "6498407ce56c1457", "to": "28cabdd067e6007d", "created_utc": "2026-09-25T17:54:12.842076+00:00", "reason": "iterative ordered OOM batch splitting and disabled KV cache in persuasion reads; no prompts, item selection, estimators, labels or metric definitions changed", "smoke_run": "oom-retry-smoke-20260925", "retained_outputs": "all complete earlier results remain valid; failing q35_27b persuasion had no output" }, { "from": "28cabdd067e6007d", "to": "333a8071da6e62c0", "created_utc": "2026-09-26T09:29:03.236759+00:00", "reason": "validation distinguishes capped category-balanced selections from exhaustive smaller intersection-known pools; manuscript now discloses their uneven categories; no experiment or estimator changed", "smoke_run": "validation-fix-smoke-20260926", "retained_outputs": "all 11-model raw results and completed analysis/artifacts are unchanged; paper build must reflect manuscript edit" }, { "from": "333a8071da6e62c0", "to": "af4abc6113cff8ef", "created_utc": "2026-09-26T09:41:35.023763+00:00", "reason": "add exploratory Pareto analysis, appendix figure/table, and validation using existing head-count and strength sweeps; no changes to raw model evaluations", "smoke_run": "tradeoff-test-20260926", "retained_outputs": "all raw model measurements remain valid; analyze, figures, paper and validate will be regenerated" }, { "from": "af4abc6113cff8ef", "to": "9e209f7044052cb3", "created_utc": "2026-09-26T09:52:05.924790+00:00", "reason": "correct reported 11-model cohort and two-judge figure label; enlarge appendix trade-off table for legibility; raw data and analyzed metrics unchanged", "smoke_run": "paper-polish-test-20260926", "retained_outputs": "raw results and summary unchanged; regenerate figures, tables, paper, and validation" }, { "from": "9e209f7044052cb3", "to": "6243110480575c8f", "created_utc": "2026-09-26T10:19:12Z", "reason": "final submission citation, claim-scope, low-N and figure/table presentation corrections; source fingerprint now includes the figure generator and excludes generated tables and macOS resource forks; no experiment inputs, raw measurements, or analysis metrics changed", "retained_outputs": "all raw results and analysis summary remain valid; rerun tests, figures, paper, and validation with regenerated artifact hashes" }, { "from": "6243110480575c8f", "to": "9af6686ba5c9ef14", "created_utc": "2026-09-26T10:24:34Z", "reason": "fixed appendix figures and trade-off table at their source positions after visual review found floats before headings and sparse pages; no experiment or metric changed", "retained_outputs": "raw results and analysis summary remain valid; regenerate the trade-off table, rebuild the paper, and revalidate" }, { "from": "9af6686ba5c9ef14", "to": "d0b9e12c84657b30", "created_utc": "2026-09-26T10:28:55Z", "reason": "remove draft-only prose, hide citation link boxes, and prevent stretched appendix gaps; no figure data, experiment input, or metric changed", "retained_outputs": "raw results, analysis summary, and generated data figures and tables remain valid; rebuild the paper and revalidate" }, { "from": "d0b9e12c84657b30", "to": "310bdd546e203564", "created_utc": "2026-09-26T11:58:35Z", "reason": "add an aligned opposite-fold item bootstrap for the already measured signed head maps; redesign data figures and rewrite manuscript with narrower speaker, patch-panel, and direction claims; no model forwards, prompts, item selection, or raw measurements changed", "parent_summary_sha256": "ee29384264e2d13e3a830f58a028ed4e43ecd227a66caa98e4e7c751d8271b50", "parent_validation_sha256": "16d98673276a22bc607016d57ca1eb4c20fa1fed8281c3cd50f9db5fe0baf7ec", "checked_raw_sidecar": "eapig_items_q35_4b_EN.npz", "checked_raw_sha256": "a16885bd658681370fd91a07f0bf572f6ca02ac36f0421e0eca2fd6d82e09dac", "smoke_run": "paper-revision-test-hcc", "retained_outputs": "all validated raw results are unchanged; rerun analysis, figures, paper build, and validation from the parent run" }, { "from": "310bdd546e203564", "to": "8d3985e8c2eee520", "created_utc": "2026-09-26T12:14:04Z", "reason": "correct a visually verified teaser-label collision and clarify the >=300-item figure title; add an appendix display of saved ZH/AR source-selected interventions with their three-checkpoint scope; no prompts, selected items, raw measurements, estimators, or analysis summary changed", "parent_summary_sha256": "1b02ca5e297afa7c3222ea4cb988325510258e297a0996d687df98b12e7cbcf7", "parent_validation_sha256": "841faff636670c5c7a4997fa6bf21c5e39fa7c4f703db91270fc43e51613cb24", "checked_raw_sidecar": "eapig_items_q35_4b_EN.npz", "checked_raw_sha256": "a16885bd658681370fd91a07f0bf572f6ca02ac36f0421e0eca2fd6d82e09dac", "retained_outputs": "validated raw measurements and unchanged analysis summary remain valid; rerun tests, figures, paper build, and validation" }, { "from": "8d3985e8c2eee520", "to": "6e6e078c27d6a65f", "created_utc": "2026-09-26T15:32:51Z", "reason": "rewrite the abstract, introduction, related work, results interpretation, discussion, and conclusion; move the illustrative figure into methods and secondary direction analysis to the appendix; no prompts, selected items, raw measurements, estimators, figures, or analysis metrics changed", "parent_summary_sha256": "1b02ca5e297afa7c3222ea4cb988325510258e297a0996d687df98b12e7cbcf7", "parent_validation_sha256": "ed7f8b1e72d5de1033dfaf7f06ded3780d0b7658377694a695f5e7698a3920d3", "checked_raw_sidecar": "eapig_items_q35_4b_EN.npz", "checked_raw_sha256": "a16885bd658681370fd91a07f0bf572f6ca02ac36f0421e0eca2fd6d82e09dac", "retained_outputs": "validated raw measurements, analysis summary, and generated data figures and tables remain unchanged; rebuild the paper and revalidate" }, { "from": "6e6e078c27d6a65f", "to": "57b0d2ddaec0e934", "created_utc": "2026-09-26T15:43:40Z", "reason": "repair duplicated author-year citations and place the unchanged teaser after the map definition; prevent floats from crossing sections during PDF layout review; no experiment inputs, raw results, analysis, or figure data changed", "parent_summary_sha256": "1b02ca5e297afa7c3222ea4cb988325510258e297a0996d687df98b12e7cbcf7", "parent_validation_sha256": "8df5f7f705dd9ed0a4f25e65d84328d06296defcbe86fc1839040cb38ca2824f", "checked_raw_sidecar": "eapig_items_q35_4b_EN.npz", "checked_raw_sha256": "a16885bd658681370fd91a07f0bf572f6ca02ac36f0421e0eca2fd6d82e09dac", "retained_outputs": "validated raw results, analysis, figures, and tables remain unchanged; rerun tests, paper, and validation" }, { "from": "57b0d2ddaec0e934", "to": "2f659d313ab0ef46", "created_utc": "2026-09-26T15:49:48Z", "reason": "keep the methods teaser ahead of the patch explanation, move the unchanged appendix setup table after the coverage figure to avoid a sparse reference page, and describe patch overshoot accurately; no measurement, estimator, or figure-data change", "parent_summary_sha256": "1b02ca5e297afa7c3222ea4cb988325510258e297a0996d687df98b12e7cbcf7", "parent_validation_sha256": "fe5f01b671c15e972d9e2b290ef05a6c546d82944dca4d149d116a16ca40b551", "checked_raw_sidecar": "eapig_items_q35_4b_EN.npz", "checked_raw_sha256": "a16885bd658681370fd91a07f0bf572f6ca02ac36f0421e0eca2fd6d82e09dac", "retained_outputs": "validated raw measurements and derived analysis/figures/tables remain unchanged; rebuild paper and revalidate" }, { "from": "2f659d313ab0ef46", "to": "a5bfda4c5f82680f", "created_utc": "2026-09-26T15:53:51Z", "reason": "keep the specificity explanation adjacent to its unchanged figure without splitting a sentence around a page-top float; presentation only, with all measurements and data artifacts unchanged", "parent_summary_sha256": "1b02ca5e297afa7c3222ea4cb988325510258e297a0996d687df98b12e7cbcf7", "parent_validation_sha256": "877b4f5b03914c0298ae17731fcb522df801cd3a281f8985a7260ebe9c3d0b24", "checked_raw_sidecar": "eapig_items_q35_4b_EN.npz", "checked_raw_sha256": "a16885bd658681370fd91a07f0bf572f6ca02ac36f0421e0eca2fd6d82e09dac", "retained_outputs": "validated raw results, analysis summary, figures, and tables unchanged; rerun tests, paper, validation" }, { "from": "a5bfda4c5f82680f", "to": "65d5ac597322ce2a", "created_utc": "2026-09-26T16:54:26Z", "reason": "replace decorative off-white figure and axes surfaces with plain white and replace the hard-coded near-chance AUROC shading with a labeled 0.5 reference line; presentation only, with all measurements and analysis unchanged", "parent_summary_sha256": "1b02ca5e297afa7c3222ea4cb988325510258e297a0996d687df98b12e7cbcf7", "parent_validation_sha256": "5235500102d1fd29da9c33648e8dc80e45e6b81f9a72d6d8bb34e62c21b97512", "checked_raw_sidecar": "eapig_items_q35_4b_EN.npz", "checked_raw_sha256": "a16885bd658681370fd91a07f0bf572f6ca02ac36f0421e0eca2fd6d82e09dac", "retained_outputs": "validated raw model results and analysis summary unchanged; regenerate figures, tables, paper, and validation" } ], "compatible_config_migrations": [ { "created_utc": "2026-09-25T08:14:44.721682+00:00", "from_config_hash": "a5d5b6349448299f", "to_config_hash": "9e92809b6a23bb9f", "from_code_hash": "40b2d13e1e30fb61", "to_code_hash": "6498407ce56c1457", "old_judges": [ "google/gemma-4-12b-it", "Qwen/Qwen3.5-27B" ], "prompt_judges": [ "google/gemma-4-12b-it", "Qwen/Qwen3.5-27B" ], "behavior_judges": [ "google/gemma-4-31B-it", "Qwen/Qwen3.5-27B" ], "behavior_policy": "both judges score every subject; all-judge ensemble primary; cross-family sensitivity retained", "reason": "Replace Gemma-4-12B behavior judge with Gemma-4-31B after weak judge-letter diagnostic agreement; all judge-dependent behavior outputs archived and regenerated from the beginning", "archived_files": [ "freeform_judge_q35_08b.json", "freeform_judge_q35_2b.json", "freeform_judge_q35_4b.json", "persuade_q35_08b.json", "persuade_q35_2b.json" ], "archive": "archived_behavior_gemma4_12b/archive_manifest.json", "smoke_run": "judge31-behavior-smoke-20260925", "retained_outputs": "knowledge, causal maps, exact patches, interventions, symmetry controls, and transfer are judge-independent" } ] }