Release alpha-tot-v1.0.4: probe-gated freeze and 22-model channel panel

.gitattributes

@@ -66,3 +66,7 @@ abc_exp/results/sa_causal_openrouter_llama_4_maverick_n200.jsonl filter=lfs diff
 abc_exp/results/sa_causal_openrouter_llama_4_scout_n200.jsonl filter=lfs diff=lfs merge=lfs -text
 abc_exp/results/sa_causal_openrouter_mistral_small_4_n200.jsonl filter=lfs diff=lfs merge=lfs -text
 abc_exp/results/sa_causal_router_deepseek-v4-pro.jsonl filter=lfs diff=lfs merge=lfs -text
+abc_exp/results/gate_rerun/gate_default_vllm_phi-4-mini_s0.jsonl filter=lfs diff=lfs merge=lfs -text
+abc_exp/results/gate_rerun/gate_default_vllm_phi-4-mini_s1.jsonl filter=lfs diff=lfs merge=lfs -text
+abc_exp/results/gate_rerun/gate_default_vllm_qwen3-4b_s0.jsonl filter=lfs diff=lfs merge=lfs -text
+abc_exp/results/gate_rerun/gate_default_vllm_qwen3-4b_s1.jsonl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -51,6 +51,7 @@ hf download XINLI1997/DebateLedger --repo-type dataset --local-dir .
 | `cross_benchmark_*.jsonl` (16 files) | GPQA, TruthfulQA and ARC-Challenge stress checks | 1,033 debates |
 | `block0_*`, `block1_*`, `block3_*` | Early pilot blocks | 1,705 records |
 | `r5_gemini_alpha_panel_*.jsonl` | Three-day closed-API repeatability panel | 190 records |
+| `gate_rerun/*.jsonl` (10 files) | Probe scores of Llama-3.1-8B, Phi-4-mini, Qwen3-4B and Qwen3-8B collected again for the probe-gated freeze and the channel panel: default condition on the debate questions, all three conditions on the 200-question pool | 24,335 records |
 
 Also included: the aggregate tables and audits behind the paper
 (`abc_exp/results/*.json|md|csv|tex`), the datasheet
@@ -92,10 +93,13 @@ it. The full text of the 57 files that contained it, of the thinking-off rerun,
 (convince-wrong) and social-pressure probe templates, are in the
 [gated tier](https://huggingface.co/datasets/XINLI1997/DebateLedger-gated).
 
+In the gate-rerun probe files (`abc_exp/results/gate_rerun/`) the model output (`initial_text` and `probe_results[].text`) is `null`; that text is not released in either tier.
+
 No model weights are included.
 
 ## Versions
 
+- `alpha-tot-v1.0.4`: adds the gate-rerun probe scores (`abc_exp/results/gate_rerun/`) and the results of the probe-gated freeze and the 22-model channel panel. Earlier traces are unchanged.
 - `alpha-tot-v1.0.3`: adds the N=9 sensitivity results. Traces are unchanged.
 - `alpha-tot-v1.0.2`: aggregate results now read every row at the paper's endpoint, Round 3 majority with alphabetic ties.
   Traces are unchanged: rows with a logged fourth round still store its majority in `majority_answer`, and

README_REPRO.md

@@ -2,6 +2,26 @@
 
 This repository contains the code and data for *Measuring Collapse and Correction in Homogeneous-Panel LLM Debate* (NeurIPS 2026, Evaluations and Datasets Track). The analyses below are zero-API: they read existing traces and write derived tables only.
 
+## Probe-Gated Freeze and Channel Panel (alpha-tot-v1.0.4)
+
+The probe-gated freeze, the 22-model channel panel, the collapse-diagnostic AUC table and the probe-reliability check of Section 4, and the N=9 partial/Holm rows rebuild from the open tier without API calls. The probe scores of the four local models (Llama-3.1-8B, Phi-4-mini, Qwen3-4B, Qwen3-8B) were collected again for this analysis with `abc_exp/scripts/run_alpha_probe_vllm.py` (vLLM; initial answers sampled at 0.5/0.7/1.0 with an 8,192-token cap, probe replies greedy with a 2,048-token cap, `--stable-seeds`): `abc_exp/results/gate_rerun/gate_default_vllm_*_s{0,1}.jsonl` hold the default condition on the debate questions outside the 200-question high-FR pool (`--questions abc_exp/results/gate_rerun/qids_<model>_s<k>.json --conditions default`), and `abc_exp/results/gate_rerun/sa_causal_vllm_*_rerun.jsonl` hold all three conditions on the high-FR pool. Model text is null in the open tier. The gate is a post-hoc reconstruction on these new probe scores; the scripts and probe inputs of the original pilot were not retained.
+
+```bash
+python abc_exp/experiments/gate_rerun_acceptance.py abc_exp/results/gate_rerun          # completeness, truncation and unparsed-reply rates
+for m in paper flip drop; do
+  python abc_exp/experiments/probe_gated_lomo.py --debates abc_exp/results --probes abc_exp/results/gate_rerun \
+    --none-mode $m --out abc_exp/results/gate_rerun/gate_lomo_$m.json
+done
+python abc_exp/experiments/gate_tex.py abc_exp/results/gate_rerun/gate_lomo_{paper,flip,drop}.json > abc_exp/results/gate_rerun/gate_tex.txt
+python abc_exp/experiments/channel_panel.py --out abc_exp/results/channel_panel/channel_panel_22.json --tex abc_exp/results/channel_panel/channel_panel_22.tex
+python abc_exp/experiments/fig2_channel_panel.py abc_exp/results/channel_panel/channel_panel_22.json abc_exp/results/figures/fig2_channel_manipulation
+python abc_exp/experiments/construct_auc_public_n200.py abc_exp/results abc_exp/results/construct_auc_public_n200.json
+python abc_exp/experiments/reliability_recheck.py
+python abc_exp/experiments/n9_partial_holm.py
+```
+
+In this release `conditional_collapse_table.py` keeps the recorded counts of the rows whose debate files are not public instead of dropping them, and leaves out the 25 Gemini 2.5 Flash rows that hold only an API error (118 valid rows); `post_a6_r1_replication.py`, `cascade_r1_auc.py` and `build_per_debate_features.py` read fixed input lists. `cascade_r1_bootstrap.py` writes the question-clustered intervals into `cascade_r1_auc.json`. Scripts that still do not run on the open tier: the Bayesian MLM scripts need the arviz/xarray versions they were written for, the `block*` experiment drivers call APIs, `split_alpha.py` needs a local label file, and the very-strong-drop and log-ratio checks (`T2_1_drop_very_strong.py`, `T2_4_log_ratio.py`) need N=9 probe files that are not public; their checked-in outputs are kept.
+
 ## Zero-API Rebuild
 
 `abc_exp/experiments/n9_gee_sensitivity.py` and `abc_exp/experiments/n9_perm_lomo_check.py` rebuild the GEE and GLM tables, Fisher 90% intervals, permutation null and leave-one-model-out R^2 of the N=9 sensitivity check (`abc_exp/results/N9_GEE_SENSITIVITY.md`, `n9_gee_sensitivity.json`, `n9_perm_r2_check.json`). `abc_exp/scripts/B_pilot_monotonicity.py` is the harness of the B pilot; it needs a GPU with vLLM.
@@ -66,7 +86,7 @@ These scripts recompute the family-level exact permutation test, N14 errors-in-v
 - `abc_exp/results/debate_traces_*.jsonl`: round-by-round debate traces used for parser and provenance audits on the realized expansion rows and the response-period checks. In the open tier the `reasoning` fields are `null`.
 - `abc_exp/results/sa_causal_*.jsonl`: probe traces used for alpha/parser audits. In the open tier the `initial_response_prefix` field is `null`.
 - `abc_exp/results/conditional_collapse_table.json`: sealed-vintage debate diagnostics for older rows.
-- `abc_exp/results/per_debate_r1_features.jsonl`: row-level R1 feature matrix for the reported Bayesian MLM. It was rebuilt on 2026-09-28 from the six primary-cohort trace files with `abc_exp/experiments/build_per_debate_features.py` and reproduces the MLM input exactly (3,145 initially correct debates from Llama-3.1-8B, Phi-4-mini, Qwen3-4B and Qwen3-8B, 251 collapses, 1,509 questions). The fitted MLM summary is checked in as `abc_exp/results/bayesian_multilevel_logistic.json`. This matrix is not required for the headline transition tables, parser audits, family-rank analyses, or signed-replay ledgers; `abc_exp/experiments/bayesian_multilevel_logistic.py` reads it directly. To rebuild it, run `build_per_debate_features.py` with `ABC_RESULTS` pointing to a folder that holds only the six primary-cohort files, because the script reads every `debate_traces_*.jsonl` in that folder.
+- `abc_exp/results/per_debate_r1_features.jsonl`: row-level Round 1 feature matrix of the six primary-cohort trace files (6,925 debates, 253 collapses), built with `abc_exp/experiments/build_per_debate_features.py`. The reported Bayesian MLM uses its 3,145 initially correct debates from Llama-3.1-8B, Phi-4-mini, Qwen3-4B and Qwen3-8B (251 collapses, 1,509 questions). The fitted MLM summary is checked in as `abc_exp/results/bayesian_multilevel_logistic.json`. This matrix is not required for the headline transition tables, parser audits, family-rank analyses, or signed-replay ledgers; `abc_exp/experiments/bayesian_multilevel_logistic.py` reads it directly. To rebuild it, run `build_per_debate_features.py`; it reads the six primary-cohort files by name.
 
 ## Protocol Pointers
 
@@ -108,8 +128,8 @@ The script validates that the card contains identity, question-pool, selection-t
 
 Results in the paper fall into four tiers (paper Appendix G.6):
 
-1. **Zero-API rebuilds** from the released traces and scripts: the family association, transition tables, parser audits, and the response-period summaries.
-2. **Aggregate-only results**: the 6,525-debate probe-gated leave-one-model-out ledger is released as a frozen aggregate, and its row-level decision matrix is not part of the release. The Round 1 replay rows use a separate 1,255-trace cohort.
+1. **Zero-API rebuilds** from the released traces and scripts: the family association, transition tables, parser audits, the response-period summaries, and the 6,525-debate probe-gated leave-one-model-out ledger, whose gate script and per-question probe scores are released.
+2. **Aggregate-only results**: the Round 1 replay rows use a separate 1,255-trace cohort.
 3. **Provider-dependent reruns**: closed-API rows can drift as providers update models.
 4. **Gated fields**: convince-wrong probe text and full transcripts are in the gated dataset. The transcript-free per-debate Round 1 feature matrix (`abc_exp/results/per_debate_r1_features.jsonl`) is open.
 
@@ -173,4 +193,4 @@ The release has three parts:
 - Open data tier (https://huggingface.co/datasets/XINLI1997/DebateLedger), CC BY 4.0: the aggregate files plus every probe and debate trace with the model-generated text fields set to `null`. Parsed answers, correctness labels, round structure, token counts and costs are kept. The non-social weak, moderate and strong probe templates are open.
 - Gated data tier (https://huggingface.co/datasets/XINLI1997/DebateLedger-gated), research-use license with no redistribution and no fine-tuning of models that target real users without independent safety review: the full-text versions of the traces that contain model text, and the very-strong (convince-wrong) probe and social-pressure suffix. These can be reused as a misleading-answer corpus, so access is reviewed on request.
 
-The first public release is versioned `alpha-tot-v1.0`; the current one is `alpha-tot-v1.0.3`. Closed-API snapshot reruns and corrections are released as new minor versions rather than in-place replacements.
+The first public release is versioned `alpha-tot-v1.0`; the current one is `alpha-tot-v1.0.4`. Closed-API snapshot reruns and corrections are released as new minor versions rather than in-place replacements.

abc_exp/results/CONDITIONAL_COLLAPSE_TABLE.md

@@ -7,7 +7,7 @@ N models = 11
 | Claude Sonnet 4.5 | Anthropic | 120 | 77.5% | 0.705 | +0.15 | 1.67 [0.46, 5.87] | 2.15 [0.59, 7.51] | 11.11 | 1.50 |
 | Claude Haiku 4.5 | Anthropic | 500 | 82.2% | 0.512 | -0.03 | 9.20 [6.97, 12.05] | 11.19 [8.50, 14.61] | 19.10 | 0.37 |
 | GPT-4o-mini | OpenAI | 498 | 68.5% | 0.458 | +0.05 | 1.61 [0.82, 3.14] | 2.35 [1.19, 4.56] | 10.19 | 2.00 |
-| Gemini 2.5 Flash | Google | 200 | 36.5% | 0.648 | +0.14 | 0.50 [0.09, 2.78] | 1.37 [0.24, 7.36] | 11.81 | 15.00 |
+| Gemini 2.5 Flash | Google | 118 | 61.9% | 0.648 | +0.14 | 0.85 [0.15, 4.64] | 1.37 [0.24, 7.36] | 33.33 | 15.00 |
 | GPT-5.4-mini | OpenAI | 200 | 71.5% | 0.350 | +0.05 | 0.50 [0.09, 2.78] | 0.70 [0.12, 3.85] | 17.54 | 10.00 |
 | Gemini 3-flash | Google | 200 | 85.0% | 0.249 | -0.08 | 0.50 [0.09, 2.78] | 0.59 [0.10, 3.26] | 6.67 | 2.00 |
 | Phi-4-mini | Microsoft | 2161 | 50.6% | 0.451 | +0.23 | 5.18 [4.33, 6.20] | 10.24 [8.58, 12.18] | 16.31 | 1.55 |
@@ -18,10 +18,10 @@ N models = 11
 
 ## Cross-model rank tests
 - **N_models**: 11
-- **spearman_FR_raw**: stat=0.087, p=0.7984
+- **spearman_FR_raw**: stat=0.151, p=0.6583
 - **spearman_FR_cond**: stat=0.196, p=0.5637
-- **spearman_FR_recover**: stat=-0.091, p=0.7899
+- **spearman_FR_recover**: stat=0.036, p=0.9153
 - **spearman_SA_cond**: stat=0.251, p=0.4574
 - **kendall_FR_cond**: stat=0.147, p=0.5322
-- **spearman_init_acc_cond**: stat=-0.218, p=0.5192
-- **spearman_init_acc_raw**: stat=0.037, p=0.9147
+- **spearman_init_acc_cond**: stat=-0.409, p=0.2115
+- **spearman_init_acc_raw**: stat=-0.232, p=0.4918

abc_exp/results/DEBATE_EVAL_ARTIFACT_DATASHEET.md

@@ -44,7 +44,7 @@ Answer extraction uses the parser documented in `README_REPRO.md` and Appendix C
 
 ## Maintenance And Versioning
 
-The initial release is versioned as `alpha-tot-v1.0`; the current one is `alpha-tot-v1.0.3`; since `alpha-tot-v1.0.2` every row is read at Round 3 with alphabetic ties. The archived rows are immutable. Future closed-API reruns, provider snapshot changes, parser changes, or benchmark-pool revisions should be released as new minor versions with a changelog rather than overwriting the archived artifact. Each public version is archived with a DOI and cross-links the code repository, the open and gated datasets, this datasheet, and the metadata files.
+The initial release is versioned as `alpha-tot-v1.0`; the current one is `alpha-tot-v1.0.4`; since `alpha-tot-v1.0.2` every row is read at Round 3 with alphabetic ties. The archived rows are immutable. Future closed-API reruns, provider snapshot changes, parser changes, or benchmark-pool revisions should be released as new minor versions with a changelog rather than overwriting the archived artifact. Each public version is archived with a DOI and cross-links the code repository, the open and gated datasets, this datasheet, and the metadata files.
 
 ## Known Limitations
 

abc_exp/results/DEBATE_EVAL_CROISSANT_VALIDATION.md

@@ -5,7 +5,7 @@
 - Metadata file: `abc_exp/results/debate_eval_croissant_rai_metadata.json`.
 - Top-level Croissant file: `croissant.json`.
 - Dataset URL in metadata: `https://huggingface.co/datasets/XINLI1997/DebateLedger`.
-- Files hashed: **21**.
+- Files hashed: **23**.
 
 The metadata URL is set to the public Hugging Face dataset. If the URL changes, rerun the validator with `DEBATELEDGER_DATASET_URL` set to the new URL.
 
@@ -30,13 +30,13 @@ The metadata URL is set to the public Hugging Face dataset. If the URL changes,
 
 | File | SHA256 prefix |
 |---|---:|
-| `README_REPRO.md` | `c6e9a51e08df` |
+| `README_REPRO.md` | `5e0b142177ed` |
 | `requirements-artifact.txt` | `a092c219f714` |
-| `abc_exp/results/DEBATE_EVAL_ARTIFACT_DATASHEET.md` | `a5aa8fc5f835` |
+| `abc_exp/results/DEBATE_EVAL_ARTIFACT_DATASHEET.md` | `e4ab47578fca` |
 | `abc_exp/results/debate_eval_reuse_card.schema.json` | `4d6c028c58cd` |
 | `abc_exp/results/cohort_n14_headline.json` | `9727061b9b7c` |
 | `abc_exp/results/family_level_headline_aggregation.json` | `958e2ebbe781` |
-| `abc_exp/results/conditional_collapse_table.json` | `2df5c2ae2f80` |
+| `abc_exp/results/conditional_collapse_table.json` | `d6a11d548753` |
 | `abc_exp/results/reviewer_hardening_analyses.json` | `2a7da77d2ce2` |
 | `abc_exp/results/N14_REVIEWER_HARDENING_ARTIFACTS.md` | `797a36b82322` |
 | `abc_exp/results/n14_reviewer_hardening_artifacts.json` | `167509bdabc0` |
@@ -51,3 +51,5 @@ The metadata URL is set to the public Hugging Face dataset. If the URL changes,
 | `abc_exp/results/block0_debates.jsonl` | `c3c05b9b3ca3` |
 | `abc_exp/results/block1_abc_vs_baselines.jsonl` | `7cba65bb6ab8` |
 | `abc_exp/results/block3_multi_benchmark.jsonl` | `72a7397af3a6` |
+| `abc_exp/results/gate_rerun/gate_lomo_paper.json` | `1350c4e952a1` |
+| `abc_exp/results/channel_panel/channel_panel_22.json` | `5320543ee013` |

abc_exp/results/POST_A6_R1_REPLICATION.md

@@ -6,16 +6,16 @@ Majority-level metrics from `debate_traces_*.jsonl`. Collapse onset uses natural
 
 | Cohort | N | Init Maj Correct | Collapses | Cond Collapse | Corrections | Final Acc | Collapse Onset |
 |---|---:|---:|---:|---:|---:|---:|---|
-| Legacy/partial traces | 82 | 63 | 2 | 3.2% | 5 | 80.5% | R4: 1, unlocalized: 1 |
+| Legacy/partial traces | 85 | 65 | 2 | 3.1% | 5 | 80.0% | R4: 1, unlocalized: 1 |
 | Post-A6 200-row traces | 1200 | 662 | 75 | 11.3% | 144 | 60.9% | R1: 40, R2: 10, R3: 11, R4: 14 |
 
 ## Pooled All Traces
 
-- N debates: 1282
-- Initial majority correct: 725 (56.6%)
+- N debates: 1285
+- Initial majority correct: 727 (56.6%)
 - Collapses: 77 (10.6% conditional on initial majority correct)
-- Corrections: 149 (26.8% of initial-majority-wrong debates)
-- Final accuracy: 797 / 1282 (62.2%)
+- Corrections: 149 (26.7% of initial-majority-wrong debates)
+- Final accuracy: 799 / 1285 (62.2%)
 - Collapse onset: R1: 40, R2: 10, R3: 11, R4: 15, unlocalized: 1
 
 ## Per Model
@@ -28,7 +28,7 @@ Majority-level metrics from `debate_traces_*.jsonl`. Collapse onset uses natural
 | Qwen/Qwen3.6-27B-FP8 | post_a6_200row | 200 | 122 | 10 | 8.2% | 36 | 74.0% | R1: 1, R2: 2, R3: 4, R4: 3 |
 | Qwen/Qwen3.6-35B-A3B-FP8 | post_a6_200row | 200 | 93 | 10 | 10.8% | 47 | 65.0% | R2: 3, R3: 2, R4: 5 |
 | deepseek-v4-flash | post_a6_200row | 200 | 149 | 0 | 0.0% | 7 | 78.0% | - |
-| deepseek-v4-pro | legacy_or_partial | 78 | 59 | 1 | 1.7% | 5 | 80.8% | R4: 1 |
+| deepseek-v4-pro | legacy_or_partial | 81 | 61 | 1 | 1.6% | 5 | 80.2% | R4: 1 |
 | google/gemma-4-31b-it | post_a6_200row | 200 | 159 | 0 | 0.0% | 5 | 82.0% | - |
 | kimi-k2.6 | legacy_or_partial | 3 | 3 | 1 | 33.3% | 0 | 66.7% | - |
 
@@ -37,7 +37,7 @@ Majority-level metrics from `debate_traces_*.jsonl`. Collapse onset uses natural
 | File | Rows | Cohort |
 |---|---:|---|
 | `debate_traces_openrouter_deepseek-v4-flash.jsonl` | 200 | post_a6_200row |
-| `debate_traces_openrouter_deepseek-v4-pro.jsonl` | 78 | legacy_or_partial |
+| `debate_traces_openrouter_deepseek-v4-pro.jsonl` | 81 | legacy_or_partial |
 | `debate_traces_openrouter_kimi-k2.6.jsonl` | 3 | legacy_or_partial |
 | `debate_traces_smoke_qwen3-8b.jsonl` | 1 | legacy_or_partial |
 | `debate_traces_vllm_gemma-4-31b-it.jsonl` | 200 | post_a6_200row |

abc_exp/results/cascade_r1_auc.json

@@ -325,78 +325,84 @@
     }
   },
   "bootstrap_delta_auc_clustered": {
-    "status": "infeasible_missing_reported_cohort_row_level_data",
+    "status": "computed",
     "seed": 20260424,
     "B": 1000,
     "oof_seed": 42,
-    "observed_delta_source": "cascade_r1_auc.json:bootstrap_delta_auc",
-    "input_audit": [
-      "Missing reported-cohort raw trace files: results/debate_traces_gemini_3-flash.jsonl, results/debate_traces_openai_gpt-5.4-mini.jsonl, results/debate_traces_vllm_llama-3.1-8b.jsonl, results/debate_traces_vllm_phi-4-mini.jsonl, results/debate_traces_vllm_qwen3-4b.jsonl, results/debate_traces_vllm_qwen3-8b.jsonl",
-      "results/per_debate_r1_features.jsonl exists but is empty"
-    ],
+    "source": "reported_raw_trace_files",
+    "source_notes": [],
+    "n_debates": 6925,
+    "n_collapsed": 253,
+    "pooled_oof_auc": {
+      "baseline": 0.6690917029222472,
+      "r1_trajectory": 0.768466649605217,
+      "full_trajectory": 0.845866686097762
+    },
     "question_clustered": {
-      "status": "infeasible_missing_reported_cohort_row_level_data",
+      "status": "computed",
+      "cluster_key": "question_id",
+      "n_clusters": 2161,
       "B": 1000,
-      "B_used": 0,
+      "B_used": 1000,
+      "B_skipped_degenerate": 0,
       "delta_r1_minus_baseline": {
         "obs": 0.09937494668296987,
-        "median": null,
-        "ci_2.5": null,
-        "ci_97.5": null
+        "median": 0.09936596945247778,
+        "ci_2.5": 0.08087539133302094,
+        "ci_97.5": 0.11856359787105843
       },
       "delta_full_minus_r1": {
         "obs": 0.07740003649254501,
-        "median": null,
-        "ci_2.5": null,
-        "ci_97.5": null
-      },
-      "caveat": "The checked-out cascade_r1_auc.json has aggregate AUCs and the old debate-index bootstrap, but the exact reported-cohort row-level traces/cache with question_id and model identifiers are absent or empty. Clustered CIs cannot be recovered from aggregate AUC summaries without changing the estimand.",
-      "cluster_key": "question_id",
-      "n_clusters": null
+        "median": 0.07732057764107919,
+        "ci_2.5": 0.06376175232599098,
+        "ci_97.5": 0.09218039115988531
+      }
     },
     "model_clustered": {
-      "status": "infeasible_missing_reported_cohort_row_level_data",
+      "status": "computed",
+      "cluster_key": "model",
+      "n_clusters": 6,
       "B": 1000,
-      "B_used": 0,
+      "B_used": 1000,
+      "B_skipped_degenerate": 0,
       "delta_r1_minus_baseline": {
         "obs": 0.09937494668296987,
-        "median": null,
-        "ci_2.5": null,
-        "ci_97.5": null
+        "median": 0.09790560951867966,
+        "ci_2.5": 0.07388329834214147,
+        "ci_97.5": 0.12058461896873063
       },
       "delta_full_minus_r1": {
         "obs": 0.07740003649254501,
-        "median": null,
-        "ci_2.5": null,
-        "ci_97.5": null
-      },
-      "caveat": "The checked-out cascade_r1_auc.json has aggregate AUCs and the old debate-index bootstrap, but the exact reported-cohort row-level traces/cache with question_id and model identifiers are absent or empty. Clustered CIs cannot be recovered from aggregate AUC summaries without changing the estimand.",
-      "cluster_key": "model",
-      "n_clusters": null
+        "median": 0.08070226153778348,
+        "ci_2.5": 0.06859480199811763,
+        "ci_97.5": 0.09896733590391349
+      }
     },
     "two_way_like_clustered": {
-      "status": "infeasible_missing_reported_cohort_row_level_data",
+      "status": "computed",
+      "method": "crossed_cluster_resample_with_question_model_multiplicity_product",
+      "cluster_keys": [
+        "question_id",
+        "model"
+      ],
+      "n_question_clusters": 2161,
+      "n_model_clusters": 6,
       "B": 1000,
-      "B_used": 0,
+      "B_used": 998,
+      "B_skipped_degenerate": 2,
       "delta_r1_minus_baseline": {
         "obs": 0.09937494668296987,
-        "median": null,
-        "ci_2.5": null,
-        "ci_97.5": null
+        "median": 0.09743493774327006,
+        "ci_2.5": 0.05463957569635729,
+        "ci_97.5": 0.13500663672350027
       },
       "delta_full_minus_r1": {
         "obs": 0.07740003649254501,
-        "median": null,
-        "ci_2.5": null,
-        "ci_97.5": null
+        "median": 0.08038112605969788,
+        "ci_2.5": 0.05703069600172932,
+        "ci_97.5": 0.11124668033979467
       },
-      "caveat": "The checked-out cascade_r1_auc.json has aggregate AUCs and the old debate-index bootstrap, but the exact reported-cohort row-level traces/cache with question_id and model identifiers are absent or empty. Clustered CIs cannot be recovered from aggregate AUC summaries without changing the estimand.",
-      "cluster_keys": [
-        "question_id",
-        "model"
-      ],
-      "n_question_clusters": null,
-      "n_model_clusters": null
+      "caveat": "Two-way-like percentile bootstrap resamples question and model clusters independently using fixed out-of-fold predictions; it is a design-aware sensitivity, not a full refit exact two-way clustered bootstrap."
     }
   }
 }

abc_exp/results/channel_panel/channel_panel_22.json

@@ -0,0 +1,698 @@
+{
+ "summary": {
+  "models": 22,
+  "AA_gt_AS": 22,
+  "holm_sig": 20,
+  "strict_order": 18,
+  "mean_d": 0.7581603463423218,
+  "family_mean_d": {
+   "Google": 0.6336727557237534,
+   "DeepSeek": 0.33550202258285333,
+   "Zhipu": 1.5264401201263489,
+   "xAI": 0.49747040928574393,
+   "Tencent": 1.037212115456171,
+   "Mistral": 2.549272831429903,
+   "Meta": 0.9168727314675933,
+   "Microsoft": 0.6596764763471628,
+   "Qwen": 0.5051813818506653
+  },
+  "group_mean_d": {
+   "API models": 0.8791145741171492,
+   "Local open-weight models (vLLM)": 0.6130152730125289
+  },
+  "group_counts": {
+   "API models": 12,
+   "Local open-weight models (vLLM)": 10
+  },
+  "shared_questions": 188,
+  "sa_positive_per_model_pool": 21,
+  "sa_positive_shared": 21,
+  "sa_ci_excludes_zero_pos": 13,
+  "spearman_rho": 0.9909655561829477,
+  "spearman_p": 6.29840888794503e-19,
+  "sign_flips": [],
+  "rerun_dir": "abc_exp/results/gate_rerun"
+ },
+ "per_model": {
+  "Gemini 3-flash": {
+   "family": "Google",
+   "group": "API models",
+   "file": "rel:sa_causal_gemini_3_flash_n200.jsonl",
+   "N": 1800,
+   "pairs": 600,
+   "FR_D": 0.14020833333333332,
+   "FR_AS": 0.03208333333333333,
+   "FR_AA": 0.384375,
+   "order": "check",
+   "d": 1.360310278049752,
+   "p": 1.656367866361095e-138,
+   "sa_default": 0.15791640208381125,
+   "sa_ci": [
+    0.12916644745873265,
+    0.1870830196047337
+   ],
+   "sa_questions": 200,
+   "p_holm": 3.1470989460860803e-137,
+   "holm_sig": true,
+   "shared": {
+    "records": 564,
+    "questions": 188,
+    "alpha": 0.12078900709219857,
+    "S_sens": 0.10150709219858157,
+    "A_sens": 0.08732269503546099,
+    "SA": 0.1533685372345071,
+    "InitAcc": 0.8581560283687943
+   }
+  },
+  "Gemini 3.1-flash-lite": {
+   "family": "Google",
+   "group": "API models",
+   "file": "rel:sa_causal_gemini_3_1_flash_lite_n200.jsonl",
+   "N": 1800,
+   "pairs": 600,
+   "FR_D": 0.30645833333333333,
+   "FR_AS": 0.27458333333333335,
+   "FR_AA": 0.32416666666666666,
+   "order": "check",
+   "d": 0.19302509010521768,
+   "p": 2.8279431675340454e-06,
+   "sa_default": 0.028749936250133747,
+   "sa_ci": [
+    -0.02124995914591339,
+    0.07624984306281832
+   ],
+   "sa_questions": 200,
+   "p_holm": 1.1311772670136182e-05,
+   "holm_sig": true,
+   "shared": {
+    "records": 564,
+    "questions": 188,
+    "alpha": 0.2910017730496454,
+    "S_sens": 0.01374113475177305,
+    "A_sens": 0.11835106382978723,
+    "SA": 0.03280134751787765,
+    "InitAcc": 0.8492907801418439
+   }
+  },
+  "Gemini 3.1-pro": {
+   "family": "Google",
+   "group": "API models",
+   "file": "rel:sa_causal_gemini_3_1_pro_n200.jsonl",
+   "N": 1791,
+   "pairs": 597,
+   "FR_D": 0.15117252931323283,
+   "FR_AS": 0.13295644891122277,
+   "FR_AA": 0.31972361809045224,
+   "order": "check",
+   "d": 0.6342806868514714,
+   "p": 9.146279706587667e-46,
+   "sa_default": -0.01340030318264824,
+   "sa_ci": [
+    -0.033919523890439215,
+    0.007118926737845415
+   ],
+   "sa_questions": 199,
+   "p_holm": 1.09755356479052e-44,
+   "holm_sig": true,
+   "shared": {
+    "records": 564,
+    "questions": 188,
+    "alpha": 0.13054078014184395,
+    "S_sens": -0.00044326241134751853,
+    "A_sens": 0.0022163120567375884,
+    "SA": -0.007978700354663118,
+    "InitAcc": 0.8652482269503545
+   }
+  },
+  "DeepSeek V4 Flash": {
+   "family": "DeepSeek",
+   "group": "API models",
+   "file": "rel:sa_causal_router_deepseek-v4-flash.jsonl",
+   "N": 1793,
+   "pairs": 598,
+   "FR_D": 0.11723993288590603,
+   "FR_AS": 0.03574414715719063,
+   "FR_AA": 0.18468280467445744,
+   "order": "check",
+   "d": 0.6562866443684916,
+   "p": 1.8829024514262912e-48,
+   "sa_default": 0.061666561250192924,
+   "sa_ci": [
+    0.037499943333428334,
+    0.08666651708360874
+   ],
+   "sa_questions": 200,
+   "p_holm": 2.447773186854179e-47,
+   "holm_sig": true,
+   "shared": {
+    "records": 563,
+    "questions": 188,
+    "alpha": 0.10992907801418439,
+    "S_sens": 0.035460992907801414,
+    "A_sens": 0.09397163120567377,
+    "SA": 0.05629422828032314,
+    "InitAcc": 0.8492907801418439
+   }
+  },
+  "DeepSeek V4 Pro": {
+   "family": "DeepSeek",
+   "group": "API models",
+   "file": "rel:sa_causal_router_deepseek-v4-pro.jsonl",
+   "N": 1614,
+   "pairs": 523,
+   "FR_D": 0.03997227356746765,
+   "FR_AS": 0.04460966542750929,
+   "FR_AA": 0.046261682242990654,
+   "order": "sim",
+   "d": 0.014717400797215022,
+   "p": 0.7365725067900777,
+   "sa_default": 0.01693401559720633,
+   "sa_ci": [
+    -0.0031194206773823502,
+    0.03698745374344714
+   ],
+   "sa_questions": 187,
+   "p_holm": 0.7365725067900777,
+   "holm_sig": false,
+   "shared": {
+    "records": 541,
+    "questions": 187,
+    "alpha": 0.038658645276292335,
+    "S_sens": 0.010026737967914439,
+    "A_sens": -0.0015597147950089107,
+    "SA": 0.01693401559720633,
+    "InitAcc": 0.8716577540106952
+   }
+  },
+  "GLM-4.6 (no reasoning)": {
+   "family": "Zhipu",
+   "group": "API models",
+   "file": "rel:sa_causal_openrouter_glm_4_6_noreason_n200.jsonl",
+   "N": 1800,
+   "pairs": 600,
+   "FR_D": 0.54125,
+   "FR_AS": 0.2864583333333333,
+   "FR_AA": 0.716875,
+   "order": "check",
+   "d": 1.5264401201263489,
+   "p": 9.236647647368552e-159,
+   "sa_default": 0.02208330708336792,
+   "sa_ci": [
+    -0.009999962583425578,
+    0.0529165963334385
+   ],
+   "sa_questions": 200,
+   "p_holm": 1.8473295294737104e-157,
+   "holm_sig": true,
+   "shared": {
+    "records": 564,
+    "questions": 188,
+    "alpha": 0.5283687943262411,
+    "S_sens": 0.018617021276595747,
+    "A_sens": 0.45035460992907794,
+    "SA": 0.022606352393663567,
+    "InitAcc": 0.6843971631205673
+   }
+  },
+  "Grok 4.1 Fast": {
+   "family": "xAI",
+   "group": "API models",
+   "file": "rel:sa_causal_openrouter_grok_4_1_fast_n200.jsonl",
+   "N": 1799,
+   "pairs": 599,
+   "FR_D": 0.18291666666666667,
+   "FR_AS": 0.06833333333333333,
+   "FR_AA": 0.18447412353923207,
+   "order": "check",
+   "d": 0.49747040928574393,
+   "p": 1.2726557462531835e-30,
+   "sa_default": 0.003333333333326666,
+   "sa_ci": [
+    -0.0266666141876041,
+    0.03291660958343875
+   ],
+   "sa_questions": 200,
+   "p_holm": 1.0181245970025468e-29,
+   "holm_sig": true,
+   "shared": {
+    "records": 564,
+    "questions": 188,
+    "alpha": 0.16843971631205673,
+    "S_sens": 0.0044326241134751785,
+    "A_sens": 0.11702127659574468,
+    "SA": 0.0008865310283484103,
+    "InitAcc": 0.7163120567375888
+   }
+  },
+  "HY3 Preview": {
+   "family": "Tencent",
+   "group": "API models",
+   "file": "rel:sa_causal_openrouter_hy3_preview_free_n200.jsonl",
+   "N": 1785,
+   "pairs": 590,
+   "FR_D": 0.3123949579831933,
+   "FR_AS": 0.08340336134453781,
+   "FR_AA": 0.4016806722689076,
+   "order": "check",
+   "d": 1.037212115456171,
+   "p": 1.3626786726276777e-95,
+   "sa_default": 0.024374973125031882,
+   "sa_ci": [
+    -0.009166628354263266,
+    0.05770824481265015
+   ],
+   "sa_questions": 200,
+   "p_holm": 2.316553743467052e-94,
+   "holm_sig": true,
+   "shared": {
+    "records": 559,
+    "questions": 188,
+    "alpha": 0.28734485815602834,
+    "S_sens": 0.021941489361702128,
+    "A_sens": 0.18107269503546097,
+    "SA": 0.02194146742023471,
+    "InitAcc": 0.801418439716312
+   }
+  },
+  "Mistral Small 4": {
+   "family": "Mistral",
+   "group": "API models",
+   "file": "rel:sa_causal_openrouter_mistral_small_4_n200.jsonl",
+   "N": 1800,
+   "pairs": 600,
+   "FR_D": 0.420625,
+   "FR_AS": 0.06604166666666667,
+   "FR_AA": 0.6447916666666667,
+   "order": "check",
+   "d": 2.549272831429903,
+   "p": 1.9770419770592286e-264,
+   "sa_default": 0.03791659708346625,
+   "sa_ci": [
+    9.145805896781952e-09,
+    0.0754269547398181
+   ],
+   "sa_questions": 200,
+   "p_holm": 4.349492349530303e-263,
+   "holm_sig": true,
+   "shared": {
+    "records": 564,
+    "questions": 188,
+    "alpha": 0.42220744680851063,
+    "S_sens": 0.023492907801418435,
+    "A_sens": 0.21675531914893617,
+    "SA": 0.0407800664895062,
+    "InitAcc": 0.6648936170212765
+   }
+  },
+  "Llama 3.3 70B": {
+   "family": "Meta",
+   "group": "API models",
+   "file": "rel:sa_causal_openrouter_llama_3_3_70b_instruct_n200.jsonl",
+   "N": 1797,
+   "pairs": 598,
+   "FR_D": 0.47036727879799667,
+   "FR_AS": 0.19145833333333334,
+   "FR_AA": 0.5796404682274248,
+   "order": "check",
+   "d": 1.3064661232671453,
+   "p": 2.430831809605083e-131,
+   "sa_default": 0.07458322000019083,
+   "sa_ci": [
+    0.04374994537507637,
+    0.10541649379197446
+   ],
+   "sa_questions": 200,
+   "p_holm": 4.375497257289149e-130,
+   "holm_sig": true,
+   "shared": {
+    "records": 564,
+    "questions": 188,
+    "alpha": 0.4532358156028369,
+    "S_sens": 0.05363475177304964,
+    "A_sens": 0.2646276595744681,
+    "SA": 0.07180840248245034,
+    "InitAcc": 0.6187943262411348
+   }
+  },
+  "Llama 4 Scout": {
+   "family": "Meta",
+   "group": "API models",
+   "file": "rel:sa_causal_openrouter_llama_4_scout_n200.jsonl",
+   "N": 1800,
+   "pairs": 600,
+   "FR_D": 0.43416666666666665,
+   "FR_AS": 0.425625,
+   "FR_AA": 0.5741666666666667,
+   "order": "check",
+   "d": 0.5400995102531039,
+   "p": 3.095527782959672e-35,
+   "sa_default": 0.03708326375013458,
+   "sa_ci": [
+    0.0029166569166900872,
+    0.07249985277113002
+   ],
+   "sa_questions": 200,
+   "p_holm": 2.7859750046637047e-34,
+   "holm_sig": true,
+   "shared": {
+    "records": 564,
+    "questions": 188,
+    "alpha": 0.4273049645390071,
+    "S_sens": 0.024822695035460994,
+    "A_sens": 0.16489361702127658,
+    "SA": 0.042996374556889635,
+    "InitAcc": 0.7074468085106383
+   }
+  },
+  "Llama 4 Maverick": {
+   "family": "Meta",
+   "group": "API models",
+   "file": "rel:sa_causal_openrouter_llama_4_maverick_n200.jsonl",
+   "N": 1794,
+   "pairs": 598,
+   "FR_D": 0.5727768456375839,
+   "FR_AS": 0.6220784641068448,
+   "FR_AA": 0.6798831385642737,
+   "order": "sim",
+   "d": 0.23379367941522572,
+   "p": 1.7120815256254766e-08,
+   "sa_default": 0.07499988583352585,
+   "sa_ci": [
+    0.03312496428129091,
+    0.11708836267217773
+   ],
+   "sa_questions": 200,
+   "p_holm": 8.560407628127382e-08,
+   "holm_sig": true,
+   "shared": {
+    "records": 560,
+    "questions": 188,
+    "alpha": 0.5675975177304964,
+    "S_sens": 0.05851063829787234,
+    "A_sens": 0.17154255319148937,
+    "SA": 0.08156015602858864,
+    "InitAcc": 0.772163120567376
+   }
+  },
+  "Gemma 4 31B": {
+   "family": "Google",
+   "group": "Local open-weight models (vLLM)",
+   "file": "rel:sa_causal_vllm_gemma-4-31b-it-awq.jsonl",
+   "N": 1800,
+   "pairs": 600,
+   "FR_D": 0.296875,
+   "FR_AS": 0.265,
+   "FR_AA": 0.37,
+   "order": "check",
+   "d": 0.3470749678885722,
+   "p": 1.4911765708273656e-16,
+   "sa_default": 0.016249970416722916,
+   "sa_ci": [
+    -0.021666632520892435,
+    0.0541665716876766
+   ],
+   "sa_questions": 200,
+   "p_holm": 8.947059424964194e-16,
+   "holm_sig": true,
+   "shared": {
+    "records": 564,
+    "questions": 188,
+    "alpha": 0.28812056737588654,
+    "S_sens": 0.0035460992907801418,
+    "A_sens": 0.07624113475177305,
+    "SA": 0.008421967641881653,
+    "InitAcc": 0.824468085106383
+   }
+  },
+  "Llama-3.1-8B": {
+   "family": "Meta",
+   "group": "Local open-weight models (vLLM)",
+   "file": "rerun:sa_causal_vllm_llama-3.1-8b_rerun.jsonl",
+   "N": 1800,
+   "pairs": 600,
+   "FR_D": 0.68875,
+   "FR_AS": 0.4014583333333333,
+   "FR_AA": 0.8122916666666666,
+   "order": "check",
+   "d": 1.5871316129348985,
+   "p": 5.995486703811289e-166,
+   "sa_default": 0.05124989458354708,
+   "sa_ci": [
+    0.005416661895836811,
+    0.09833315250034584
+   ],
+   "sa_questions": 200,
+   "p_holm": 1.2590522078003708e-164,
+   "holm_sig": true,
+   "shared": {
+    "records": 564,
+    "questions": 188,
+    "alpha": 0.6808510638297872,
+    "S_sens": 0.018617021276595747,
+    "A_sens": 0.29609929078014185,
+    "SA": 0.046099189716523055,
+    "InitAcc": 0.3404255319148936
+   }
+  },
+  "Phi-4-mini": {
+   "family": "Microsoft",
+   "group": "Local open-weight models (vLLM)",
+   "file": "rerun:sa_causal_vllm_phi-4-mini_rerun.jsonl",
+   "N": 1800,
+   "pairs": 600,
+   "FR_D": 0.45458333333333334,
+   "FR_AS": 0.29583333333333334,
+   "FR_AA": 0.49666666666666665,
+   "order": "check",
+   "d": 0.6596764763471628,
+   "p": 5.14773141087264e-49,
+   "sa_default": 0.05958326125009708,
+   "sa_ci": [
+    0.015416668708296374,
+    0.10458318885439095
+   ],
+   "sa_questions": 200,
+   "p_holm": 7.206823975221696e-48,
+   "holm_sig": true,
+   "shared": {
+    "records": 564,
+    "questions": 188,
+    "alpha": 0.45257092198581567,
+    "S_sens": 0.05585106382978723,
+    "A_sens": 0.24468085106382978,
+    "SA": 0.06294318528379345,
+    "InitAcc": 0.3882978723404255
+   }
+  },
+  "Qwen3-4B": {
+   "family": "Qwen",
+   "group": "Local open-weight models (vLLM)",
+   "file": "rerun:sa_causal_vllm_qwen3-4b_rerun.jsonl",
+   "N": 1800,
+   "pairs": 600,
+   "FR_D": 0.65125,
+   "FR_AS": 0.62625,
+   "FR_AA": 0.794375,
+   "order": "check",
+   "d": 0.7975424200505937,
+   "p": 3.967118458034035e-66,
+   "sa_default": 0.2441662858339875,
+   "sa_ci": [
+    0.2037496945630088,
+    0.28416620416748584
+   ],
+   "sa_questions": 200,
+   "p_holm": 5.950677687051052e-65,
+   "holm_sig": true,
+   "shared": {
+    "records": 564,
+    "questions": 188,
+    "alpha": 0.6376329787234043,
+    "S_sens": 0.18129432624113476,
+    "A_sens": 0.3896276595744681,
+    "SA": 0.2531024401602584,
+    "InitAcc": 0.5212765957446808
+   }
+  },
+  "Qwen3-8B": {
+   "family": "Qwen",
+   "group": "Local open-weight models (vLLM)",
+   "file": "rerun:sa_causal_vllm_qwen3-8b_rerun.jsonl",
+   "N": 1800,
+   "pairs": 600,
+   "FR_D": 0.34208333333333335,
+   "FR_AS": 0.18625,
+   "FR_AA": 0.4845833333333333,
+   "order": "check",
+   "d": 0.9131072702927213,
+   "p": 5.164393210373414e-81,
+   "sa_default": 0.10458317125027708,
+   "sa_ci": [
+    0.060416577833479,
+    0.14666643008374977
+   ],
+   "sa_questions": 200,
+   "p_holm": 8.263029136597463e-80,
+   "holm_sig": true,
+   "shared": {
+    "records": 564,
+    "questions": 188,
+    "alpha": 0.3249113475177305,
+    "S_sens": 0.07535460992907801,
+    "A_sens": 0.1897163120567376,
+    "SA": 0.10372324379460017,
+    "InitAcc": 0.597517730496454
+   }
+  },
+  "Qwen3-32B": {
+   "family": "Qwen",
+   "group": "Local open-weight models (vLLM)",
+   "file": "rel:sa_causal_vllm_qwen3_32b_a7_n200.jsonl",
+   "N": 1800,
+   "pairs": 600,
+   "FR_D": 0.6545833333333333,
+   "FR_AS": 0.6754166666666667,
+   "FR_AA": 0.7002083333333333,
+   "order": "sim",
+   "d": 0.08093099086681814,
+   "p": 0.04789138765468257,
+   "sa_default": 0.006666660000006665,
+   "sa_ci": [
+    -0.038749929604300355,
+    0.052916576333498506
+   ],
+   "sa_questions": 200,
+   "p_holm": 0.09578277530936515,
+   "holm_sig": false,
+   "shared": {
+    "records": 564,
+    "questions": 188,
+    "alpha": 0.6527039007092199,
+    "S_sens": 0.007535460992907801,
+    "A_sens": 0.22384751773049644,
+    "SA": 0.00930849778370789,
+    "InitAcc": 0.35992907801418433
+   }
+  },
+  "Qwen3.5-4B": {
+   "family": "Qwen",
+   "group": "Local open-weight models (vLLM)",
+   "file": "off:sa_causal_vllm_qwen3.5-4b_thinkoff.jsonl",
+   "N": 1800,
+   "pairs": 600,
+   "FR_D": 0.360625,
+   "FR_AS": 0.3304166666666667,
+   "FR_AA": 0.5010416666666667,
+   "order": "check",
+   "d": 0.6285384355615422,
+   "p": 2.617234244550622e-45,
+   "sa_default": 0.027916620416749582,
+   "sa_ci": [
+    -0.004166653354198271,
+    0.05999989252103576
+   ],
+   "sa_questions": 200,
+   "p_holm": 2.8789576690056844e-44,
+   "holm_sig": true,
+   "shared": {
+    "records": 564,
+    "questions": 188,
+    "alpha": 0.3426418439716312,
+    "S_sens": 0.01773049645390071,
+    "A_sens": 0.11524822695035461,
+    "SA": 0.023049611702182626,
+    "InitAcc": 0.7092198581560283
+   }
+  },
+  "Qwen3.5-9B": {
+   "family": "Qwen",
+   "group": "Local open-weight models (vLLM)",
+   "file": "off:sa_causal_vllm_qwen3.5-9b_thinkoff.jsonl",
+   "N": 1800,
+   "pairs": 600,
+   "FR_D": 0.29791666666666666,
+   "FR_AS": 0.18145833333333333,
+   "FR_AA": 0.336875,
+   "order": "check",
+   "d": 0.5588911263397918,
+   "p": 2.592486041976998e-37,
+   "sa_default": 0.04416659250013416,
+   "sa_ci": [
+    0.009999982479199228,
+    0.07874985708360457
+   ],
+   "sa_questions": 200,
+   "p_holm": 2.592486041976998e-36,
+   "holm_sig": true,
+   "shared": {
+    "records": 564,
+    "questions": 188,
+    "alpha": 0.27615248226950356,
+    "S_sens": 0.028368794326241138,
+    "A_sens": 0.10992907801418439,
+    "SA": 0.042996381649066925,
+    "InitAcc": 0.7517730496453902
+   }
+  },
+  "Qwen3.6-27B": {
+   "family": "Qwen",
+   "group": "Local open-weight models (vLLM)",
+   "file": "off:sa_causal_vllm_qwen3.6-27b-fp8_thinkoff.jsonl",
+   "N": 1800,
+   "pairs": 600,
+   "FR_D": 0.10291666666666667,
+   "FR_AS": 0.07895833333333334,
+   "FR_AA": 0.16625,
+   "order": "check",
+   "d": 0.4307436393733259,
+   "p": 5.4926247580465e-24,
+   "sa_default": 0.028749952083419586,
+   "sa_ci": [
+    0.009583323750009583,
+    0.047916583833486
+   ],
+   "sa_questions": 200,
+   "p_holm": 3.84483733063255e-23,
+   "holm_sig": true,
+   "shared": {
+    "records": 564,
+    "questions": 188,
+    "alpha": 0.08488475177304966,
+    "S_sens": 0.020833333333333336,
+    "A_sens": 0.05629432624113476,
+    "SA": 0.031028317375978717,
+    "InitAcc": 0.8351063829787234
+   }
+  },
+  "Qwen3.6-35B-A3B": {
+   "family": "Qwen",
+   "group": "Local open-weight models (vLLM)",
+   "file": "off:sa_causal_vllm_qwen3.6-35b-a3b-fp8_thinkoff.jsonl",
+   "N": 1800,
+   "pairs": 600,
+   "FR_D": 0.12479166666666666,
+   "FR_AS": 0.08875,
+   "FR_AA": 0.11458333333333333,
+   "order": "sim",
+   "d": 0.1265157904698635,
+   "p": 0.002032852363501612,
+   "sa_default": 0.06166656250018916,
+   "sa_ci": [
+    0.032083273708448046,
+    0.09291650793779035
+   ],
+   "sa_questions": 200,
+   "p_holm": 0.0060985570905048355,
+   "holm_sig": true,
+   "shared": {
+    "records": 564,
+    "questions": 188,
+    "alpha": 0.11148049645390071,
+    "S_sens": 0.03235815602836879,
+    "A_sens": 0.053634751773049646,
+    "SA": 0.0496453058512172,
+    "InitAcc": 0.8333333333333335
+   }
+  }
+ }
+}

abc_exp/results/channel_panel/channel_panel_22.tex

@@ -0,0 +1,75 @@
+% tab:full12 rows
+\multicolumn{8}{l}{\emph{API models (12 models, mean $d = 0.879$)}} \\
+Gemini 3-flash & 1,800 & .140 & .032 & .384 & $\checkmark$ & 1.360 & 1.7e-138$^\star$ \\
+Gemini 3.1-flash-lite & 1,800 & .306 & .275 & .324 & $\checkmark$ & 0.193 & 2.8e-6$^\star$ \\
+Gemini 3.1-pro & 1,791 & .151 & .133 & .320 & $\checkmark$ & 0.634 & 9.1e-46$^\star$ \\
+DeepSeek V4 Flash & 1,793 & .117 & .036 & .185 & $\checkmark$ & 0.656 & 1.9e-48$^\star$ \\
+DeepSeek V4 Pro & 1,614 & .040 & .045 & .046 & $\sim$ & 0.015 & .737 \\
+GLM-4.6 (no reasoning) & 1,800 & .541 & .286 & .717 & $\checkmark$ & 1.526 & 9.2e-159$^\star$ \\
+Grok 4.1 Fast & 1,799 & .183 & .068 & .184 & $\checkmark$ & 0.497 & 1.3e-30$^\star$ \\
+HY3 Preview & 1,785 & .312 & .083 & .402 & $\checkmark$ & 1.037 & 1.4e-95$^\star$ \\
+Mistral Small 4 & 1,800 & .421 & .066 & .645 & $\checkmark$ & 2.549 & 2.0e-264$^\star$ \\
+Llama 3.3 70B & 1,797 & .470 & .191 & .580 & $\checkmark$ & 1.306 & 2.4e-131$^\star$ \\
+Llama 4 Scout & 1,800 & .434 & .426 & .574 & $\checkmark$ & 0.540 & 3.1e-35$^\star$ \\
+Llama 4 Maverick & 1,794 & .573 & .622 & .680 & $\sim$ & 0.234 & 1.7e-8$^\star$ \\
+\midrule
+\multicolumn{8}{l}{\emph{Local open-weight models (vLLM) (10 models, mean $d = 0.613$)}} \\
+Gemma 4 31B & 1,800 & .297 & .265 & .370 & $\checkmark$ & 0.347 & 1.5e-16$^\star$ \\
+Llama-3.1-8B & 1,800 & .689 & .401 & .812 & $\checkmark$ & 1.587 & 6.0e-166$^\star$ \\
+Phi-4-mini & 1,800 & .455 & .296 & .497 & $\checkmark$ & 0.660 & 5.1e-49$^\star$ \\
+Qwen3-4B & 1,800 & .651 & .626 & .794 & $\checkmark$ & 0.798 & 4.0e-66$^\star$ \\
+Qwen3-8B & 1,800 & .342 & .186 & .485 & $\checkmark$ & 0.913 & 5.2e-81$^\star$ \\
+Qwen3-32B & 1,800 & .655 & .675 & .700 & $\sim$ & 0.081 & .048 \\
+Qwen3.5-4B & 1,800 & .361 & .330 & .501 & $\checkmark$ & 0.629 & 2.6e-45$^\star$ \\
+Qwen3.5-9B & 1,800 & .298 & .181 & .337 & $\checkmark$ & 0.559 & 2.6e-37$^\star$ \\
+Qwen3.6-27B & 1,800 & .103 & .079 & .166 & $\checkmark$ & 0.431 & 5.5e-24$^\star$ \\
+Qwen3.6-35B-A3B & 1,800 & .125 & .089 & .115 & $\sim$ & 0.127 & .002$^\star$ \\
+\midrule
+
+% tab:sa_ci rows
+Gemini 3-flash & $+0.158$ & [$+0.129$, $+0.187$] \\
+Gemini 3.1-flash-lite & $+0.029$ & [$-0.021$, $+0.076$] \\
+Gemini 3.1-pro & $-0.013$ & [$-0.034$, $+0.007$] \\
+DeepSeek V4 Flash & $+0.062$ & [$+0.037$, $+0.087$] \\
+DeepSeek V4 Pro & $+0.017$ & [$-0.003$, $+0.037$] \\
+GLM-4.6 (no reasoning) & $+0.022$ & [$-0.010$, $+0.053$] \\
+Grok 4.1 Fast & $+0.003$ & [$-0.027$, $+0.033$] \\
+HY3 Preview & $+0.024$ & [$-0.009$, $+0.058$] \\
+Mistral Small 4 & $+0.038$ & [$+0.000$, $+0.075$] \\
+Llama 3.3 70B & $+0.075$ & [$+0.044$, $+0.105$] \\
+Llama 4 Scout & $+0.037$ & [$+0.003$, $+0.072$] \\
+Llama 4 Maverick & $+0.075$ & [$+0.033$, $+0.117$] \\
+Gemma 4 31B & $+0.016$ & [$-0.022$, $+0.054$] \\
+Llama-3.1-8B & $+0.051$ & [$+0.005$, $+0.098$] \\
+Phi-4-mini & $+0.060$ & [$+0.015$, $+0.105$] \\
+Qwen3-4B & $+0.244$ & [$+0.204$, $+0.284$] \\
+Qwen3-8B & $+0.105$ & [$+0.060$, $+0.147$] \\
+Qwen3-32B & $+0.007$ & [$-0.039$, $+0.053$] \\
+Qwen3.5-4B & $+0.028$ & [$-0.004$, $+0.060$] \\
+Qwen3.5-9B & $+0.044$ & [$+0.010$, $+0.079$] \\
+Qwen3.6-27B & $+0.029$ & [$+0.010$, $+0.048$] \\
+Qwen3.6-35B-A3B & $+0.062$ & [$+0.032$, $+0.093$] \\
+
+% tab:shared_pool_sa rows
+Gemini 3-flash & 0.121 & $+0.102$ & $+0.087$ & $+0.153$ & 85.8\% \\
+Gemini 3.1-flash-lite & 0.291 & $+0.014$ & $+0.118$ & $+0.033$ & 84.9\% \\
+Gemini 3.1-pro & 0.131 & $-0.000$ & $+0.002$ & $-0.008$ & 86.5\% \\
+DeepSeek V4 Flash & 0.110 & $+0.035$ & $+0.094$ & $+0.056$ & 84.9\% \\
+DeepSeek V4 Pro & 0.039 & $+0.010$ & $-0.002$ & $+0.017$ & 87.2\% \\
+GLM-4.6 (no reasoning) & 0.528 & $+0.019$ & $+0.450$ & $+0.023$ & 68.4\% \\
+Grok 4.1 Fast & 0.168 & $+0.004$ & $+0.117$ & $+0.001$ & 71.6\% \\
+HY3 Preview & 0.287 & $+0.022$ & $+0.181$ & $+0.022$ & 80.1\% \\
+Mistral Small 4 & 0.422 & $+0.023$ & $+0.217$ & $+0.041$ & 66.5\% \\
+Llama 3.3 70B & 0.453 & $+0.054$ & $+0.265$ & $+0.072$ & 61.9\% \\
+Llama 4 Scout & 0.427 & $+0.025$ & $+0.165$ & $+0.043$ & 70.7\% \\
+Llama 4 Maverick & 0.568 & $+0.059$ & $+0.172$ & $+0.082$ & 77.2\% \\
+Gemma 4 31B & 0.288 & $+0.004$ & $+0.076$ & $+0.008$ & 82.4\% \\
+Llama-3.1-8B & 0.681 & $+0.019$ & $+0.296$ & $+0.046$ & 34.0\% \\
+Phi-4-mini & 0.453 & $+0.056$ & $+0.245$ & $+0.063$ & 38.8\% \\
+Qwen3-4B & 0.638 & $+0.181$ & $+0.390$ & $+0.253$ & 52.1\% \\
+Qwen3-8B & 0.325 & $+0.075$ & $+0.190$ & $+0.104$ & 59.8\% \\
+Qwen3-32B & 0.653 & $+0.008$ & $+0.224$ & $+0.009$ & 36.0\% \\
+Qwen3.5-4B & 0.343 & $+0.018$ & $+0.115$ & $+0.023$ & 70.9\% \\
+Qwen3.5-9B & 0.276 & $+0.028$ & $+0.110$ & $+0.043$ & 75.2\% \\
+Qwen3.6-27B & 0.085 & $+0.021$ & $+0.056$ & $+0.031$ & 83.5\% \\
+Qwen3.6-35B-A3B & 0.111 & $+0.032$ & $+0.054$ & $+0.050$ & 83.3\% \\

abc_exp/results/conditional_collapse_table.json

@@ -20,7 +20,8 @@
       "recover_hi": 28.057800818279823,
       "n_cond_coll": 2,
       "n_recover": 3,
-      "GB_ratio": 1.5
+      "GB_ratio": 1.5,
+      "source": "recorded_counts"
     },
     {
       "model": "Claude Haiku 4.5",
@@ -42,7 +43,8 @@
       "recover_hi": 28.4772682171176,
       "n_cond_coll": 46,
       "n_recover": 17,
-      "GB_ratio": 0.3695652173913043
+      "GB_ratio": 0.3695652173913043,
+      "source": "recorded_counts"
     },
     {
       "model": "GPT-4o-mini",
@@ -64,26 +66,28 @@
       "recover_hi": 15.912956144771895,
       "n_cond_coll": 8,
       "n_recover": 16,
-      "GB_ratio": 2.0
+      "GB_ratio": 2.0,
+      "source": "recorded_counts"
     },
     {
       "model": "Gemini 2.5 Flash",
       "family": "Google",
       "FR": 0.648,
       "SA": 0.14,
-      "n_deb": 200,
+      "source": "block1_abc_vs_baselines.jsonl",
+      "n_deb": 118,
       "n_init_correct": 73,
-      "n_init_wrong": 127,
-      "raw_collapse_pct": 0.5,
-      "raw_lo": 0.08831687156009763,
-      "raw_hi": 2.777370439789293,
+      "n_init_wrong": 45,
+      "raw_collapse_pct": 0.847457627118644,
+      "raw_lo": 0.14975405219652543,
+      "raw_hi": 4.6445572415402046,
       "cond_collapse_pct": 1.36986301369863,
       "cond_lo": 0.24222514664838657,
       "cond_hi": 7.359737270963818,
-      "init_acc": 0.365,
-      "recover_pct": 11.811023622047244,
-      "recover_lo": 7.289704388252161,
-      "recover_hi": 18.574772457019534,
+      "init_acc": 0.6186440677966102,
+      "recover_pct": 33.33333333333333,
+      "recover_lo": 21.358874934330032,
+      "recover_hi": 47.92951169009958,
       "n_cond_coll": 1,
       "n_recover": 15,
       "GB_ratio": 15.0
@@ -93,6 +97,7 @@
       "family": "OpenAI",
       "FR": 0.35,
       "SA": 0.05,
+      "source": "debate_traces_openai_gpt-5.4-mini.jsonl",
       "n_deb": 200,
       "n_init_correct": 143,
       "n_init_wrong": 57,
@@ -115,6 +120,7 @@
       "family": "Google",
       "FR": 0.249,
       "SA": -0.08,
+      "source": "debate_traces_gemini_3-flash.jsonl",
       "n_deb": 200,
       "n_init_correct": 170,
       "n_init_wrong": 30,
@@ -137,6 +143,7 @@
       "family": "Microsoft",
       "FR": 0.451,
       "SA": 0.23,
+      "source": "debate_traces_vllm_phi-4-mini.jsonl",
       "n_deb": 2161,
       "n_init_correct": 1094,
       "n_init_wrong": 1067,
@@ -159,6 +166,7 @@
       "family": "Qwen",
       "FR": 0.596,
       "SA": 0.08,
+      "source": "debate_traces_vllm_qwen3-4b.jsonl",
       "n_deb": 2161,
       "n_init_correct": 979,
       "n_init_wrong": 1182,
@@ -181,6 +189,7 @@
       "family": "Meta",
       "FR": 0.726,
       "SA": 0.06,
+      "source": "debate_traces_vllm_llama-3.1-8b.jsonl",
       "n_deb": 2003,
       "n_init_correct": 1016,
       "n_init_wrong": 987,
@@ -203,6 +212,7 @@
       "family": "Qwen",
       "FR": 0.596,
       "SA": -0.05,
+      "source": "debate_traces_vllm_qwen3-8b.jsonl",
       "n_deb": 200,
       "n_init_correct": 56,
       "n_init_wrong": 144,
@@ -240,29 +250,30 @@
       "recover_hi": 45.308339269675926,
       "n_cond_coll": 26,
       "n_recover": 97,
-      "GB_ratio": 3.730769230769231
+      "GB_ratio": 3.730769230769231,
+      "source": "recorded_counts"
     }
   ],
   "stats": {
     "N_models": 11,
     "spearman_FR_raw": {
-      "statistic": 0.08735839935247217,
-      "pvalue": 0.7984118639344213,
-      "correlation": 0.08735839935247217
+      "statistic": 0.15068493150684933,
+      "pvalue": 0.6583104144348021,
+      "correlation": 0.15068493150684933
     },
     "spearman_FR_cond": {
       "statistic": 0.19590028045862823,
-      "pvalue": 0.5637442329087623,
+      "pvalue": 0.5637442329087622,
       "correlation": 0.19590028045862823
     },
     "spearman_FR_recover": {
-      "statistic": -0.09111640951564104,
-      "pvalue": 0.7899039060352314,
-      "correlation": -0.09111640951564104
+      "statistic": 0.03644656380625642,
+      "pvalue": 0.9152761385497562,
+      "correlation": 0.03644656380625642
     },
     "spearman_SA_cond": {
       "statistic": 0.25057012616801283,
-      "pvalue": 0.4573852261533532,
+      "pvalue": 0.4573852261533529,
       "correlation": 0.25057012616801283
     },
     "kendall_FR_cond": {
@@ -271,14 +282,14 @@
       "correlation": 0.14679516868351475
     },
     "spearman_init_acc_cond": {
-      "statistic": -0.21818181818181823,
-      "pvalue": 0.519248247554982,
-      "correlation": -0.21818181818181823
+      "statistic": -0.4090909090909091,
+      "pvalue": 0.21154501034209602,
+      "correlation": -0.4090909090909091
     },
     "spearman_init_acc_raw": {
-      "statistic": 0.03669879217087869,
-      "pvalue": 0.9146916426486271,
-      "correlation": 0.03669879217087869
+      "statistic": -0.23234684426488467,
+      "pvalue": 0.4917654140586253,
+      "correlation": -0.23234684426488467
     }
   }
 }

abc_exp/results/construct_auc_public_n200.json

@@ -0,0 +1,87 @@
+{
+ "definition": "tab:construct on the public 200-question runs (zero API). The table is defined on public data; it does not reproduce\nan earlier version of the table, whose logs were not retained.\n\nUnit: one debate (model x question) whose initial majority is correct; outcome = collapse at Round 3 (parsed Round-3\nmajority wrong; a debate with no parsed Round-3 answer is not counted as a collapse).\nPer-question features from the same model's public probe file (same 200 questions), default condition:\n  mean probe FR = mean share of revised probes over the three agents; S/A = mean over agents of the paper's S/A formula\n  recomputed from probe_results (rec_scores);\n  difficulty = 1 - mean probe initial_correct over all rows of the question (all conditions and agents);\n  initial disagreement = 1 - (largest initial-answer count / number of parsed initial answers) in the debate.\nSingle predictors: AUC of the raw score (not re-oriented, as in the printed table). Combined: logistic regression,\nout-of-fold predictions with 5 question folds (pooled: same folds across models, plus model indicators).",
+ "per_model": {
+  "Gemini 3-flash": {
+   "n": 166,
+   "collapses": 1,
+   "fr": 0.7181818181818181,
+   "sa": 0.8212121212121212,
+   "dis": 0.9878787878787879
+  },
+  "Gemini 3.1-flash-lite": {
+   "n": 164,
+   "collapses": 0,
+   "fr": null,
+   "sa": null,
+   "dis": null
+  },
+  "Gemini 3.1-pro": {
+   "n": 145,
+   "collapses": 0,
+   "fr": null,
+   "sa": null,
+   "dis": null
+  },
+  "GLM-4.6 (no reasoning)": {
+   "n": 152,
+   "collapses": 6,
+   "fr": 0.533675799086758,
+   "sa": 0.3852739726027397,
+   "dis": 0.9178082191780822,
+   "base": 0.8698630136986302,
+   "full": 0.8242009132420092,
+   "delta": -0.045662100456621
+  },
+  "Grok 4.1 Fast": {
+   "n": 167,
+   "collapses": 0,
+   "fr": null,
+   "sa": null,
+   "dis": null
+  },
+  "Llama 3.3 70B": {
+   "n": 127,
+   "collapses": 2,
+   "fr": 0.672,
+   "sa": 0.38,
+   "dis": 0.88
+  },
+  "Llama 4 Maverick": {
+   "n": 158,
+   "collapses": 5,
+   "fr": 0.581045751633987,
+   "sa": 0.3836601307189543,
+   "dis": 0.7705882352941177,
+   "base": 0.550326797385621,
+   "full": 0.5908496732026144,
+   "delta": 0.04052287581699343
+  },
+  "Llama 4 Scout": {
+   "n": 146,
+   "collapses": 7,
+   "fr": 0.5801644398766701,
+   "sa": 0.6217882836587872,
+   "dis": 0.7944501541623844,
+   "base": 0.6659815005138746,
+   "full": 0.6135662898252826,
+   "delta": -0.05241521068859201
+  },
+  "Mistral Small 4": {
+   "n": 140,
+   "collapses": 4,
+   "fr": 0.7058823529411765,
+   "sa": 0.27849264705882354,
+   "dis": 0.9889705882352942
+  },
+  "Pooled": {
+   "n": 1365,
+   "collapses": 25,
+   "fr": 0.7069552238805971,
+   "sa": 0.446089552238806,
+   "dis": 0.8784776119402985,
+   "base": 0.8984776119402985,
+   "full": 0.8784776119402985,
+   "delta": -0.020000000000000018
+  }
+ }
+}

abc_exp/results/debate_eval_croissant_rai_metadata.json

@@ -50,7 +50,7 @@
   "name": "Debate Collapse/Correction Evaluation Artifact",
   "description": "Validated Croissant/RAI metadata for the collapse/correction evaluation protocol bundle used in a homogeneous-panel LLM debate audit.",
   "conformsTo": "http://mlcommons.org/croissant/1.0",
-  "citeAs": "Xin Li, Mengbing Liu, and Chau Yuen. Measuring Collapse and Correction in Homogeneous-Panel LLM Debate. NeurIPS 2026 Evaluations and Datasets Track. DebateLedger artifact version 1.0.3.",
+  "citeAs": "Xin Li, Mengbing Liu, and Chau Yuen. Measuring Collapse and Correction in Homogeneous-Panel LLM Debate. NeurIPS 2026 Evaluations and Datasets Track. DebateLedger artifact version 1.0.4.",
   "dateCreated": "2026-05-02",
   "datePublished": "2026-09-28",
   "inLanguage": "en",
@@ -71,14 +71,14 @@
   "rai:dataCollectionType": "machine-generated traces from public benchmark questions",
   "rai:dataLimitations": "Not a benchmark-general law; not a calibrated per-question oracle; not a deployable controller; gated tier needed for convince-wrong templates and full transcripts.",
   "rai:dataPreprocessingProtocol": "Rule-based MCQ answer extraction; majority voting after filtering unparsed answers; alphabetic tie-breaking; deterministic collapse/correction/onset labels.",
-  "rai:dataReleaseMaintenancePlan": "Versioned public release (alpha-tot-v1.0.3): open data on Hugging Face, code on GitHub, and a gated Hugging Face dataset for full transcripts and convince-wrong templates. Corrections and provider or parser changes are released as new minor versions.",
+  "rai:dataReleaseMaintenancePlan": "Versioned public release (alpha-tot-v1.0.4): open data on Hugging Face, code on GitHub, and a gated Hugging Face dataset for full transcripts and convince-wrong templates. Corrections and provider or parser changes are released as new minor versions.",
   "rai:dataSocialImpact": "Helps audit debate interventions that may reduce collapses while discarding corrections; gated release mitigates misuse as a misleading-answer corpus.",
   "rai:dataUseCases": "Zero-API recomputation of aggregate tables; applying the transition-table and signed-utility protocol to new model-scaffold pairs; documenting gated release boundaries; public artifact hosting on Hugging Face with code on GitHub.",
   "rai:hasSyntheticData": true,
   "rai:machineAnnotationTools": "Rule-based MCQ answer parser included in the artifact; no LLM judge step is used.",
   "rai:personalSensitiveInformation": "none",
   "url": "https://huggingface.co/datasets/XINLI1997/DebateLedger",
-  "version": "1.0.3",
+  "version": "1.0.4",
   "distribution": [
     {
       "@type": "cr:FileObject",
@@ -86,7 +86,7 @@
       "name": "README_REPRO.md",
       "contentUrl": "README_REPRO.md",
       "encodingFormat": "text/markdown",
-      "sha256": "c6e9a51e08dfd87c7b19f3ba63e40ceecabb3df06abacafc9907a523d5189c06"
+      "sha256": "5e0b142177ed0ca20479c161e0243d9f6560579c99404fef657951c21f74ca9a"
     },
     {
       "@type": "cr:FileObject",
@@ -102,7 +102,7 @@
       "name": "DEBATE_EVAL_ARTIFACT_DATASHEET.md",
       "contentUrl": "abc_exp/results/DEBATE_EVAL_ARTIFACT_DATASHEET.md",
       "encodingFormat": "text/markdown",
-      "sha256": "a5aa8fc5f8355743d836e609420f4a47f4e8d67c66c6a65caf231fa340848e96"
+      "sha256": "e4ab47578fca06c1904f680d78da011f43e96dcc58d0cb874a5cc734bb8fcbd3"
     },
     {
       "@type": "cr:FileObject",
@@ -134,7 +134,7 @@
       "name": "conditional_collapse_table.json",
       "contentUrl": "abc_exp/results/conditional_collapse_table.json",
       "encodingFormat": "application/json",
-      "sha256": "2df5c2ae2f8047c2d5706fe82339229774d4614c0583351042034f07ccaa5eba"
+      "sha256": "d6a11d548753d89b695d700e7f21755c414a5def11c2204acb7382192e9bbd21"
     },
     {
       "@type": "cr:FileObject",
@@ -248,6 +248,22 @@
       "encodingFormat": "application/jsonlines",
       "sha256": "72a7397af3a6469e64cc2412b58faca14c9dbd3366c8d3114b2d2093d82f9e5e"
     },
+    {
+      "@type": "cr:FileObject",
+      "@id": "file_abc_exp_results_gate_rerun_gate_lomo_paper_json",
+      "name": "gate_lomo_paper.json",
+      "contentUrl": "abc_exp/results/gate_rerun/gate_lomo_paper.json",
+      "encodingFormat": "application/json",
+      "sha256": "1350c4e952a1e03995a9eb65f422b02fa4174af35a371d45e7d2fcef746c18f9"
+    },
+    {
+      "@type": "cr:FileObject",
+      "@id": "file_abc_exp_results_channel_panel_channel_panel_22_json",
+      "name": "channel_panel_22.json",
+      "contentUrl": "abc_exp/results/channel_panel/channel_panel_22.json",
+      "encodingFormat": "application/json",
+      "sha256": "5320543ee013297441def772f7ac62545cf406f81c55d699a8908e86801d1027"
+    },
     {
       "@type": "cr:FileSet",
       "@id": "fileset_debate_traces",
@@ -303,6 +319,14 @@
       "description": "8-probe screen traces of the thinking-off rerun of the four Qwen3.5/3.6 rows. Model output text is null in the open tier.",
       "encodingFormat": "application/jsonlines",
       "includes": "abc_exp/results/thinking_off_rerun/sa_causal_*.jsonl"
+    },
+    {
+      "@type": "cr:FileSet",
+      "@id": "fileset_gate_rerun_probe_traces",
+      "name": "fileset_gate_rerun_probe_traces",
+      "description": "8-probe screen traces of the four local models collected for the probe-gated freeze and the channel panel. Model output text is null in the open tier.",
+      "encodingFormat": "application/jsonlines",
+      "includes": "abc_exp/results/gate_rerun/*.jsonl"
     }
   ],
   "recordSet": [
@@ -356,7 +380,7 @@
           "artifact_path": "README_REPRO.md",
           "artifact_role": "zero-api rebuild card",
           "encoding_format": "text/markdown",
-          "sha256": "c6e9a51e08dfd87c7b19f3ba63e40ceecabb3df06abacafc9907a523d5189c06",
+          "sha256": "5e0b142177ed0ca20479c161e0243d9f6560579c99404fef657951c21f74ca9a",
           "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset"
         },
         {
@@ -370,7 +394,7 @@
           "artifact_path": "abc_exp/results/DEBATE_EVAL_ARTIFACT_DATASHEET.md",
           "artifact_role": "artifact datasheet",
           "encoding_format": "text/markdown",
-          "sha256": "a5aa8fc5f8355743d836e609420f4a47f4e8d67c66c6a65caf231fa340848e96",
+          "sha256": "e4ab47578fca06c1904f680d78da011f43e96dcc58d0cb874a5cc734bb8fcbd3",
           "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset"
         },
         {
@@ -398,7 +422,7 @@
           "artifact_path": "abc_exp/results/conditional_collapse_table.json",
           "artifact_role": "collapse/correction transition table",
           "encoding_format": "application/json",
-          "sha256": "2df5c2ae2f8047c2d5706fe82339229774d4614c0583351042034f07ccaa5eba",
+          "sha256": "d6a11d548753d89b695d700e7f21755c414a5def11c2204acb7382192e9bbd21",
           "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset"
         },
         {
@@ -498,6 +522,20 @@
           "encoding_format": "application/jsonlines",
           "sha256": "72a7397af3a6469e64cc2412b58faca14c9dbd3366c8d3114b2d2093d82f9e5e",
           "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset"
+        },
+        {
+          "artifact_path": "abc_exp/results/gate_rerun/gate_lomo_paper.json",
+          "artifact_role": "probe-gated freeze ledger",
+          "encoding_format": "application/json",
+          "sha256": "1350c4e952a1e03995a9eb65f422b02fa4174af35a371d45e7d2fcef746c18f9",
+          "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset"
+        },
+        {
+          "artifact_path": "abc_exp/results/channel_panel/channel_panel_22.json",
+          "artifact_role": "22-model channel panel",
+          "encoding_format": "application/json",
+          "sha256": "5320543ee013297441def772f7ac62545cf406f81c55d699a8908e86801d1027",
+          "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset"
         }
       ]
     },
@@ -1164,6 +1202,110 @@
           }
         }
       ]
+    },
+    {
+      "@type": "cr:RecordSet",
+      "@id": "gate_rerun_probe_traces",
+      "name": "gate_rerun_probe_traces",
+      "description": "8-probe screen traces of the four local models collected for the probe-gated freeze and the channel panel. Model output text is null in the open tier.",
+      "field": [
+        {
+          "@type": "cr:Field",
+          "@id": "gate_rerun_probe_traces/question_id",
+          "name": "question_id",
+          "description": "Benchmark question identifier.",
+          "dataType": "sc:Text",
+          "repeated": false,
+          "source": {
+            "fileSet": {
+              "@id": "fileset_gate_rerun_probe_traces"
+            },
+            "extract": {
+              "column": "question_id"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "gate_rerun_probe_traces/agent_idx",
+          "name": "agent_idx",
+          "description": "Agent index within the panel.",
+          "dataType": "sc:Integer",
+          "repeated": false,
+          "source": {
+            "fileSet": {
+              "@id": "fileset_gate_rerun_probe_traces"
+            },
+            "extract": {
+              "column": "agent_idx"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "gate_rerun_probe_traces/condition",
+          "name": "condition",
+          "description": "System-prompt condition of the probe run.",
+          "dataType": "sc:Text",
+          "repeated": false,
+          "source": {
+            "fileSet": {
+              "@id": "fileset_gate_rerun_probe_traces"
+            },
+            "extract": {
+              "column": "condition"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "gate_rerun_probe_traces/initial_answer",
+          "name": "initial_answer",
+          "description": "Parsed answer before any probe.",
+          "dataType": "sc:Text",
+          "repeated": false,
+          "source": {
+            "fileSet": {
+              "@id": "fileset_gate_rerun_probe_traces"
+            },
+            "extract": {
+              "column": "initial_answer"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "gate_rerun_probe_traces/initial_correct",
+          "name": "initial_correct",
+          "description": "Whether the initial answer matches the key.",
+          "dataType": "sc:Boolean",
+          "repeated": false,
+          "source": {
+            "fileSet": {
+              "@id": "fileset_gate_rerun_probe_traces"
+            },
+            "extract": {
+              "column": "initial_correct"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "gate_rerun_probe_traces/flip_rate",
+          "name": "flip_rate",
+          "description": "Fraction of the eight probes that changed the answer.",
+          "dataType": "sc:Float",
+          "repeated": false,
+          "source": {
+            "fileSet": {
+              "@id": "fileset_gate_rerun_probe_traces"
+            },
+            "extract": {
+              "column": "flip_rate"
+            }
+          }
+        }
+      ]
     }
   ]
 }

abc_exp/results/debate_eval_croissant_validation_receipt.json

@@ -9,7 +9,7 @@
   "files": [
     {
       "path": "README_REPRO.md",
-      "sha256": "c6e9a51e08dfd87c7b19f3ba63e40ceecabb3df06abacafc9907a523d5189c06"
+      "sha256": "5e0b142177ed0ca20479c161e0243d9f6560579c99404fef657951c21f74ca9a"
     },
     {
       "path": "requirements-artifact.txt",
@@ -17,7 +17,7 @@
     },
     {
       "path": "abc_exp/results/DEBATE_EVAL_ARTIFACT_DATASHEET.md",
-      "sha256": "a5aa8fc5f8355743d836e609420f4a47f4e8d67c66c6a65caf231fa340848e96"
+      "sha256": "e4ab47578fca06c1904f680d78da011f43e96dcc58d0cb874a5cc734bb8fcbd3"
     },
     {
       "path": "abc_exp/results/debate_eval_reuse_card.schema.json",
@@ -33,7 +33,7 @@
     },
     {
       "path": "abc_exp/results/conditional_collapse_table.json",
-      "sha256": "2df5c2ae2f8047c2d5706fe82339229774d4614c0583351042034f07ccaa5eba"
+      "sha256": "d6a11d548753d89b695d700e7f21755c414a5def11c2204acb7382192e9bbd21"
     },
     {
       "path": "abc_exp/results/reviewer_hardening_analyses.json",
@@ -90,6 +90,14 @@
     {
       "path": "abc_exp/results/block3_multi_benchmark.jsonl",
       "sha256": "72a7397af3a6469e64cc2412b58faca14c9dbd3366c8d3114b2d2093d82f9e5e"
+    },
+    {
+      "path": "abc_exp/results/gate_rerun/gate_lomo_paper.json",
+      "sha256": "1350c4e952a1e03995a9eb65f422b02fa4174af35a371d45e7d2fcef746c18f9"
+    },
+    {
+      "path": "abc_exp/results/channel_panel/channel_panel_22.json",
+      "sha256": "5320543ee013297441def772f7ac62545cf406f81c55d699a8908e86801d1027"
     }
   ],
   "rai_fields_populated": [

abc_exp/results/gate_rerun/acceptance.json

@@ -0,0 +1,182 @@
+{
+ "sa_causal_vllm_llama-3.1-8b_rerun.jsonl": {
+  "present": true,
+  "cap_probe": [
+   2048
+  ],
+  "expected": 1800,
+  "rows": 1800,
+  "missing": 0,
+  "duplicates": 0,
+  "rows_not_8": 0,
+  "init_trunc_pct": 10.17,
+  "probe_trunc_pct": 15.84,
+  "trunc_loop_share": 0.0,
+  "init_none": 0,
+  "post_none_pct": 3.74,
+  "verdict": "OVER_2PCT",
+  "complete": true
+ },
+ "sa_causal_vllm_phi-4-mini_rerun.jsonl": {
+  "present": true,
+  "cap_probe": [
+   2048
+  ],
+  "expected": 1800,
+  "rows": 1800,
+  "missing": 0,
+  "duplicates": 0,
+  "rows_not_8": 0,
+  "init_trunc_pct": 0.39,
+  "probe_trunc_pct": 1.8,
+  "trunc_loop_share": 0.0,
+  "init_none": 0,
+  "post_none_pct": 1.06,
+  "verdict": "OK",
+  "complete": true
+ },
+ "sa_causal_vllm_qwen3-4b_rerun.jsonl": {
+  "present": true,
+  "cap_probe": [
+   2048
+  ],
+  "expected": 1800,
+  "rows": 1800,
+  "missing": 0,
+  "duplicates": 0,
+  "rows_not_8": 0,
+  "init_trunc_pct": 0.22,
+  "probe_trunc_pct": 0.07,
+  "trunc_loop_share": 0.0,
+  "init_none": 0,
+  "post_none_pct": 0.01,
+  "verdict": "OK",
+  "complete": true
+ },
+ "sa_causal_vllm_qwen3-8b_rerun.jsonl": {
+  "present": true,
+  "cap_probe": [
+   2048
+  ],
+  "expected": 1800,
+  "rows": 1800,
+  "missing": 0,
+  "duplicates": 0,
+  "rows_not_8": 0,
+  "init_trunc_pct": 0.28,
+  "probe_trunc_pct": 1.18,
+  "trunc_loop_share": 0.0,
+  "init_none": 0,
+  "post_none_pct": 0.24,
+  "verdict": "OK",
+  "complete": true
+ },
+ "gate_default_vllm_llama-3.1-8b_s0.jsonl": {
+  "present": true,
+  "cap_probe": [
+   2048
+  ],
+  "expected": 2685,
+  "rows": 2685,
+  "missing": 0,
+  "duplicates": 0,
+  "rows_not_8": 0,
+  "init_trunc_pct": 3.87,
+  "probe_trunc_pct": 12.77,
+  "trunc_loop_share": 0.0,
+  "init_none": 0,
+  "post_none_pct": 5.36,
+  "verdict": "OVER_2PCT",
+  "complete": true
+ },
+ "gate_default_vllm_llama-3.1-8b_s1.jsonl": {
+  "present": true,
+  "cap_probe": [
+   2048
+  ],
+  "expected": 2685,
+  "rows": 2685,
+  "missing": 0,
+  "duplicates": 0,
+  "rows_not_8": 0,
+  "init_trunc_pct": 4.99,
+  "probe_trunc_pct": 15.82,
+  "trunc_loop_share": 0.0,
+  "init_none": 0,
+  "post_none_pct": 6.48,
+  "verdict": "OVER_2PCT",
+  "complete": true
+ },
+ "gate_default_vllm_phi-4-mini_s0.jsonl": {
+  "present": true,
+  "cap_probe": [
+   2048
+  ],
+  "expected": 2943,
+  "rows": 2943,
+  "missing": 0,
+  "duplicates": 0,
+  "rows_not_8": 0,
+  "init_trunc_pct": 0.2,
+  "probe_trunc_pct": 1.06,
+  "trunc_loop_share": 0.0,
+  "init_none": 0,
+  "post_none_pct": 0.9,
+  "verdict": "OK",
+  "complete": true
+ },
+ "gate_default_vllm_phi-4-mini_s1.jsonl": {
+  "present": true,
+  "cap_probe": [
+   2048
+  ],
+  "expected": 2940,
+  "rows": 2940,
+  "missing": 0,
+  "duplicates": 0,
+  "rows_not_8": 0,
+  "init_trunc_pct": 0.51,
+  "probe_trunc_pct": 2.4,
+  "trunc_loop_share": 0.0,
+  "init_none": 0,
+  "post_none_pct": 1.72,
+  "verdict": "OVER_2PCT",
+  "complete": true
+ },
+ "gate_default_vllm_qwen3-4b_s0.jsonl": {
+  "present": true,
+  "cap_probe": [
+   2048
+  ],
+  "expected": 2943,
+  "rows": 2943,
+  "missing": 0,
+  "duplicates": 0,
+  "rows_not_8": 0,
+  "init_trunc_pct": 0.24,
+  "probe_trunc_pct": 0.17,
+  "trunc_loop_share": 0.0,
+  "init_none": 0,
+  "post_none_pct": 0.0,
+  "verdict": "OK",
+  "complete": true
+ },
+ "gate_default_vllm_qwen3-4b_s1.jsonl": {
+  "present": true,
+  "cap_probe": [
+   2048
+  ],
+  "expected": 2940,
+  "rows": 2939,
+  "missing": 1,
+  "duplicates": 0,
+  "rows_not_8": 0,
+  "init_trunc_pct": 0.31,
+  "probe_trunc_pct": 0.26,
+  "trunc_loop_share": 0.0,
+  "init_none": 0,
+  "post_none_pct": 0.03,
+  "verdict": "OK",
+  "complete": true
+ }
+}

abc_exp/results/gate_rerun/gate_default_vllm_phi-4-mini_s0.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:259780dd157d1a5fd6b1dce21d6e5b48080947aeb0d7017acf1c83104334733b
+size 10753798

abc_exp/results/gate_rerun/gate_default_vllm_phi-4-mini_s1.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e3c3f0751e290f20d6baa882999feb50d3d7c880c4c5c44517567f6369066d52
+size 10754344

abc_exp/results/gate_rerun/gate_default_vllm_qwen3-4b_s0.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0cf8673f92e40fecce002e0966ddffd990a8e27b0e872a6da363447b4726cef4
+size 10702553

abc_exp/results/gate_rerun/gate_default_vllm_qwen3-4b_s1.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:86924421edded2ca56750d3c49980c07350ac3f7c478d449484cc5594c887b5a
+size 10693677

abc_exp/results/gate_rerun/gate_lomo_drop.json

@@ -0,0 +1,725 @@
+{
+ "selftest": false,
+ "notes": "Probe-gated freeze with leave-one-model-out selection on the archived local-model debates (zero API).\nThe scripts and probe inputs of the original pilot were not retained. This re-implements the pilot rule on newly\ncollected probe scores (abc_exp/results/gate_rerun/); it is a post-hoc reconstruction, not a reproduction of the pilot.\n\nRules of the original pilot: tau in {0.05,0.1,0.15,0.2,0.3,0.4,0.5,0.6,0.7}; mode in {freeze_initial, freeze_if_stable}; for each held-out\nmodel, (tau*, mode*) maximizes net = prevented - lost on the other models; then frozen and applied to the held-out model.\nChoices the original pilot does not record (made here):\n  * classifier: logistic regression of `collapsed` on per-debate probe features (mean, max, min of the three agents'\n    default-condition 8-probe flip rate on that question), fitted on the training models; the gate fires when p >= tau.\n  * freeze_initial: when the gate fires, the debate returns its initial majority.\n  * freeze_if_stable: when the gate fires, an agent whose Round-1 answer equals its initial answer keeps its initial answer;\n    the other agents keep their final answers; the majority is recomputed (unparsed answers dropped, ties alphabetical).\nOutputs per held-out model: tau*, mode*, gate%, prevented, lost, net, delta-acc (pp) with a debate-bootstrap 95% CI, and the\noracle delta-acc (best tau/mode chosen on the held-out model itself); pooled totals; weight ladder (r * prevented - lost) / n.",
+ "variants": {
+  "e5_rule": {
+   "per_model": {
+    "llama-3.1-8b": {
+     "n": 2003,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5132,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.05,
+      "mode": "freeze_if_stable",
+      "dacc_pp": 0.3
+     }
+    },
+    "phi-4-mini": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5349,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    },
+    "qwen3-4b": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.6432,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    },
+    "qwen3-8b": {
+     "n": 200,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.62,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    }
+   },
+   "pooled": {
+    "n": 6525,
+    "prevented": 0,
+    "lost": 0,
+    "net": 0,
+    "dacc_pp": 0.0,
+    "ci95": [
+     0.0,
+     0.0
+    ],
+    "weight_ladder_pp": {
+     "0.5": 0.0,
+     "1": 0.0,
+     "2": 0.0,
+     "4": 0.0,
+     "8": 0.0
+    },
+    "break_even_r": null
+   }
+  },
+  "freeze_initial_only": {
+   "per_model": {
+    "llama-3.1-8b": {
+     "n": 2003,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5132,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    },
+    "phi-4-mini": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5349,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    },
+    "qwen3-4b": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.6432,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    },
+    "qwen3-8b": {
+     "n": 200,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.62,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    }
+   },
+   "pooled": {
+    "n": 6525,
+    "prevented": 0,
+    "lost": 0,
+    "net": 0,
+    "dacc_pp": 0.0,
+    "ci95": [
+     0.0,
+     0.0
+    ],
+    "weight_ladder_pp": {
+     "0.5": 0.0,
+     "1": 0.0,
+     "2": 0.0,
+     "4": 0.0,
+     "8": 0.0
+    },
+    "break_even_r": null
+   }
+  },
+  "nested_lomo": {
+   "per_model": {
+    "llama-3.1-8b": {
+     "n": 2003,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5132,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.05,
+      "mode": "freeze_if_stable",
+      "dacc_pp": 0.3
+     }
+    },
+    "phi-4-mini": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5349,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    },
+    "qwen3-4b": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": 0.05,
+     "mode": "freeze_if_stable",
+     "gate_pct": 31.79,
+     "baseline_acc": 0.6432,
+     "prevented": 8,
+     "lost": 41,
+     "other_net": 1,
+     "net": -32,
+     "dacc_pp": -1.481,
+     "ci95": [
+      -2.129,
+      -0.833
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    },
+    "qwen3-8b": {
+     "n": 200,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.62,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    }
+   },
+   "pooled": {
+    "n": 6525,
+    "prevented": 8,
+    "lost": 41,
+    "net": -32,
+    "dacc_pp": -0.49,
+    "ci95": [
+     -0.705,
+     -0.29
+    ],
+    "weight_ladder_pp": {
+     "0.5": -0.567,
+     "1": -0.506,
+     "2": -0.383,
+     "4": -0.138,
+     "8": 0.352
+    },
+    "break_even_r": 5.125
+   }
+  },
+  "with_noop": {
+   "per_model": {
+    "llama-3.1-8b": {
+     "n": 2003,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "no_op",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5132,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.05,
+      "mode": "freeze_if_stable",
+      "dacc_pp": 0.3
+     }
+    },
+    "phi-4-mini": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "no_op",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5349,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    },
+    "qwen3-4b": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "no_op",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.6432,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    },
+    "qwen3-8b": {
+     "n": 200,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "no_op",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.62,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    }
+   },
+   "pooled": {
+    "n": 6525,
+    "prevented": 0,
+    "lost": 0,
+    "net": 0,
+    "dacc_pp": 0.0,
+    "ci95": [
+     0.0,
+     0.0
+    ],
+    "weight_ladder_pp": {
+     "0.5": 0.0,
+     "1": 0.0,
+     "2": 0.0,
+     "4": 0.0,
+     "8": 0.0
+    },
+    "break_even_r": null
+   }
+  },
+  "always_freeze": {
+   "per_model": {
+    "llama-3.1-8b": {
+     "n": 2003,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "freeze_initial",
+     "gate_pct": 100.0,
+     "baseline_acc": 0.5132,
+     "prevented": 86,
+     "lost": 98,
+     "other_net": 0,
+     "net": -12,
+     "dacc_pp": -0.599,
+     "ci95": [
+      -1.898,
+      0.75
+     ]
+    },
+    "phi-4-mini": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "freeze_initial",
+     "gate_pct": 100.0,
+     "baseline_acc": 0.5349,
+     "prevented": 112,
+     "lost": 174,
+     "other_net": 0,
+     "net": -62,
+     "dacc_pp": -2.869,
+     "ci95": [
+      -4.396,
+      -1.296
+     ]
+    },
+    "qwen3-4b": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "freeze_initial",
+     "gate_pct": 100.0,
+     "baseline_acc": 0.6432,
+     "prevented": 50,
+     "lost": 461,
+     "other_net": 0,
+     "net": -411,
+     "dacc_pp": -19.019,
+     "ci95": [
+      -20.916,
+      -17.075
+     ]
+    },
+    "qwen3-8b": {
+     "n": 200,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "freeze_initial",
+     "gate_pct": 100.0,
+     "baseline_acc": 0.62,
+     "prevented": 3,
+     "lost": 71,
+     "other_net": 0,
+     "net": -68,
+     "dacc_pp": -34.0,
+     "ci95": [
+      -41.0,
+      -27.0
+     ]
+    }
+   },
+   "pooled": {
+    "n": 6525,
+    "prevented": 251,
+    "lost": 804,
+    "net": -553,
+    "dacc_pp": -8.475,
+    "ci95": [
+     -9.484,
+     -7.484
+    ],
+    "weight_ladder_pp": {
+     "0.5": -10.398,
+     "1": -8.475,
+     "2": -4.628,
+     "4": 3.065,
+     "8": 18.452
+    },
+    "break_even_r": 3.203
+   }
+  },
+  "raw_fr_gate": {
+   "per_model": {
+    "llama-3.1-8b": {
+     "n": 2003,
+     "n_missing_probe": 0,
+     "tau": 0.7,
+     "mode": "freeze_initial",
+     "gate_pct": 51.42,
+     "baseline_acc": 0.5132,
+     "prevented": 56,
+     "lost": 61,
+     "other_net": 0,
+     "net": -5,
+     "dacc_pp": -0.25,
+     "ci95": [
+      -1.295,
+      0.8
+     ]
+    },
+    "phi-4-mini": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": 0.7,
+     "mode": "freeze_initial",
+     "gate_pct": 18.65,
+     "baseline_acc": 0.5349,
+     "prevented": 21,
+     "lost": 47,
+     "other_net": 0,
+     "net": -26,
+     "dacc_pp": -1.203,
+     "ci95": [
+      -1.99,
+      -0.508
+     ]
+    },
+    "qwen3-4b": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": 0.7,
+     "mode": "freeze_initial",
+     "gate_pct": 37.02,
+     "baseline_acc": 0.6432,
+     "prevented": 29,
+     "lost": 187,
+     "other_net": 0,
+     "net": -158,
+     "dacc_pp": -7.311,
+     "ci95": [
+      -8.562,
+      -6.015
+     ]
+    },
+    "qwen3-8b": {
+     "n": 200,
+     "n_missing_probe": 0,
+     "tau": 0.7,
+     "mode": "freeze_initial",
+     "gate_pct": 12.5,
+     "baseline_acc": 0.62,
+     "prevented": 2,
+     "lost": 2,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      -2.0,
+      2.0
+     ]
+    }
+   },
+   "pooled": {
+    "n": 6525,
+    "prevented": 108,
+    "lost": 297,
+    "net": -189,
+    "dacc_pp": -2.897,
+    "ci95": [
+     -3.496,
+     -2.336
+    ],
+    "weight_ladder_pp": {
+     "0.5": -3.724,
+     "1": -2.897,
+     "2": -1.241,
+     "4": 2.069,
+     "8": 8.69
+    },
+    "break_even_r": 2.75
+   }
+  },
+  "random_same_rate": {
+   "per_model": {
+    "llama-3.1-8b": {
+     "n": 2003,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5132,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ]
+    },
+    "phi-4-mini": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5349,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ]
+    },
+    "qwen3-4b": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.6432,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ]
+    },
+    "qwen3-8b": {
+     "n": 200,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.62,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ]
+    }
+   },
+   "pooled": {
+    "n": 6525,
+    "prevented": 0,
+    "lost": 0,
+    "net": 0,
+    "dacc_pp": 0.0,
+    "ci95": [
+     0.0,
+     0.0
+    ],
+    "weight_ladder_pp": {
+     "0.5": 0.0,
+     "1": 0.0,
+     "2": 0.0,
+     "4": 0.0,
+     "8": 0.0
+    },
+    "break_even_r": null
+   }
+  }
+ },
+ "none_mode": "drop",
+ "probes_dir": "abc_exp/results/gate_rerun"
+}

abc_exp/results/gate_rerun/gate_lomo_flip.json

@@ -0,0 +1,725 @@
+{
+ "selftest": false,
+ "notes": "Probe-gated freeze with leave-one-model-out selection on the archived local-model debates (zero API).\nThe scripts and probe inputs of the original pilot were not retained. This re-implements the pilot rule on newly\ncollected probe scores (abc_exp/results/gate_rerun/); it is a post-hoc reconstruction, not a reproduction of the pilot.\n\nRules of the original pilot: tau in {0.05,0.1,0.15,0.2,0.3,0.4,0.5,0.6,0.7}; mode in {freeze_initial, freeze_if_stable}; for each held-out\nmodel, (tau*, mode*) maximizes net = prevented - lost on the other models; then frozen and applied to the held-out model.\nChoices the original pilot does not record (made here):\n  * classifier: logistic regression of `collapsed` on per-debate probe features (mean, max, min of the three agents'\n    default-condition 8-probe flip rate on that question), fitted on the training models; the gate fires when p >= tau.\n  * freeze_initial: when the gate fires, the debate returns its initial majority.\n  * freeze_if_stable: when the gate fires, an agent whose Round-1 answer equals its initial answer keeps its initial answer;\n    the other agents keep their final answers; the majority is recomputed (unparsed answers dropped, ties alphabetical).\nOutputs per held-out model: tau*, mode*, gate%, prevented, lost, net, delta-acc (pp) with a debate-bootstrap 95% CI, and the\noracle delta-acc (best tau/mode chosen on the held-out model itself); pooled totals; weight ladder (r * prevented - lost) / n.",
+ "variants": {
+  "e5_rule": {
+   "per_model": {
+    "llama-3.1-8b": {
+     "n": 2003,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5132,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.05,
+      "mode": "freeze_if_stable",
+      "dacc_pp": 0.2
+     }
+    },
+    "phi-4-mini": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5349,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    },
+    "qwen3-4b": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.6432,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    },
+    "qwen3-8b": {
+     "n": 200,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.62,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    }
+   },
+   "pooled": {
+    "n": 6525,
+    "prevented": 0,
+    "lost": 0,
+    "net": 0,
+    "dacc_pp": 0.0,
+    "ci95": [
+     0.0,
+     0.0
+    ],
+    "weight_ladder_pp": {
+     "0.5": 0.0,
+     "1": 0.0,
+     "2": 0.0,
+     "4": 0.0,
+     "8": 0.0
+    },
+    "break_even_r": null
+   }
+  },
+  "freeze_initial_only": {
+   "per_model": {
+    "llama-3.1-8b": {
+     "n": 2003,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5132,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    },
+    "phi-4-mini": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5349,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    },
+    "qwen3-4b": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.6432,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    },
+    "qwen3-8b": {
+     "n": 200,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.62,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    }
+   },
+   "pooled": {
+    "n": 6525,
+    "prevented": 0,
+    "lost": 0,
+    "net": 0,
+    "dacc_pp": 0.0,
+    "ci95": [
+     0.0,
+     0.0
+    ],
+    "weight_ladder_pp": {
+     "0.5": 0.0,
+     "1": 0.0,
+     "2": 0.0,
+     "4": 0.0,
+     "8": 0.0
+    },
+    "break_even_r": null
+   }
+  },
+  "nested_lomo": {
+   "per_model": {
+    "llama-3.1-8b": {
+     "n": 2003,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5132,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.05,
+      "mode": "freeze_if_stable",
+      "dacc_pp": 0.2
+     }
+    },
+    "phi-4-mini": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5349,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    },
+    "qwen3-4b": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": 0.05,
+     "mode": "freeze_if_stable",
+     "gate_pct": 24.43,
+     "baseline_acc": 0.6432,
+     "prevented": 7,
+     "lost": 31,
+     "other_net": 1,
+     "net": -23,
+     "dacc_pp": -1.064,
+     "ci95": [
+      -1.621,
+      -0.509
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    },
+    "qwen3-8b": {
+     "n": 200,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.62,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    }
+   },
+   "pooled": {
+    "n": 6525,
+    "prevented": 7,
+    "lost": 31,
+    "net": -23,
+    "dacc_pp": -0.352,
+    "ci95": [
+     -0.536,
+     -0.168
+    ],
+    "weight_ladder_pp": {
+     "0.5": -0.421,
+     "1": -0.368,
+     "2": -0.261,
+     "4": -0.046,
+     "8": 0.383
+    },
+    "break_even_r": 4.429
+   }
+  },
+  "with_noop": {
+   "per_model": {
+    "llama-3.1-8b": {
+     "n": 2003,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "no_op",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5132,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.05,
+      "mode": "freeze_if_stable",
+      "dacc_pp": 0.2
+     }
+    },
+    "phi-4-mini": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "no_op",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5349,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    },
+    "qwen3-4b": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "no_op",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.6432,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    },
+    "qwen3-8b": {
+     "n": 200,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "no_op",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.62,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    }
+   },
+   "pooled": {
+    "n": 6525,
+    "prevented": 0,
+    "lost": 0,
+    "net": 0,
+    "dacc_pp": 0.0,
+    "ci95": [
+     0.0,
+     0.0
+    ],
+    "weight_ladder_pp": {
+     "0.5": 0.0,
+     "1": 0.0,
+     "2": 0.0,
+     "4": 0.0,
+     "8": 0.0
+    },
+    "break_even_r": null
+   }
+  },
+  "always_freeze": {
+   "per_model": {
+    "llama-3.1-8b": {
+     "n": 2003,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "freeze_initial",
+     "gate_pct": 100.0,
+     "baseline_acc": 0.5132,
+     "prevented": 86,
+     "lost": 98,
+     "other_net": 0,
+     "net": -12,
+     "dacc_pp": -0.599,
+     "ci95": [
+      -1.898,
+      0.75
+     ]
+    },
+    "phi-4-mini": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "freeze_initial",
+     "gate_pct": 100.0,
+     "baseline_acc": 0.5349,
+     "prevented": 112,
+     "lost": 174,
+     "other_net": 0,
+     "net": -62,
+     "dacc_pp": -2.869,
+     "ci95": [
+      -4.396,
+      -1.296
+     ]
+    },
+    "qwen3-4b": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "freeze_initial",
+     "gate_pct": 100.0,
+     "baseline_acc": 0.6432,
+     "prevented": 50,
+     "lost": 461,
+     "other_net": 0,
+     "net": -411,
+     "dacc_pp": -19.019,
+     "ci95": [
+      -20.916,
+      -17.075
+     ]
+    },
+    "qwen3-8b": {
+     "n": 200,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "freeze_initial",
+     "gate_pct": 100.0,
+     "baseline_acc": 0.62,
+     "prevented": 3,
+     "lost": 71,
+     "other_net": 0,
+     "net": -68,
+     "dacc_pp": -34.0,
+     "ci95": [
+      -41.0,
+      -27.0
+     ]
+    }
+   },
+   "pooled": {
+    "n": 6525,
+    "prevented": 251,
+    "lost": 804,
+    "net": -553,
+    "dacc_pp": -8.475,
+    "ci95": [
+     -9.484,
+     -7.484
+    ],
+    "weight_ladder_pp": {
+     "0.5": -10.398,
+     "1": -8.475,
+     "2": -4.628,
+     "4": 3.065,
+     "8": 18.452
+    },
+    "break_even_r": 3.203
+   }
+  },
+  "raw_fr_gate": {
+   "per_model": {
+    "llama-3.1-8b": {
+     "n": 2003,
+     "n_missing_probe": 0,
+     "tau": 0.7,
+     "mode": "freeze_initial",
+     "gate_pct": 54.37,
+     "baseline_acc": 0.5132,
+     "prevented": 60,
+     "lost": 71,
+     "other_net": 0,
+     "net": -11,
+     "dacc_pp": -0.549,
+     "ci95": [
+      -1.643,
+      0.598
+     ]
+    },
+    "phi-4-mini": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": 0.7,
+     "mode": "freeze_initial",
+     "gate_pct": 19.76,
+     "baseline_acc": 0.5349,
+     "prevented": 24,
+     "lost": 49,
+     "other_net": 0,
+     "net": -25,
+     "dacc_pp": -1.157,
+     "ci95": [
+      -1.944,
+      -0.37
+     ]
+    },
+    "qwen3-4b": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": 0.7,
+     "mode": "freeze_initial",
+     "gate_pct": 37.02,
+     "baseline_acc": 0.6432,
+     "prevented": 29,
+     "lost": 187,
+     "other_net": 0,
+     "net": -158,
+     "dacc_pp": -7.311,
+     "ci95": [
+      -8.562,
+      -6.015
+     ]
+    },
+    "qwen3-8b": {
+     "n": 200,
+     "n_missing_probe": 0,
+     "tau": 0.7,
+     "mode": "freeze_initial",
+     "gate_pct": 12.5,
+     "baseline_acc": 0.62,
+     "prevented": 2,
+     "lost": 2,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      -2.0,
+      2.0
+     ]
+    }
+   },
+   "pooled": {
+    "n": 6525,
+    "prevented": 115,
+    "lost": 309,
+    "net": -194,
+    "dacc_pp": -2.973,
+    "ci95": [
+     -3.587,
+     -2.37
+    ],
+    "weight_ladder_pp": {
+     "0.5": -3.854,
+     "1": -2.973,
+     "2": -1.211,
+     "4": 2.314,
+     "8": 9.364
+    },
+    "break_even_r": 2.687
+   }
+  },
+  "random_same_rate": {
+   "per_model": {
+    "llama-3.1-8b": {
+     "n": 2003,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5132,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ]
+    },
+    "phi-4-mini": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5349,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ]
+    },
+    "qwen3-4b": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.6432,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ]
+    },
+    "qwen3-8b": {
+     "n": 200,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.62,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ]
+    }
+   },
+   "pooled": {
+    "n": 6525,
+    "prevented": 0,
+    "lost": 0,
+    "net": 0,
+    "dacc_pp": 0.0,
+    "ci95": [
+     0.0,
+     0.0
+    ],
+    "weight_ladder_pp": {
+     "0.5": 0.0,
+     "1": 0.0,
+     "2": 0.0,
+     "4": 0.0,
+     "8": 0.0
+    },
+    "break_even_r": null
+   }
+  }
+ },
+ "none_mode": "flip",
+ "probes_dir": "abc_exp/results/gate_rerun"
+}

abc_exp/results/gate_rerun/gate_lomo_paper.json

@@ -0,0 +1,725 @@
+{
+ "selftest": false,
+ "notes": "Probe-gated freeze with leave-one-model-out selection on the archived local-model debates (zero API).\nThe scripts and probe inputs of the original pilot were not retained. This re-implements the pilot rule on newly\ncollected probe scores (abc_exp/results/gate_rerun/); it is a post-hoc reconstruction, not a reproduction of the pilot.\n\nRules of the original pilot: tau in {0.05,0.1,0.15,0.2,0.3,0.4,0.5,0.6,0.7}; mode in {freeze_initial, freeze_if_stable}; for each held-out\nmodel, (tau*, mode*) maximizes net = prevented - lost on the other models; then frozen and applied to the held-out model.\nChoices the original pilot does not record (made here):\n  * classifier: logistic regression of `collapsed` on per-debate probe features (mean, max, min of the three agents'\n    default-condition 8-probe flip rate on that question), fitted on the training models; the gate fires when p >= tau.\n  * freeze_initial: when the gate fires, the debate returns its initial majority.\n  * freeze_if_stable: when the gate fires, an agent whose Round-1 answer equals its initial answer keeps its initial answer;\n    the other agents keep their final answers; the majority is recomputed (unparsed answers dropped, ties alphabetical).\nOutputs per held-out model: tau*, mode*, gate%, prevented, lost, net, delta-acc (pp) with a debate-bootstrap 95% CI, and the\noracle delta-acc (best tau/mode chosen on the held-out model itself); pooled totals; weight ladder (r * prevented - lost) / n.",
+ "variants": {
+  "e5_rule": {
+   "per_model": {
+    "llama-3.1-8b": {
+     "n": 2003,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5132,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.05,
+      "mode": "freeze_if_stable",
+      "dacc_pp": 0.15
+     }
+    },
+    "phi-4-mini": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5349,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.05,
+      "mode": "freeze_if_stable",
+      "dacc_pp": 0.093
+     }
+    },
+    "qwen3-4b": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.6432,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    },
+    "qwen3-8b": {
+     "n": 200,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.62,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    }
+   },
+   "pooled": {
+    "n": 6525,
+    "prevented": 0,
+    "lost": 0,
+    "net": 0,
+    "dacc_pp": 0.0,
+    "ci95": [
+     0.0,
+     0.0
+    ],
+    "weight_ladder_pp": {
+     "0.5": 0.0,
+     "1": 0.0,
+     "2": 0.0,
+     "4": 0.0,
+     "8": 0.0
+    },
+    "break_even_r": null
+   }
+  },
+  "freeze_initial_only": {
+   "per_model": {
+    "llama-3.1-8b": {
+     "n": 2003,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5132,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    },
+    "phi-4-mini": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5349,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    },
+    "qwen3-4b": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.6432,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    },
+    "qwen3-8b": {
+     "n": 200,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.62,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    }
+   },
+   "pooled": {
+    "n": 6525,
+    "prevented": 0,
+    "lost": 0,
+    "net": 0,
+    "dacc_pp": 0.0,
+    "ci95": [
+     0.0,
+     0.0
+    ],
+    "weight_ladder_pp": {
+     "0.5": 0.0,
+     "1": 0.0,
+     "2": 0.0,
+     "4": 0.0,
+     "8": 0.0
+    },
+    "break_even_r": null
+   }
+  },
+  "nested_lomo": {
+   "per_model": {
+    "llama-3.1-8b": {
+     "n": 2003,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5132,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.05,
+      "mode": "freeze_if_stable",
+      "dacc_pp": 0.15
+     }
+    },
+    "phi-4-mini": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5349,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.05,
+      "mode": "freeze_if_stable",
+      "dacc_pp": 0.093
+     }
+    },
+    "qwen3-4b": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": 0.05,
+     "mode": "freeze_if_stable",
+     "gate_pct": 20.87,
+     "baseline_acc": 0.6432,
+     "prevented": 6,
+     "lost": 26,
+     "other_net": 1,
+     "net": -19,
+     "dacc_pp": -0.879,
+     "ci95": [
+      -1.435,
+      -0.37
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    },
+    "qwen3-8b": {
+     "n": 200,
+     "n_missing_probe": 0,
+     "tau": 0.1,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.62,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    }
+   },
+   "pooled": {
+    "n": 6525,
+    "prevented": 6,
+    "lost": 26,
+    "net": -19,
+    "dacc_pp": -0.291,
+    "ci95": [
+     -0.461,
+     -0.122
+    ],
+    "weight_ladder_pp": {
+     "0.5": -0.352,
+     "1": -0.307,
+     "2": -0.215,
+     "4": -0.031,
+     "8": 0.337
+    },
+    "break_even_r": 4.333
+   }
+  },
+  "with_noop": {
+   "per_model": {
+    "llama-3.1-8b": {
+     "n": 2003,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "no_op",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5132,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.05,
+      "mode": "freeze_if_stable",
+      "dacc_pp": 0.15
+     }
+    },
+    "phi-4-mini": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "no_op",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5349,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.05,
+      "mode": "freeze_if_stable",
+      "dacc_pp": 0.093
+     }
+    },
+    "qwen3-4b": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "no_op",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.6432,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    },
+    "qwen3-8b": {
+     "n": 200,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "no_op",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.62,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ],
+     "oracle": {
+      "tau": 0.1,
+      "mode": "freeze_initial",
+      "dacc_pp": 0.0
+     }
+    }
+   },
+   "pooled": {
+    "n": 6525,
+    "prevented": 0,
+    "lost": 0,
+    "net": 0,
+    "dacc_pp": 0.0,
+    "ci95": [
+     0.0,
+     0.0
+    ],
+    "weight_ladder_pp": {
+     "0.5": 0.0,
+     "1": 0.0,
+     "2": 0.0,
+     "4": 0.0,
+     "8": 0.0
+    },
+    "break_even_r": null
+   }
+  },
+  "always_freeze": {
+   "per_model": {
+    "llama-3.1-8b": {
+     "n": 2003,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "freeze_initial",
+     "gate_pct": 100.0,
+     "baseline_acc": 0.5132,
+     "prevented": 86,
+     "lost": 98,
+     "other_net": 0,
+     "net": -12,
+     "dacc_pp": -0.599,
+     "ci95": [
+      -1.898,
+      0.75
+     ]
+    },
+    "phi-4-mini": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "freeze_initial",
+     "gate_pct": 100.0,
+     "baseline_acc": 0.5349,
+     "prevented": 112,
+     "lost": 174,
+     "other_net": 0,
+     "net": -62,
+     "dacc_pp": -2.869,
+     "ci95": [
+      -4.396,
+      -1.296
+     ]
+    },
+    "qwen3-4b": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "freeze_initial",
+     "gate_pct": 100.0,
+     "baseline_acc": 0.6432,
+     "prevented": 50,
+     "lost": 461,
+     "other_net": 0,
+     "net": -411,
+     "dacc_pp": -19.019,
+     "ci95": [
+      -20.916,
+      -17.075
+     ]
+    },
+    "qwen3-8b": {
+     "n": 200,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "freeze_initial",
+     "gate_pct": 100.0,
+     "baseline_acc": 0.62,
+     "prevented": 3,
+     "lost": 71,
+     "other_net": 0,
+     "net": -68,
+     "dacc_pp": -34.0,
+     "ci95": [
+      -41.0,
+      -27.0
+     ]
+    }
+   },
+   "pooled": {
+    "n": 6525,
+    "prevented": 251,
+    "lost": 804,
+    "net": -553,
+    "dacc_pp": -8.475,
+    "ci95": [
+     -9.484,
+     -7.484
+    ],
+    "weight_ladder_pp": {
+     "0.5": -10.398,
+     "1": -8.475,
+     "2": -4.628,
+     "4": 3.065,
+     "8": 18.452
+    },
+    "break_even_r": 3.203
+   }
+  },
+  "raw_fr_gate": {
+   "per_model": {
+    "llama-3.1-8b": {
+     "n": 2003,
+     "n_missing_probe": 0,
+     "tau": 0.7,
+     "mode": "freeze_initial",
+     "gate_pct": 41.99,
+     "baseline_acc": 0.5132,
+     "prevented": 45,
+     "lost": 46,
+     "other_net": 0,
+     "net": -1,
+     "dacc_pp": -0.05,
+     "ci95": [
+      -0.998,
+      0.898
+     ]
+    },
+    "phi-4-mini": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": 0.7,
+     "mode": "freeze_initial",
+     "gate_pct": 17.58,
+     "baseline_acc": 0.5349,
+     "prevented": 21,
+     "lost": 45,
+     "other_net": 0,
+     "net": -24,
+     "dacc_pp": -1.111,
+     "ci95": [
+      -1.852,
+      -0.416
+     ]
+    },
+    "qwen3-4b": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": 0.7,
+     "mode": "freeze_initial",
+     "gate_pct": 37.02,
+     "baseline_acc": 0.6432,
+     "prevented": 29,
+     "lost": 187,
+     "other_net": 0,
+     "net": -158,
+     "dacc_pp": -7.311,
+     "ci95": [
+      -8.562,
+      -6.015
+     ]
+    },
+    "qwen3-8b": {
+     "n": 200,
+     "n_missing_probe": 0,
+     "tau": 0.7,
+     "mode": "freeze_initial",
+     "gate_pct": 12.0,
+     "baseline_acc": 0.62,
+     "prevented": 2,
+     "lost": 2,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      -2.0,
+      2.0
+     ]
+    }
+   },
+   "pooled": {
+    "n": 6525,
+    "prevented": 97,
+    "lost": 280,
+    "net": -183,
+    "dacc_pp": -2.805,
+    "ci95": [
+     -3.406,
+     -2.24
+    ],
+    "weight_ladder_pp": {
+     "0.5": -3.548,
+     "1": -2.805,
+     "2": -1.318,
+     "4": 1.655,
+     "8": 7.602
+    },
+    "break_even_r": 2.887
+   }
+  },
+  "random_same_rate": {
+   "per_model": {
+    "llama-3.1-8b": {
+     "n": 2003,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5132,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ]
+    },
+    "phi-4-mini": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.5349,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ]
+    },
+    "qwen3-4b": {
+     "n": 2161,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.6432,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ]
+    },
+    "qwen3-8b": {
+     "n": 200,
+     "n_missing_probe": 0,
+     "tau": null,
+     "mode": "freeze_initial",
+     "gate_pct": 0.0,
+     "baseline_acc": 0.62,
+     "prevented": 0,
+     "lost": 0,
+     "other_net": 0,
+     "net": 0,
+     "dacc_pp": 0.0,
+     "ci95": [
+      0.0,
+      0.0
+     ]
+    }
+   },
+   "pooled": {
+    "n": 6525,
+    "prevented": 0,
+    "lost": 0,
+    "net": 0,
+    "dacc_pp": 0.0,
+    "ci95": [
+     0.0,
+     0.0
+    ],
+    "weight_ladder_pp": {
+     "0.5": 0.0,
+     "1": 0.0,
+     "2": 0.0,
+     "4": 0.0,
+     "8": 0.0
+    },
+    "break_even_r": null
+   }
+  }
+ },
+ "none_mode": "paper",
+ "probes_dir": "abc_exp/results/gate_rerun"
+}

abc_exp/results/gate_rerun/gate_tex.txt

@@ -0,0 +1,25 @@
+% tab:pilot_gated: e5_rule per held-out model, then pooled rows of the main rule and the comparisons
+Llama-3.1-8B (86 col / 98 corr) & 0.10 & freeze_initial & 0.0\% & 0/86 & 0/98 & $+0$ & $+0.00$ [+0.00,+0.00] & $+0.15$ \\
+Phi-4-mini (112 col / 174 corr) & 0.10 & freeze_initial & 0.0\% & 0/112 & 0/174 & $+0$ & $+0.00$ [+0.00,+0.00] & $+0.09$ \\
+Qwen3-4B (50 col / 461 corr) & 0.10 & freeze_initial & 0.0\% & 0/50 & 0/461 & $+0$ & $+0.00$ [+0.00,+0.00] & $+0.00$ \\
+Qwen3-8B (3 col / 71 corr) & 0.10 & freeze_initial & 0.0\% & 0/3 & 0/71 & $+0$ & $+0.00$ [+0.00,+0.00] & $+0.00$ \\
+\midrule
+\textbf{Pooled (frozen LOMO policy)} & & & & 0/251 & 0/804 & $+0$ & $+0.00$ [+0.00,+0.00] & n/a \\
+\textbf{Nested LOMO selection} & & & & 6/251 & 26/804 & $-20$ & $-0.29$ [-0.46,-0.12] & n/a \\
+\textbf{Raw mean-FR threshold} & & & & 97/251 & 280/804 & $-183$ & $-2.80$ [-3.41,-2.24] & n/a \\
+\textbf{Always freeze} & & & & 251/251 & 804/804 & $-553$ & $-8.48$ [-9.48,-7.48] & n/a \\
+
+% tab:weight_sensitivity rows (r = 0.5, 1, 2, 4, 8, breakeven)
+Probe-gated freeze & $+0.00$ & $+0.00$ & $+0.00$ & $+0.00$ & $+0.00$ & n/a \\
+Always freeze & $-10.40$ & $-8.48$ & $-4.63$ & $+3.07$ & $+18.45$ & 3.20 \\
+Raw mean-FR threshold & $-3.55$ & $-2.80$ & $-1.32$ & $+1.66$ & $+7.60$ & 2.89 \\
+
+% tab:control_summary rows
+Always freeze & $6{,}525$ OSS & none & $-8.48$pp & $251$ & $804$ & $-553$ & $3.20$ \\
+Probe-gated freeze & $6{,}525$ OSS & LOMO & $+0.00$pp & $0$ & $0$ & $+0$ & $n/a$ \\
+
+% Unparsed-reply bounds: pooled prevented / lost / net per variant
+% e5_rule: paper 0/0/+0 (acc +0); flip 0/0/+0 (acc +0); drop 0/0/+0 (acc +0)
+% nested_lomo: paper 6/26/-20 (acc -19); flip 7/31/-24 (acc -23); drop 8/41/-33 (acc -32)
+% raw_fr_gate: paper 97/280/-183 (acc -183); flip 115/309/-194 (acc -194); drop 108/297/-189 (acc -189)
+% always_freeze: paper 251/804/-553 (acc -553); flip 251/804/-553 (acc -553); drop 251/804/-553 (acc -553)

abc_exp/results/gate_rerun/qids_llama-3.1-8b_s0.json

@@ -0,0 +1 @@
+["mmlu_pro_10", "mmlu_pro_19", "mmlu_pro_20", "mmlu_pro_27", "mmlu_pro_29", "mmlu_pro_39", "mmlu_pro_48", "mmlu_pro_49", "mmlu_pro_52", "mmlu_pro_60", "mmlu_pro_73", "mmlu_pro_102", "mmlu_pro_88", "mmlu_pro_107", "mmlu_pro_112", "mmlu_pro_123", "mmlu_pro_117", "mmlu_pro_142", "mmlu_pro_154", "mmlu_pro_157", "mmlu_pro_167", "mmlu_pro_165", "mmlu_pro_169", "mmlu_pro_171", "mmlu_pro_177", "mmlu_pro_180", "mmlu_pro_201", "mmlu_pro_181", "mmlu_pro_211", "mmlu_pro_213", "mmlu_pro_220", "mmlu_pro_233", "mmlu_pro_222", "mmlu_pro_234", "mmlu_pro_238", "mmlu_pro_240", "mmlu_pro_245", "mmlu_pro_256", "mmlu_pro_257", "mmlu_pro_261", "mmlu_pro_254", "mmlu_pro_264", "mmlu_pro_269", "mmlu_pro_267", "mmlu_pro_273", "mmlu_pro_279", "mmlu_pro_275", "mmlu_pro_291", "mmlu_pro_293", "mmlu_pro_295", "mmlu_pro_296", "mmlu_pro_303", "mmlu_pro_308", "mmlu_pro_299", "mmlu_pro_309", "mmlu_pro_318", "mmlu_pro_321", "mmlu_pro_320", "mmlu_pro_328", "mmlu_pro_334", "mmlu_pro_325", "mmlu_pro_364", "mmlu_pro_344", "mmlu_pro_369", "mmlu_pro_370", "mmlu_pro_377", "mmlu_pro_372", "mmlu_pro_382", "mmlu_pro_412", "mmlu_pro_384", "mmlu_pro_413", "mmlu_pro_416", "mmlu_pro_420", "mmlu_pro_425", "mmlu_pro_430", "mmlu_pro_431", "mmlu_pro_436", "mmlu_pro_440", "mmlu_pro_458", "mmlu_pro_451", "mmlu_pro_459", "mmlu_pro_476", "mmlu_pro_480", "mmlu_pro_491", "mmlu_pro_498", "mmlu_pro_485", "mmlu_pro_517", "mmlu_pro_521", "mmlu_pro_522", "mmlu_pro_526", "mmlu_pro_539", "mmlu_pro_542", "mmlu_pro_530", "mmlu_pro_552", "mmlu_pro_554", "mmlu_pro_556", "mmlu_pro_558", "mmlu_pro_577", "mmlu_pro_590", "mmlu_pro_602", "mmlu_pro_596", "mmlu_pro_610", "mmlu_pro_612", "mmlu_pro_604", "mmlu_pro_615", "mmlu_pro_627", "mmlu_pro_631", "mmlu_pro_633", "mmlu_pro_646", "mmlu_pro_658", "mmlu_pro_657", "mmlu_pro_673", "mmlu_pro_661", "mmlu_pro_678", "mmlu_pro_683", "mmlu_pro_686", "mmlu_pro_693", "mmlu_pro_703", "mmlu_pro_716", "mmlu_pro_712", "mmlu_pro_726", "mmlu_pro_743", "mmlu_pro_753", "mmlu_pro_754", "mmlu_pro_755", "mmlu_pro_756", "mmlu_pro_758", "mmlu_pro_765", "mmlu_pro_766", "mmlu_pro_768", "mmlu_pro_791", "mmlu_pro_796", "mmlu_pro_797", "mmlu_pro_822", "mmlu_pro_833", "mmlu_pro_830", "mmlu_pro_835", "mmlu_pro_841", "mmlu_pro_852", "mmlu_pro_859", "mmlu_pro_863", "mmlu_pro_866", "mmlu_pro_883", "mmlu_pro_880", "mmlu_pro_886", "mmlu_pro_890", "mmlu_pro_895", "mmlu_pro_891", "mmlu_pro_896", "mmlu_pro_903", "mmlu_pro_913", "mmlu_pro_910", "mmlu_pro_919", "mmlu_pro_932", "mmlu_pro_943", "mmlu_pro_954", "mmlu_pro_972", "mmlu_pro_968", "mmlu_pro_978", "mmlu_pro_982", "mmlu_pro_996", "mmlu_pro_1001", "mmlu_pro_1011", "mmlu_pro_1014", "mmlu_pro_1027", "mmlu_pro_1029", "mmlu_pro_1038", "mmlu_pro_1034", "mmlu_pro_1040", "mmlu_pro_1055", "mmlu_pro_1061", "mmlu_pro_1070", "mmlu_pro_1081", "mmlu_pro_1082", "mmlu_pro_1092", "mmlu_pro_1105", "mmlu_pro_1112", "mmlu_pro_1113", "mmlu_pro_1117", "mmlu_pro_1123", "mmlu_pro_1156", "mmlu_pro_1173", "mmlu_pro_1174", "mmlu_pro_1179", "mmlu_pro_1181", "mmlu_pro_1191", "mmlu_pro_1198", "mmlu_pro_1199", "mmlu_pro_1209", "mmlu_pro_1212", "mmlu_pro_1217", "mmlu_pro_1221", "mmlu_pro_1227", "mmlu_pro_1233", "mmlu_pro_1241", "mmlu_pro_1243", "mmlu_pro_1247", "mmlu_pro_1252", "mmlu_pro_1255", "mmlu_pro_1261", "mmlu_pro_1263", "mmlu_pro_1265", "mmlu_pro_1268", "mmlu_pro_1278", "mmlu_pro_1299", "mmlu_pro_1305", "mmlu_pro_1306", "mmlu_pro_1308", "mmlu_pro_1326", "mmlu_pro_1327", "mmlu_pro_1339", "mmlu_pro_1341", "mmlu_pro_1351", "mmlu_pro_1353", "mmlu_pro_1355", "mmlu_pro_1358", "mmlu_pro_1360", "mmlu_pro_1362", "mmlu_pro_1365", "mmlu_pro_1367", "mmlu_pro_1378", "mmlu_pro_1380", "mmlu_pro_1383", "mmlu_pro_1388", "mmlu_pro_1390", "mmlu_pro_1396", "mmlu_pro_1399", "mmlu_pro_1402", "mmlu_pro_1410", "mmlu_pro_1413", "mmlu_pro_1414", "mmlu_pro_1416", "mmlu_pro_1418", "mmlu_pro_1429", "mmlu_pro_1430", "mmlu_pro_1433", "mmlu_pro_1436", "mmlu_pro_1442", "mmlu_pro_1451", "mmlu_pro_1453", "mmlu_pro_1462", "mmlu_pro_1469", "mmlu_pro_1479", "mmlu_pro_1481", "mmlu_pro_1482", "mmlu_pro_1485", "mmlu_pro_1500", "mmlu_pro_1502", "mmlu_pro_1509", "mmlu_pro_1511", "mmlu_pro_1515", "mmlu_pro_1519", "mmlu_pro_1522", "mmlu_pro_1534", "mmlu_pro_1541", "mmlu_pro_1543", "mmlu_pro_1544", "mmlu_pro_1547", "mmlu_pro_1548", "mmlu_pro_1563", "mmlu_pro_1567", "mmlu_pro_1568", "mmlu_pro_1576", "mmlu_pro_1589", "mmlu_pro_1590", "mmlu_pro_1606", "mmlu_pro_1612", "mmlu_pro_1615", "mmlu_pro_1618", "mmlu_pro_1626", "mmlu_pro_1628", "mmlu_pro_1639", "mmlu_pro_1646", "mmlu_pro_1648", "mmlu_pro_1649", "mmlu_pro_1652", "mmlu_pro_1661", "mmlu_pro_1663", "mmlu_pro_1664", "mmlu_pro_1676", "mmlu_pro_1681", "mmlu_pro_1689", "mmlu_pro_1694", "mmlu_pro_1710", "mmlu_pro_1715", "mmlu_pro_1716", "mmlu_pro_1721", "mmlu_pro_1727", "mmlu_pro_1729", "mmlu_pro_1730", "mmlu_pro_1733", "mmlu_pro_1750", "mmlu_pro_1766", "mmlu_pro_1772", "mmlu_pro_1774", "mmlu_pro_1776", "mmlu_pro_1793", "mmlu_pro_1801", "mmlu_pro_1815", "mmlu_pro_1838", "mmlu_pro_1847", "mmlu_pro_1855", "mmlu_pro_1863", "mmlu_pro_1865", "mmlu_pro_1869", "mmlu_pro_1879", "mmlu_pro_1891", "mmlu_pro_1897", "mmlu_pro_1905", "mmlu_pro_1907", "mmlu_pro_1911", "mmlu_pro_1916", "mmlu_pro_1917", "mmlu_pro_1918", "mmlu_pro_1938", "mmlu_pro_1927", "mmlu_pro_1939", "mmlu_pro_1949", "mmlu_pro_1959", "mmlu_pro_1954", "mmlu_pro_1969", "mmlu_pro_1971", "mmlu_pro_1980", "mmlu_pro_1983", "mmlu_pro_1991", "mmlu_pro_1994", "mmlu_pro_1999", "mmlu_pro_2006", "mmlu_pro_2011", "mmlu_pro_2014", "mmlu_pro_2017", "mmlu_pro_2022", "mmlu_pro_2023", "mmlu_pro_2026", "mmlu_pro_2051", "mmlu_pro_2054", "mmlu_pro_2057", "mmlu_pro_2066", "mmlu_pro_2068", "mmlu_pro_2073", "mmlu_pro_2076", "mmlu_pro_2078", "mmlu_pro_2081", "mmlu_pro_2105", "mmlu_pro_2114", "mmlu_pro_2107", "mmlu_pro_2117", "mmlu_pro_2119", "mmlu_pro_2144", "mmlu_pro_2156", "mmlu_pro_2160", "mmlu_pro_2165", "mmlu_pro_2177", "mmlu_pro_2167", "mmlu_pro_2178", "mmlu_pro_2181", "mmlu_pro_2184", "mmlu_pro_2191", "mmlu_pro_2193", "mmlu_pro_2199", "mmlu_pro_2208", "mmlu_pro_2227", "mmlu_pro_2238", "mmlu_pro_2240", "mmlu_pro_2241", "mmlu_pro_2242", "mmlu_pro_2260", "mmlu_pro_2274", "mmlu_pro_2276", "mmlu_pro_2303", "mmlu_pro_2306", "mmlu_pro_2313", "mmlu_pro_2315", "mmlu_pro_2334", "mmlu_pro_2337", "mmlu_pro_2338", "mmlu_pro_2340", "mmlu_pro_2352", "mmlu_pro_2360", "mmlu_pro_2362", "mmlu_pro_2365", "mmlu_pro_2366", "mmlu_pro_2381", "mmlu_pro_2388", "mmlu_pro_2390", "mmlu_pro_2395", "mmlu_pro_2407", "mmlu_pro_2424", "mmlu_pro_2423", "mmlu_pro_2429", "mmlu_pro_2440", "mmlu_pro_2458", "mmlu_pro_2468", "mmlu_pro_2472", "mmlu_pro_2480", "mmlu_pro_2481", "mmlu_pro_2487", "mmlu_pro_2490", "mmlu_pro_2492", "mmlu_pro_2496", "mmlu_pro_2497", "mmlu_pro_2498", "mmlu_pro_2511", "mmlu_pro_2516", "mmlu_pro_2521", "mmlu_pro_2522", "mmlu_pro_2535", "mmlu_pro_2539", "mmlu_pro_2540", "mmlu_pro_2544", "mmlu_pro_2545", "mmlu_pro_2553", "mmlu_pro_2556", "mmlu_pro_2567", "mmlu_pro_2574", "mmlu_pro_2577", "mmlu_pro_2579", "mmlu_pro_2582", "mmlu_pro_2586", "mmlu_pro_2588", "mmlu_pro_2594", "mmlu_pro_2596", "mmlu_pro_2598", "mmlu_pro_2606", "mmlu_pro_2610", "mmlu_pro_2621", "mmlu_pro_2627", "mmlu_pro_2630", "mmlu_pro_2633", "mmlu_pro_2634", "mmlu_pro_2638", "mmlu_pro_2652", "mmlu_pro_2651", "mmlu_pro_2654", "mmlu_pro_2665", "mmlu_pro_2666", "mmlu_pro_2672", "mmlu_pro_2676", "mmlu_pro_2680", "mmlu_pro_2694", "mmlu_pro_2692", "mmlu_pro_2706", "mmlu_pro_2701", "mmlu_pro_2709", "mmlu_pro_2718", "mmlu_pro_2727", "mmlu_pro_2720", "mmlu_pro_2732", "mmlu_pro_2734", "mmlu_pro_2747", "mmlu_pro_2753", "mmlu_pro_2757", "mmlu_pro_2761", "mmlu_pro_2770", "mmlu_pro_2786", "mmlu_pro_2790", "mmlu_pro_2803", "mmlu_pro_2802", "mmlu_pro_2811", "mmlu_pro_2815", "mmlu_pro_2832", "mmlu_pro_2831", "mmlu_pro_2833", "mmlu_pro_2837", "mmlu_pro_2835", "mmlu_pro_2841", "mmlu_pro_2845", "mmlu_pro_2848", "mmlu_pro_2858", "mmlu_pro_2860", "mmlu_pro_2861", "mmlu_pro_2864", "mmlu_pro_2876", "mmlu_pro_2898", "mmlu_pro_2906", "mmlu_pro_2914", "mmlu_pro_2918", "mmlu_pro_2929", "mmlu_pro_2939", "mmlu_pro_2930", "mmlu_pro_2953", "mmlu_pro_2965", "mmlu_pro_2970", "mmlu_pro_2980", "mmlu_pro_2985", "mmlu_pro_2986", "mmlu_pro_2987", "mmlu_pro_2988", "mmlu_pro_3019", "mmlu_pro_3022", "mmlu_pro_3028", "mmlu_pro_3031", "mmlu_pro_3037", "mmlu_pro_3041", "mmlu_pro_3043", "mmlu_pro_3046", "mmlu_pro_3049", "mmlu_pro_3069", "mmlu_pro_3070", "mmlu_pro_3072", "mmlu_pro_3074", "mmlu_pro_3086", "mmlu_pro_3087", "mmlu_pro_3089", "mmlu_pro_3091", "mmlu_pro_3092", "mmlu_pro_3108", "mmlu_pro_3101", "mmlu_pro_3114", "mmlu_pro_3118", "mmlu_pro_3120", "mmlu_pro_3126", "mmlu_pro_3144", "mmlu_pro_3159", "mmlu_pro_3161", "mmlu_pro_3162", "mmlu_pro_3171", "mmlu_pro_3196", "mmlu_pro_3177", "mmlu_pro_3211", "mmlu_pro_3218", "mmlu_pro_3219", "mmlu_pro_3230", "mmlu_pro_3253", "mmlu_pro_3259", "mmlu_pro_3265", "mmlu_pro_3269", "mmlu_pro_3270", "mmlu_pro_3273", "mmlu_pro_3275", "mmlu_pro_3289", "mmlu_pro_3281", "mmlu_pro_3299", "mmlu_pro_3303", "mmlu_pro_3306", "mmlu_pro_3305", "mmlu_pro_3310", "mmlu_pro_3311", "mmlu_pro_3315", "mmlu_pro_3319", "mmlu_pro_3320", "mmlu_pro_3331", "mmlu_pro_3338", "mmlu_pro_3336", "mmlu_pro_3339", "mmlu_pro_3340", "mmlu_pro_3352", "mmlu_pro_3354", "mmlu_pro_3358", "mmlu_pro_3362", "mmlu_pro_3364", "mmlu_pro_3363", "mmlu_pro_3369", "mmlu_pro_3370", "mmlu_pro_3375", "mmlu_pro_3378", "mmlu_pro_3402", "mmlu_pro_3404", "mmlu_pro_3409", "mmlu_pro_3410", "mmlu_pro_3414", "mmlu_pro_3419", "mmlu_pro_3417", "mmlu_pro_3421", "mmlu_pro_3433", "mmlu_pro_3444", "mmlu_pro_3434", "mmlu_pro_3446", "mmlu_pro_3453", "mmlu_pro_3459", "mmlu_pro_3461", "mmlu_pro_3460", "mmlu_pro_3465", "mmlu_pro_3470", "mmlu_pro_3479", "mmlu_pro_3499", "mmlu_pro_3489", "mmlu_pro_3511", "mmlu_pro_3522", "mmlu_pro_3523", "mmlu_pro_3534", "mmlu_pro_3538", "mmlu_pro_3539", "mmlu_pro_3543", "mmlu_pro_3558", "mmlu_pro_3570", "mmlu_pro_3562", "mmlu_pro_3572", "mmlu_pro_3579", "mmlu_pro_3580", "mmlu_pro_3584", "mmlu_pro_3587", "mmlu_pro_3598", "mmlu_pro_3599", "mmlu_pro_3600", "mmlu_pro_3612", "mmlu_pro_3610", "mmlu_pro_3630", "mmlu_pro_3626", "mmlu_pro_3637", "mmlu_pro_3644", "mmlu_pro_3647", "mmlu_pro_3646", "mmlu_pro_3670", "mmlu_pro_3678", "mmlu_pro_3681", "mmlu_pro_3680", "mmlu_pro_3687", "mmlu_pro_3695", "mmlu_pro_3696", "mmlu_pro_3703", "mmlu_pro_3720", "mmlu_pro_3711", "mmlu_pro_3725", "mmlu_pro_3722", "mmlu_pro_3729", "mmlu_pro_3736", "mmlu_pro_3737", "mmlu_pro_3739", "mmlu_pro_3740", "mmlu_pro_3757", "mmlu_pro_3758", "mmlu_pro_3759", "mmlu_pro_3768", "mmlu_pro_3764", "mmlu_pro_3777", "mmlu_pro_3786", "mmlu_pro_3805", "mmlu_pro_3791", "mmlu_pro_3807", "mmlu_pro_3817", "mmlu_pro_3815", "mmlu_pro_3819", "mmlu_pro_3822", "mmlu_pro_3826", "mmlu_pro_3828", "mmlu_pro_3832", "mmlu_pro_3842", "mmlu_pro_3835", "mmlu_pro_3855", "mmlu_pro_3856", "mmlu_pro_3863", "mmlu_pro_3864", "mmlu_pro_3871", "mmlu_pro_3867", "mmlu_pro_3902", "mmlu_pro_3895", "mmlu_pro_3903", "mmlu_pro_3909", "mmlu_pro_3910", "mmlu_pro_3912", "mmlu_pro_3913", "mmlu_pro_3924", "mmlu_pro_3936", "mmlu_pro_3954", "mmlu_pro_3955", "mmlu_pro_3964", "mmlu_pro_3976", "mmlu_pro_3979", "mmlu_pro_3985", "mmlu_pro_3981", "mmlu_pro_3994", "mmlu_pro_3991", "mmlu_pro_4004", "mmlu_pro_4006", "mmlu_pro_4009", "mmlu_pro_4018", "mmlu_pro_4032", "mmlu_pro_4029", "mmlu_pro_4038", "mmlu_pro_4037", "mmlu_pro_4040", "mmlu_pro_4045", "mmlu_pro_4057", "mmlu_pro_4063", "mmlu_pro_4064", "mmlu_pro_4089", "mmlu_pro_4105", "mmlu_pro_4117", "mmlu_pro_4109", "mmlu_pro_4139", "mmlu_pro_4143", "mmlu_pro_4146", "mmlu_pro_4144", "mmlu_pro_4158", "mmlu_pro_4148", "mmlu_pro_4163", "mmlu_pro_4165", "mmlu_pro_4181", "mmlu_pro_4198", "mmlu_pro_4204", "mmlu_pro_4207", "mmlu_pro_4205", "mmlu_pro_4210", "mmlu_pro_4215", "mmlu_pro_4217", "mmlu_pro_4226", "mmlu_pro_4228", "mmlu_pro_4232", "mmlu_pro_4234", "mmlu_pro_4242", "mmlu_pro_4243", "mmlu_pro_4246", "mmlu_pro_4253", "mmlu_pro_4261", "mmlu_pro_4258", "mmlu_pro_4269", "mmlu_pro_4270", "mmlu_pro_4279", "mmlu_pro_4295", "mmlu_pro_4287", "mmlu_pro_4296", "mmlu_pro_4302", "mmlu_pro_4297", "mmlu_pro_4340", "mmlu_pro_4333", "mmlu_pro_4343", "mmlu_pro_4371", "mmlu_pro_4350", "mmlu_pro_4399", "mmlu_pro_4392", "mmlu_pro_4406", "mmlu_pro_4410", "mmlu_pro_4411", "mmlu_pro_4433", "mmlu_pro_4434", "mmlu_pro_4459", "mmlu_pro_4439", "mmlu_pro_4468", "mmlu_pro_4479", "mmlu_pro_4489", "mmlu_pro_4507", "mmlu_pro_4509", "mmlu_pro_4508", "mmlu_pro_4518", "mmlu_pro_4512", "mmlu_pro_4519", "mmlu_pro_4522", "mmlu_pro_4524", "mmlu_pro_4531", "mmlu_pro_4528", "mmlu_pro_4534", "mmlu_pro_4537", "mmlu_pro_4536", "mmlu_pro_4544", "mmlu_pro_4551", "mmlu_pro_4562", "mmlu_pro_4557", "mmlu_pro_4574", "mmlu_pro_4583", "mmlu_pro_4586", "mmlu_pro_4589", "mmlu_pro_4598", "mmlu_pro_4597", "mmlu_pro_4606", "mmlu_pro_4616", "mmlu_pro_4621", "mmlu_pro_4617", "mmlu_pro_4636", "mmlu_pro_4630", "mmlu_pro_4639", "mmlu_pro_4637", "mmlu_pro_4640", "mmlu_pro_4643", "mmlu_pro_4660", "mmlu_pro_4661", "mmlu_pro_4670", "mmlu_pro_4674", "mmlu_pro_4681", "mmlu_pro_4684", "mmlu_pro_4696", "mmlu_pro_4698", "mmlu_pro_4703", "mmlu_pro_4704", "mmlu_pro_4720", "mmlu_pro_4709", "mmlu_pro_4723", "mmlu_pro_4724", "mmlu_pro_4736", "mmlu_pro_4744", "mmlu_pro_4753", "mmlu_pro_4760", "mmlu_pro_4763", "mmlu_pro_4771", "mmlu_pro_4776", "mmlu_pro_4786", "mmlu_pro_4787", "mmlu_pro_4794", "mmlu_pro_4792", "mmlu_pro_4795", "mmlu_pro_4800", "mmlu_pro_4814", "mmlu_pro_4807", "mmlu_pro_4820", "mmlu_pro_4829", "mmlu_pro_4836", "mmlu_pro_4842", "mmlu_pro_4877", "mmlu_pro_4878", "mmlu_pro_4884", "mmlu_pro_4893", "mmlu_pro_4897", "mmlu_pro_4899", "mmlu_pro_4904", "mmlu_pro_4916", "mmlu_pro_4924", "mmlu_pro_4938", "mmlu_pro_4939", "mmlu_pro_4949", "mmlu_pro_4955", "mmlu_pro_4956", "mmlu_pro_4957", "mmlu_pro_4960", "mmlu_pro_4961", "mmlu_pro_4969", "mmlu_pro_4986", "mmlu_pro_4992", "mmlu_pro_5023", "mmlu_pro_5016", "mmlu_pro_5038", "mmlu_pro_5046", "mmlu_pro_5058", "mmlu_pro_5050", "mmlu_pro_5061", "mmlu_pro_5072", "mmlu_pro_5074", "mmlu_pro_5078", "mmlu_pro_5085", "mmlu_pro_5080", "mmlu_pro_5086", "mmlu_pro_5088", "mmlu_pro_5091", "mmlu_pro_5093", "mmlu_pro_5095", "mmlu_pro_5103", "mmlu_pro_5108", "mmlu_pro_5114", "mmlu_pro_5121", "mmlu_pro_5128", "mmlu_pro_5130", "mmlu_pro_5134", "mmlu_pro_5140", "mmlu_pro_5149", "mmlu_pro_5153", "mmlu_pro_5157", "mmlu_pro_5158", "mmlu_pro_5159", "mmlu_pro_5167", "mmlu_pro_5172", "mmlu_pro_5191", "mmlu_pro_5193", "mmlu_pro_5214", "mmlu_pro_5212", "mmlu_pro_5225", "mmlu_pro_5226", "mmlu_pro_5229", "mmlu_pro_5231", "mmlu_pro_5245", "mmlu_pro_5237", "mmlu_pro_5248", "mmlu_pro_5246", "mmlu_pro_5256", "mmlu_pro_5262", "mmlu_pro_5263", "mmlu_pro_5264", "mmlu_pro_5267", "mmlu_pro_5270", "mmlu_pro_5279", "mmlu_pro_5281", "mmlu_pro_5294", "mmlu_pro_5300", "mmlu_pro_5311", "mmlu_pro_5305", "mmlu_pro_5324", "mmlu_pro_5319", "mmlu_pro_5331", "mmlu_pro_5328", "mmlu_pro_5333", "mmlu_pro_5339", "mmlu_pro_5348", "mmlu_pro_5369", "mmlu_pro_5374", "mmlu_pro_5381", "mmlu_pro_5387", "mmlu_pro_5393", "mmlu_pro_5396", "mmlu_pro_5423", "mmlu_pro_5421", "mmlu_pro_5431", "mmlu_pro_5443", "mmlu_pro_5445", "mmlu_pro_5447", "mmlu_pro_5450", "mmlu_pro_5456", "mmlu_pro_5459"]

abc_exp/results/gate_rerun/qids_llama-3.1-8b_s1.json

@@ -0,0 +1 @@
+["mmlu_pro_5465", "mmlu_pro_5473", "mmlu_pro_5500", "mmlu_pro_5482", "mmlu_pro_5501", "mmlu_pro_5502", "mmlu_pro_5507", "mmlu_pro_5512", "mmlu_pro_5519", "mmlu_pro_5522", "mmlu_pro_5523", "mmlu_pro_5540", "mmlu_pro_5541", "mmlu_pro_5542", "mmlu_pro_5554", "mmlu_pro_5557", "mmlu_pro_5563", "mmlu_pro_5561", "mmlu_pro_5565", "mmlu_pro_5588", "mmlu_pro_5571", "mmlu_pro_5598", "mmlu_pro_5600", "mmlu_pro_5607", "mmlu_pro_5610", "mmlu_pro_5613", "mmlu_pro_5616", "mmlu_pro_5623", "mmlu_pro_5626", "mmlu_pro_5638", "mmlu_pro_5630", "mmlu_pro_5639", "mmlu_pro_5647", "mmlu_pro_5661", "mmlu_pro_5662", "mmlu_pro_5669", "mmlu_pro_5672", "mmlu_pro_5677", "mmlu_pro_5695", "mmlu_pro_5696", "mmlu_pro_5705", "mmlu_pro_5709", "mmlu_pro_5711", "mmlu_pro_5714", "mmlu_pro_5716", "mmlu_pro_5721", "mmlu_pro_5730", "mmlu_pro_5737", "mmlu_pro_5738", "mmlu_pro_5746", "mmlu_pro_5760", "mmlu_pro_5763", "mmlu_pro_5778", "mmlu_pro_5780", "mmlu_pro_5784", "mmlu_pro_5791", "mmlu_pro_5812", "mmlu_pro_5815", "mmlu_pro_5817", "mmlu_pro_5818", "mmlu_pro_5829", "mmlu_pro_5823", "mmlu_pro_5836", "mmlu_pro_5853", "mmlu_pro_5854", "mmlu_pro_5855", "mmlu_pro_5860", "mmlu_pro_5876", "mmlu_pro_5879", "mmlu_pro_5888", "mmlu_pro_5891", "mmlu_pro_5899", "mmlu_pro_5905", "mmlu_pro_5902", "mmlu_pro_5911", "mmlu_pro_5912", "mmlu_pro_5927", "mmlu_pro_5931", "mmlu_pro_5932", "mmlu_pro_5947", "mmlu_pro_5955", "mmlu_pro_5959", "mmlu_pro_5960", "mmlu_pro_5965", "mmlu_pro_5966", "mmlu_pro_5967", "mmlu_pro_5971", "mmlu_pro_5981", "mmlu_pro_5982", "mmlu_pro_5985", "mmlu_pro_5997", "mmlu_pro_5998", "mmlu_pro_6007", "mmlu_pro_6009", "mmlu_pro_6012", "mmlu_pro_6021", "mmlu_pro_6020", "mmlu_pro_6029", "mmlu_pro_6027", "mmlu_pro_6033", "mmlu_pro_6032", "mmlu_pro_6034", "mmlu_pro_6038", "mmlu_pro_6052", "mmlu_pro_6054", "mmlu_pro_6055", "mmlu_pro_6062", "mmlu_pro_6073", "mmlu_pro_6074", "mmlu_pro_6079", "mmlu_pro_6084", "mmlu_pro_6088", "mmlu_pro_6094", "mmlu_pro_6095", "mmlu_pro_6120", "mmlu_pro_6114", "mmlu_pro_6130", "mmlu_pro_6134", "mmlu_pro_6140", "mmlu_pro_6144", "mmlu_pro_6160", "mmlu_pro_6152", "mmlu_pro_6162", "mmlu_pro_6171", "mmlu_pro_6176", "mmlu_pro_6177", "mmlu_pro_6188", "mmlu_pro_6194", "mmlu_pro_6204", "mmlu_pro_6206", "mmlu_pro_6212", "mmlu_pro_6215", "mmlu_pro_6217", "mmlu_pro_6219", "mmlu_pro_6221", "mmlu_pro_6229", "mmlu_pro_6237", "mmlu_pro_6249", "mmlu_pro_6252", "mmlu_pro_6269", "mmlu_pro_6263", "mmlu_pro_6299", "mmlu_pro_6291", "mmlu_pro_6302", "mmlu_pro_6306", "mmlu_pro_6310", "mmlu_pro_6313", "mmlu_pro_6314", "mmlu_pro_6317", "mmlu_pro_6318", "mmlu_pro_6332", "mmlu_pro_6336", "mmlu_pro_6341", "mmlu_pro_6344", "mmlu_pro_6347", "mmlu_pro_6354", "mmlu_pro_6367", "mmlu_pro_6372", "mmlu_pro_6379", "mmlu_pro_6382", "mmlu_pro_6389", "mmlu_pro_6413", "mmlu_pro_6443", "mmlu_pro_6435", "mmlu_pro_6464", "mmlu_pro_6452", "mmlu_pro_6471", "mmlu_pro_6482", "mmlu_pro_6485", "mmlu_pro_6487", "mmlu_pro_6488", "mmlu_pro_6490", "mmlu_pro_6491", "mmlu_pro_6494", "mmlu_pro_6495", "mmlu_pro_6499", "mmlu_pro_6504", "mmlu_pro_6523", "mmlu_pro_6524", "mmlu_pro_6537", "mmlu_pro_6539", "mmlu_pro_6542", "mmlu_pro_6546", "mmlu_pro_6547", "mmlu_pro_6554", "mmlu_pro_6567", "mmlu_pro_6565", "mmlu_pro_6576", "mmlu_pro_6586", "mmlu_pro_6592", "mmlu_pro_6596", "mmlu_pro_6609", "mmlu_pro_6624", "mmlu_pro_6625", "mmlu_pro_6646", "mmlu_pro_6658", "mmlu_pro_6674", "mmlu_pro_6682", "mmlu_pro_6679", "mmlu_pro_6700", "mmlu_pro_6694", "mmlu_pro_6720", "mmlu_pro_6722", "mmlu_pro_6726", "mmlu_pro_6729", "mmlu_pro_6733", "mmlu_pro_6736", "mmlu_pro_6752", "mmlu_pro_6760", "mmlu_pro_6763", "mmlu_pro_6765", "mmlu_pro_6769", "mmlu_pro_6770", "mmlu_pro_6771", "mmlu_pro_6775", "mmlu_pro_6780", "mmlu_pro_6786", "mmlu_pro_6802", "mmlu_pro_6804", "mmlu_pro_6807", "mmlu_pro_6817", "mmlu_pro_6820", "mmlu_pro_6829", "mmlu_pro_6845", "mmlu_pro_6849", "mmlu_pro_6852", "mmlu_pro_6861", "mmlu_pro_6865", "mmlu_pro_6872", "mmlu_pro_6879", "mmlu_pro_6881", "mmlu_pro_6885", "mmlu_pro_6886", "mmlu_pro_6888", "mmlu_pro_6891", "mmlu_pro_6894", "mmlu_pro_6892", "mmlu_pro_6907", "mmlu_pro_6920", "mmlu_pro_6921", "mmlu_pro_6923", "mmlu_pro_6928", "mmlu_pro_6933", "mmlu_pro_6939", "mmlu_pro_6942", "mmlu_pro_6964", "mmlu_pro_6970", "mmlu_pro_6975", "mmlu_pro_6973", "mmlu_pro_6994", "mmlu_pro_6995", "mmlu_pro_7001", "mmlu_pro_7007", "mmlu_pro_7002", "mmlu_pro_7008", "mmlu_pro_7011", "mmlu_pro_7029", "mmlu_pro_7043", "mmlu_pro_7049", "mmlu_pro_7050", "mmlu_pro_7051", "mmlu_pro_7058", "mmlu_pro_7076", "mmlu_pro_7085", "mmlu_pro_7095", "mmlu_pro_7090", "mmlu_pro_7097", "mmlu_pro_7100", "mmlu_pro_7110", "mmlu_pro_7111", "mmlu_pro_7113", "mmlu_pro_7118", "mmlu_pro_7127", "mmlu_pro_7126", "mmlu_pro_7131", "mmlu_pro_7132", "mmlu_pro_7142", "mmlu_pro_7154", "mmlu_pro_7163", "mmlu_pro_7171", "mmlu_pro_7175", "mmlu_pro_7176", "mmlu_pro_7181", "mmlu_pro_7183", "mmlu_pro_7190", "mmlu_pro_7194", "mmlu_pro_7209", "mmlu_pro_7232", "mmlu_pro_7236", "mmlu_pro_7249", "mmlu_pro_7252", "mmlu_pro_7253", "mmlu_pro_7269", "mmlu_pro_7279", "mmlu_pro_7283", "mmlu_pro_7284", "mmlu_pro_7285", "mmlu_pro_7287", "mmlu_pro_7294", "mmlu_pro_7305", "mmlu_pro_7306", "mmlu_pro_7327", "mmlu_pro_7331", "mmlu_pro_7344", "mmlu_pro_7345", "mmlu_pro_7359", "mmlu_pro_7361", "mmlu_pro_7365", "mmlu_pro_7367", "mmlu_pro_7369", "mmlu_pro_7371", "mmlu_pro_7376", "mmlu_pro_7372", "mmlu_pro_7386", "mmlu_pro_7380", "mmlu_pro_7388", "mmlu_pro_7391", "mmlu_pro_7395", "mmlu_pro_7397", "mmlu_pro_7406", "mmlu_pro_7402", "mmlu_pro_7413", "mmlu_pro_7415", "mmlu_pro_7423", "mmlu_pro_7427", "mmlu_pro_7422", "mmlu_pro_7438", "mmlu_pro_7444", "mmlu_pro_7470", "mmlu_pro_7468", "mmlu_pro_7473", "mmlu_pro_7482", "mmlu_pro_7501", "mmlu_pro_7506", "mmlu_pro_7503", "mmlu_pro_7516", "mmlu_pro_7523", "mmlu_pro_7524", "mmlu_pro_7532", "mmlu_pro_7528", "mmlu_pro_7540", "mmlu_pro_7538", "mmlu_pro_7544", "mmlu_pro_7564", "mmlu_pro_7589", "mmlu_pro_7572", "mmlu_pro_7596", "mmlu_pro_7595", "mmlu_pro_7601", "mmlu_pro_7600", "mmlu_pro_7612", "mmlu_pro_7619", "mmlu_pro_7620", "mmlu_pro_7621", "mmlu_pro_7626", "mmlu_pro_7629", "mmlu_pro_7633", "mmlu_pro_7635", "mmlu_pro_7639", "mmlu_pro_7660", "mmlu_pro_7656", "mmlu_pro_7675", "mmlu_pro_7661", "mmlu_pro_7682", "mmlu_pro_7694", "mmlu_pro_7708", "mmlu_pro_7709", "mmlu_pro_7715", "mmlu_pro_7716", "mmlu_pro_7723", "mmlu_pro_7717", "mmlu_pro_7727", "mmlu_pro_7737", "mmlu_pro_7750", "mmlu_pro_7755", "mmlu_pro_7764", "mmlu_pro_7770", "mmlu_pro_7765", "mmlu_pro_7777", "mmlu_pro_7779", "mmlu_pro_7789", "mmlu_pro_7790", "mmlu_pro_7792", "mmlu_pro_7798", "mmlu_pro_7808", "mmlu_pro_7817", "mmlu_pro_7812", "mmlu_pro_7826", "mmlu_pro_7828", "mmlu_pro_7835", "mmlu_pro_7836", "mmlu_pro_7848", "mmlu_pro_7837", "mmlu_pro_7851", "mmlu_pro_7852", "mmlu_pro_7871", "mmlu_pro_7858", "mmlu_pro_7880", "mmlu_pro_7878", "mmlu_pro_7887", "mmlu_pro_7900", "mmlu_pro_7893", "mmlu_pro_7909", "mmlu_pro_7902", "mmlu_pro_7917", "mmlu_pro_7918", "mmlu_pro_7925", "mmlu_pro_7936", "mmlu_pro_7934", "mmlu_pro_7940", "mmlu_pro_7939", "mmlu_pro_7946", "mmlu_pro_7945", "mmlu_pro_7953", "mmlu_pro_7957", "mmlu_pro_7962", "mmlu_pro_7975", "mmlu_pro_7972", "mmlu_pro_7981", "mmlu_pro_7989", "mmlu_pro_7983", "mmlu_pro_7995", "mmlu_pro_7993", "mmlu_pro_7996", "mmlu_pro_8003", "mmlu_pro_8004", "mmlu_pro_8023", "mmlu_pro_8021", "mmlu_pro_8048", "mmlu_pro_8034", "mmlu_pro_8050", "mmlu_pro_8071", "mmlu_pro_8080", "mmlu_pro_8076", "mmlu_pro_8095", "mmlu_pro_8084", "mmlu_pro_8101", "mmlu_pro_8108", "mmlu_pro_8109", "mmlu_pro_8115", "mmlu_pro_8123", "mmlu_pro_8138", "mmlu_pro_8152", "mmlu_pro_8154", "mmlu_pro_8140", "mmlu_pro_8158", "mmlu_pro_8156", "mmlu_pro_8166", "mmlu_pro_8176", "mmlu_pro_8159", "mmlu_pro_8178", "mmlu_pro_8181", "mmlu_pro_8195", "mmlu_pro_8184", "mmlu_pro_8218", "mmlu_pro_8196", "mmlu_pro_8224", "mmlu_pro_8222", "mmlu_pro_8226", "mmlu_pro_8235", "mmlu_pro_8236", "mmlu_pro_8241", "mmlu_pro_8247", "mmlu_pro_8242", "mmlu_pro_8249", "mmlu_pro_8259", "mmlu_pro_8262", "mmlu_pro_8273", "mmlu_pro_8275", "mmlu_pro_8277", "mmlu_pro_8283", "mmlu_pro_8284", "mmlu_pro_8294", "mmlu_pro_8291", "mmlu_pro_8295", "mmlu_pro_8302", "mmlu_pro_8301", "mmlu_pro_8320", "mmlu_pro_8333", "mmlu_pro_8337", "mmlu_pro_8341", "mmlu_pro_8342", "mmlu_pro_8354", "mmlu_pro_8348", "mmlu_pro_8365", "mmlu_pro_8361", "mmlu_pro_8371", "mmlu_pro_8379", "mmlu_pro_8391", "mmlu_pro_8389", "mmlu_pro_8397", "mmlu_pro_8411", "mmlu_pro_8419", "mmlu_pro_8417", "mmlu_pro_8449", "mmlu_pro_8441", "mmlu_pro_8473", "mmlu_pro_8458", "mmlu_pro_8474", "mmlu_pro_8475", "mmlu_pro_8478", "mmlu_pro_8481", "mmlu_pro_8484", "mmlu_pro_8486", "mmlu_pro_8508", "mmlu_pro_8515", "mmlu_pro_8516", "mmlu_pro_8530", "mmlu_pro_8562", "mmlu_pro_8563", "mmlu_pro_8567", "mmlu_pro_8569", "mmlu_pro_8580", "mmlu_pro_8584", "mmlu_pro_8582", "mmlu_pro_8594", "mmlu_pro_8586", "mmlu_pro_8601", "mmlu_pro_8611", "mmlu_pro_8614", "mmlu_pro_8629", "mmlu_pro_8641", "mmlu_pro_8644", "mmlu_pro_8647", "mmlu_pro_8658", "mmlu_pro_8654", "mmlu_pro_8669", "mmlu_pro_8670", "mmlu_pro_8699", "mmlu_pro_8690", "mmlu_pro_8704", "mmlu_pro_8709", "mmlu_pro_8710", "mmlu_pro_8724", "mmlu_pro_8729", "mmlu_pro_8733", "mmlu_pro_8725", "mmlu_pro_8745", "mmlu_pro_8747", "mmlu_pro_8760", "mmlu_pro_8761", "mmlu_pro_8781", "mmlu_pro_8766", "mmlu_pro_8796", "mmlu_pro_8805", "mmlu_pro_8803", "mmlu_pro_8816", "mmlu_pro_8820", "mmlu_pro_8826", "mmlu_pro_8831", "mmlu_pro_8833", "mmlu_pro_8839", "mmlu_pro_8843", "mmlu_pro_8844", "mmlu_pro_8848", "mmlu_pro_8855", "mmlu_pro_8850", "mmlu_pro_8865", "mmlu_pro_8856", "mmlu_pro_8875", "mmlu_pro_8878", "mmlu_pro_8872", "mmlu_pro_8912", "mmlu_pro_8903", "mmlu_pro_8932", "mmlu_pro_8919", "mmlu_pro_8934", "mmlu_pro_8936", "mmlu_pro_8941", "mmlu_pro_8950", "mmlu_pro_8955", "mmlu_pro_8956", "mmlu_pro_8958", "mmlu_pro_8975", "mmlu_pro_8971", "mmlu_pro_8983", "mmlu_pro_9004", "mmlu_pro_9002", "mmlu_pro_9005", "mmlu_pro_9010", "mmlu_pro_9013", "mmlu_pro_9016", "mmlu_pro_9019", "mmlu_pro_9024", "mmlu_pro_9028", "mmlu_pro_9033", "mmlu_pro_9039", "mmlu_pro_9042", "mmlu_pro_9072", "mmlu_pro_9063", "mmlu_pro_9099", "mmlu_pro_9084", "mmlu_pro_9101", "mmlu_pro_9114", "mmlu_pro_9121", "mmlu_pro_9123", "mmlu_pro_9124", "mmlu_pro_9125", "mmlu_pro_9129", "mmlu_pro_9140", "mmlu_pro_9137", "mmlu_pro_9141", "mmlu_pro_9155", "mmlu_pro_9143", "mmlu_pro_9157", "mmlu_pro_9164", "mmlu_pro_9172", "mmlu_pro_9181", "mmlu_pro_9185", "mmlu_pro_9187", "mmlu_pro_9188", "mmlu_pro_9192", "mmlu_pro_9217", "mmlu_pro_9213", "mmlu_pro_9228", "mmlu_pro_9219", "mmlu_pro_9239", "mmlu_pro_9243", "mmlu_pro_9244", "mmlu_pro_9245", "mmlu_pro_9254", "mmlu_pro_9274", "mmlu_pro_9265", "mmlu_pro_9282", "mmlu_pro_9291", "mmlu_pro_9286", "mmlu_pro_9307", "mmlu_pro_9314", "mmlu_pro_9308", "mmlu_pro_9328", "mmlu_pro_9324", "mmlu_pro_9337", "mmlu_pro_9329", "mmlu_pro_9342", "mmlu_pro_9340", "mmlu_pro_9345", "mmlu_pro_9349", "mmlu_pro_9374", "mmlu_pro_9375", "mmlu_pro_9376", "mmlu_pro_9380", "mmlu_pro_9378", "mmlu_pro_9396", "mmlu_pro_9407", "mmlu_pro_9400", "mmlu_pro_9410", "mmlu_pro_9411", "mmlu_pro_9414", "mmlu_pro_9415", "mmlu_pro_9422", "mmlu_pro_9439", "mmlu_pro_9442", "mmlu_pro_9443", "mmlu_pro_9448", "mmlu_pro_9483", "mmlu_pro_9491", "mmlu_pro_9494", "mmlu_pro_9503", "mmlu_pro_9500", "mmlu_pro_9511", "mmlu_pro_9521", "mmlu_pro_9512", "mmlu_pro_9545", "mmlu_pro_9533", "mmlu_pro_9548", "mmlu_pro_9586", "mmlu_pro_9567", "mmlu_pro_9613", "mmlu_pro_9603", "mmlu_pro_9621", "mmlu_pro_9619", "mmlu_pro_9623", "mmlu_pro_9627", "mmlu_pro_9629", "mmlu_pro_9632", "mmlu_pro_9641", "mmlu_pro_9644", "mmlu_pro_9653", "mmlu_pro_9647", "mmlu_pro_9661", "mmlu_pro_9671", "mmlu_pro_9686", "mmlu_pro_9675", "mmlu_pro_9687", "mmlu_pro_9697", "mmlu_pro_9705", "mmlu_pro_9715", "mmlu_pro_9725", "mmlu_pro_9734", "mmlu_pro_9739", "mmlu_pro_9738", "mmlu_pro_9743", "mmlu_pro_9753", "mmlu_pro_9756", "mmlu_pro_9768", "mmlu_pro_9777", "mmlu_pro_9791", "mmlu_pro_9793", "mmlu_pro_9798", "mmlu_pro_9808", "mmlu_pro_9817", "mmlu_pro_9820", "mmlu_pro_9819", "mmlu_pro_9839", "mmlu_pro_9840", "mmlu_pro_9843", "mmlu_pro_9851", "mmlu_pro_9852", "mmlu_pro_9859", "mmlu_pro_9876", "mmlu_pro_9872", "mmlu_pro_9882", "mmlu_pro_9884", "mmlu_pro_9891", "mmlu_pro_9903", "mmlu_pro_9895", "mmlu_pro_9907", "mmlu_pro_9923", "mmlu_pro_9925", "mmlu_pro_9926", "mmlu_pro_9933", "mmlu_pro_9930", "mmlu_pro_9942", "mmlu_pro_9968", "mmlu_pro_9969", "mmlu_pro_9971", "mmlu_pro_9975", "mmlu_pro_9976", "mmlu_pro_9977", "mmlu_pro_9989", "mmlu_pro_9987", "mmlu_pro_9993", "mmlu_pro_9991", "mmlu_pro_10005", "mmlu_pro_10011", "mmlu_pro_10034", "mmlu_pro_10026", "mmlu_pro_10035", "mmlu_pro_10037", "mmlu_pro_10036", "mmlu_pro_10038", "mmlu_pro_10044", "mmlu_pro_10051", "mmlu_pro_10060", "mmlu_pro_10065", "mmlu_pro_10063", "mmlu_pro_10079", "mmlu_pro_10074", "mmlu_pro_10085", "mmlu_pro_10080", "mmlu_pro_10087", "mmlu_pro_10095", "mmlu_pro_10090", "mmlu_pro_10112", "mmlu_pro_10098", "mmlu_pro_10131", "mmlu_pro_10127", "mmlu_pro_10134", "mmlu_pro_10141", "mmlu_pro_10139", "mmlu_pro_10142", "mmlu_pro_10147", "mmlu_pro_10150", "mmlu_pro_10156", "mmlu_pro_10159", "mmlu_pro_10162", "mmlu_pro_10179", "mmlu_pro_10170", "mmlu_pro_10185", "mmlu_pro_10186", "mmlu_pro_10222", "mmlu_pro_10230", "mmlu_pro_10241", "mmlu_pro_10242", "mmlu_pro_10245", "mmlu_pro_10247", "mmlu_pro_10270", "mmlu_pro_10271", "mmlu_pro_10285", "mmlu_pro_10283", "mmlu_pro_10290", "mmlu_pro_10288", "mmlu_pro_10299", "mmlu_pro_10300", "mmlu_pro_10302", "mmlu_pro_10308", "mmlu_pro_10314", "mmlu_pro_10315", "mmlu_pro_10316", "mmlu_pro_10321", "mmlu_pro_10318", "mmlu_pro_10324", "mmlu_pro_10333", "mmlu_pro_10341", "mmlu_pro_10343", "mmlu_pro_10344", "mmlu_pro_10349", "mmlu_pro_10346", "mmlu_pro_10360", "mmlu_pro_10367", "mmlu_pro_10372", "mmlu_pro_10382", "mmlu_pro_10387", "mmlu_pro_10408", "mmlu_pro_10411", "mmlu_pro_10417", "mmlu_pro_10422", "mmlu_pro_10437", "mmlu_pro_10432", "mmlu_pro_10441", "mmlu_pro_10454", "mmlu_pro_10467", "mmlu_pro_10476", "mmlu_pro_10481", "mmlu_pro_10478", "mmlu_pro_10484", "mmlu_pro_10487", "mmlu_pro_10489", "mmlu_pro_10492", "mmlu_pro_10502", "mmlu_pro_10503", "mmlu_pro_10505", "mmlu_pro_10515", "mmlu_pro_10520", "mmlu_pro_10526", "mmlu_pro_10527", "mmlu_pro_10532", "mmlu_pro_10538", "mmlu_pro_10543", "mmlu_pro_10544", "mmlu_pro_10568", "mmlu_pro_10571", "mmlu_pro_10578", "mmlu_pro_10586", "mmlu_pro_10589", "mmlu_pro_10600", "mmlu_pro_10602", "mmlu_pro_10609", "mmlu_pro_10621", "mmlu_pro_10627", "mmlu_pro_10641", "mmlu_pro_10648", "mmlu_pro_10651", "mmlu_pro_10656", "mmlu_pro_10661", "mmlu_pro_10663", "mmlu_pro_10669", "mmlu_pro_10689", "mmlu_pro_10692", "mmlu_pro_10694", "mmlu_pro_10713", "mmlu_pro_10703", "mmlu_pro_10717", "mmlu_pro_10728", "mmlu_pro_10737", "mmlu_pro_10741", "mmlu_pro_10752", "mmlu_pro_10753", "mmlu_pro_10758", "mmlu_pro_10762", "mmlu_pro_10765", "mmlu_pro_10769", "mmlu_pro_10783", "mmlu_pro_10770", "mmlu_pro_10792", "mmlu_pro_10797", "mmlu_pro_10802", "mmlu_pro_10816", "mmlu_pro_10822", "mmlu_pro_10823", "mmlu_pro_10830", "mmlu_pro_10842", "mmlu_pro_10851", "mmlu_pro_10854", "mmlu_pro_10863", "mmlu_pro_10868", "mmlu_pro_10878", "mmlu_pro_10873", "mmlu_pro_10882", "mmlu_pro_10883", "mmlu_pro_10893", "mmlu_pro_10895", "mmlu_pro_10903", "mmlu_pro_10905", "mmlu_pro_10909", "mmlu_pro_10910", "mmlu_pro_10922", "mmlu_pro_10925", "mmlu_pro_10928", "mmlu_pro_10932", "mmlu_pro_10936", "mmlu_pro_10937", "mmlu_pro_10969", "mmlu_pro_10971", "mmlu_pro_10979"]

abc_exp/results/gate_rerun/qids_phi-4-mini_s0.json

@@ -0,0 +1 @@
+["mmlu_pro_19", "mmlu_pro_10", "mmlu_pro_27", "mmlu_pro_20", "mmlu_pro_29", "mmlu_pro_48", "mmlu_pro_49", "mmlu_pro_39", "mmlu_pro_52", "mmlu_pro_60", "mmlu_pro_73", "mmlu_pro_102", "mmlu_pro_88", "mmlu_pro_107", "mmlu_pro_112", "mmlu_pro_123", "mmlu_pro_117", "mmlu_pro_154", "mmlu_pro_142", "mmlu_pro_157", "mmlu_pro_167", "mmlu_pro_165", "mmlu_pro_169", "mmlu_pro_171", "mmlu_pro_177", "mmlu_pro_180", "mmlu_pro_201", "mmlu_pro_181", "mmlu_pro_213", "mmlu_pro_211", "mmlu_pro_220", "mmlu_pro_222", "mmlu_pro_233", "mmlu_pro_234", "mmlu_pro_238", "mmlu_pro_240", "mmlu_pro_245", "mmlu_pro_256", "mmlu_pro_257", "mmlu_pro_254", "mmlu_pro_261", "mmlu_pro_264", "mmlu_pro_269", "mmlu_pro_267", "mmlu_pro_273", "mmlu_pro_275", "mmlu_pro_279", "mmlu_pro_291", "mmlu_pro_293", "mmlu_pro_296", "mmlu_pro_295", "mmlu_pro_299", "mmlu_pro_303", "mmlu_pro_308", "mmlu_pro_318", "mmlu_pro_309", "mmlu_pro_320", "mmlu_pro_321", "mmlu_pro_328", "mmlu_pro_325", "mmlu_pro_334", "mmlu_pro_364", "mmlu_pro_344", "mmlu_pro_369", "mmlu_pro_370", "mmlu_pro_372", "mmlu_pro_377", "mmlu_pro_382", "mmlu_pro_384", "mmlu_pro_412", "mmlu_pro_413", "mmlu_pro_420", "mmlu_pro_416", "mmlu_pro_425", "mmlu_pro_430", "mmlu_pro_431", "mmlu_pro_436", "mmlu_pro_440", "mmlu_pro_458", "mmlu_pro_451", "mmlu_pro_459", "mmlu_pro_476", "mmlu_pro_480", "mmlu_pro_491", "mmlu_pro_498", "mmlu_pro_485", "mmlu_pro_517", "mmlu_pro_522", "mmlu_pro_521", "mmlu_pro_526", "mmlu_pro_539", "mmlu_pro_530", "mmlu_pro_542", "mmlu_pro_552", "mmlu_pro_554", "mmlu_pro_556", "mmlu_pro_558", "mmlu_pro_577", "mmlu_pro_590", "mmlu_pro_596", "mmlu_pro_602", "mmlu_pro_610", "mmlu_pro_604", "mmlu_pro_612", "mmlu_pro_615", "mmlu_pro_627", "mmlu_pro_631", "mmlu_pro_633", "mmlu_pro_646", "mmlu_pro_658", "mmlu_pro_661", "mmlu_pro_657", "mmlu_pro_673", "mmlu_pro_678", "mmlu_pro_683", "mmlu_pro_686", "mmlu_pro_693", "mmlu_pro_703", "mmlu_pro_716", "mmlu_pro_712", "mmlu_pro_726", "mmlu_pro_743", "mmlu_pro_754", "mmlu_pro_753", "mmlu_pro_755", "mmlu_pro_756", "mmlu_pro_758", "mmlu_pro_766", "mmlu_pro_765", "mmlu_pro_768", "mmlu_pro_791", "mmlu_pro_796", "mmlu_pro_797", "mmlu_pro_822", "mmlu_pro_830", "mmlu_pro_833", "mmlu_pro_835", "mmlu_pro_841", "mmlu_pro_852", "mmlu_pro_859", "mmlu_pro_863", "mmlu_pro_880", "mmlu_pro_866", "mmlu_pro_883", "mmlu_pro_886", "mmlu_pro_890", "mmlu_pro_891", "mmlu_pro_895", "mmlu_pro_896", "mmlu_pro_903", "mmlu_pro_910", "mmlu_pro_913", "mmlu_pro_919", "mmlu_pro_932", "mmlu_pro_954", "mmlu_pro_943", "mmlu_pro_968", "mmlu_pro_972", "mmlu_pro_978", "mmlu_pro_982", "mmlu_pro_996", "mmlu_pro_1001", "mmlu_pro_1014", "mmlu_pro_1011", "mmlu_pro_1027", "mmlu_pro_1029", "mmlu_pro_1034", "mmlu_pro_1038", "mmlu_pro_1040", "mmlu_pro_1055", "mmlu_pro_1070", "mmlu_pro_1061", "mmlu_pro_1082", "mmlu_pro_1081", "mmlu_pro_1092", "mmlu_pro_1105", "mmlu_pro_1113", "mmlu_pro_1112", "mmlu_pro_1117", "mmlu_pro_1123", "mmlu_pro_1156", "mmlu_pro_1173", "mmlu_pro_1179", "mmlu_pro_1174", "mmlu_pro_1181", "mmlu_pro_1198", "mmlu_pro_1191", "mmlu_pro_1199", "mmlu_pro_1209", "mmlu_pro_1212", "mmlu_pro_1221", "mmlu_pro_1217", "mmlu_pro_1227", "mmlu_pro_1233", "mmlu_pro_1241", "mmlu_pro_1243", "mmlu_pro_1252", "mmlu_pro_1247", "mmlu_pro_1255", "mmlu_pro_1261", "mmlu_pro_1265", "mmlu_pro_1263", "mmlu_pro_1278", "mmlu_pro_1268", "mmlu_pro_1299", "mmlu_pro_1305", "mmlu_pro_1308", "mmlu_pro_1306", "mmlu_pro_1327", "mmlu_pro_1326", "mmlu_pro_1339", "mmlu_pro_1341", "mmlu_pro_1351", "mmlu_pro_1353", "mmlu_pro_1355", "mmlu_pro_1358", "mmlu_pro_1360", "mmlu_pro_1362", "mmlu_pro_1367", "mmlu_pro_1365", "mmlu_pro_1378", "mmlu_pro_1380", "mmlu_pro_1383", "mmlu_pro_1388", "mmlu_pro_1390", "mmlu_pro_1396", "mmlu_pro_1402", "mmlu_pro_1399", "mmlu_pro_1410", "mmlu_pro_1413", "mmlu_pro_1414", "mmlu_pro_1416", "mmlu_pro_1418", "mmlu_pro_1429", "mmlu_pro_1433", "mmlu_pro_1430", "mmlu_pro_1442", "mmlu_pro_1436", "mmlu_pro_1451", "mmlu_pro_1453", "mmlu_pro_1462", "mmlu_pro_1469", "mmlu_pro_1479", "mmlu_pro_1481", "mmlu_pro_1482", "mmlu_pro_1485", "mmlu_pro_1500", "mmlu_pro_1509", "mmlu_pro_1502", "mmlu_pro_1511", "mmlu_pro_1515", "mmlu_pro_1519", "mmlu_pro_1522", "mmlu_pro_1534", "mmlu_pro_1541", "mmlu_pro_1544", "mmlu_pro_1543", "mmlu_pro_1547", "mmlu_pro_1548", "mmlu_pro_1563", "mmlu_pro_1567", "mmlu_pro_1568", "mmlu_pro_1576", "mmlu_pro_1589", "mmlu_pro_1590", "mmlu_pro_1606", "mmlu_pro_1612", "mmlu_pro_1615", "mmlu_pro_1626", "mmlu_pro_1618", "mmlu_pro_1628", "mmlu_pro_1639", "mmlu_pro_1648", "mmlu_pro_1646", "mmlu_pro_1649", "mmlu_pro_1652", "mmlu_pro_1661", "mmlu_pro_1663", "mmlu_pro_1664", "mmlu_pro_1676", "mmlu_pro_1681", "mmlu_pro_1689", "mmlu_pro_1694", "mmlu_pro_1710", "mmlu_pro_1715", "mmlu_pro_1716", "mmlu_pro_1721", "mmlu_pro_1727", "mmlu_pro_1729", "mmlu_pro_1733", "mmlu_pro_1730", "mmlu_pro_1750", "mmlu_pro_1766", "mmlu_pro_1772", "mmlu_pro_1774", "mmlu_pro_1776", "mmlu_pro_1793", "mmlu_pro_1801", "mmlu_pro_1815", "mmlu_pro_1847", "mmlu_pro_1838", "mmlu_pro_1855", "mmlu_pro_1863", "mmlu_pro_1865", "mmlu_pro_1869", "mmlu_pro_1879", "mmlu_pro_1891", "mmlu_pro_1905", "mmlu_pro_1897", "mmlu_pro_1907", "mmlu_pro_1916", "mmlu_pro_1911", "mmlu_pro_1917", "mmlu_pro_1918", "mmlu_pro_1927", "mmlu_pro_1938", "mmlu_pro_1939", "mmlu_pro_1949", "mmlu_pro_1959", "mmlu_pro_1954", "mmlu_pro_1969", "mmlu_pro_1971", "mmlu_pro_1980", "mmlu_pro_1983", "mmlu_pro_1991", "mmlu_pro_1999", "mmlu_pro_1994", "mmlu_pro_2011", "mmlu_pro_2006", "mmlu_pro_2014", "mmlu_pro_2017", "mmlu_pro_2022", "mmlu_pro_2023", "mmlu_pro_2026", "mmlu_pro_2051", "mmlu_pro_2054", "mmlu_pro_2057", "mmlu_pro_2068", "mmlu_pro_2066", "mmlu_pro_2076", "mmlu_pro_2078", "mmlu_pro_2073", "mmlu_pro_2081", "mmlu_pro_2105", "mmlu_pro_2107", "mmlu_pro_2114", "mmlu_pro_2117", "mmlu_pro_2119", "mmlu_pro_2144", "mmlu_pro_2156", "mmlu_pro_2165", "mmlu_pro_2160", "mmlu_pro_2177", "mmlu_pro_2167", "mmlu_pro_2181", "mmlu_pro_2178", "mmlu_pro_2184", "mmlu_pro_2193", "mmlu_pro_2199", "mmlu_pro_2191", "mmlu_pro_2208", "mmlu_pro_2227", "mmlu_pro_2238", "mmlu_pro_2240", "mmlu_pro_2242", "mmlu_pro_2241", "mmlu_pro_2260", "mmlu_pro_2274", "mmlu_pro_2276", "mmlu_pro_2303", "mmlu_pro_2313", "mmlu_pro_2306", "mmlu_pro_2315", "mmlu_pro_2334", "mmlu_pro_2337", "mmlu_pro_2338", "mmlu_pro_2340", "mmlu_pro_2352", "mmlu_pro_2362", "mmlu_pro_2360", "mmlu_pro_2365", "mmlu_pro_2366", "mmlu_pro_2381", "mmlu_pro_2388", "mmlu_pro_2395", "mmlu_pro_2390", "mmlu_pro_2407", "mmlu_pro_2423", "mmlu_pro_2424", "mmlu_pro_2440", "mmlu_pro_2429", "mmlu_pro_2458", "mmlu_pro_2468", "mmlu_pro_2472", "mmlu_pro_2481", "mmlu_pro_2480", "mmlu_pro_2487", "mmlu_pro_2490", "mmlu_pro_2492", "mmlu_pro_2496", "mmlu_pro_2497", "mmlu_pro_2511", "mmlu_pro_2498", "mmlu_pro_2516", "mmlu_pro_2521", "mmlu_pro_2535", "mmlu_pro_2522", "mmlu_pro_2539", "mmlu_pro_2540", "mmlu_pro_2544", "mmlu_pro_2545", "mmlu_pro_2553", "mmlu_pro_2567", "mmlu_pro_2556", "mmlu_pro_2577", "mmlu_pro_2574", "mmlu_pro_2579", "mmlu_pro_2582", "mmlu_pro_2586", "mmlu_pro_2588", "mmlu_pro_2594", "mmlu_pro_2596", "mmlu_pro_2598", "mmlu_pro_2610", "mmlu_pro_2606", "mmlu_pro_2627", "mmlu_pro_2621", "mmlu_pro_2630", "mmlu_pro_2633", "mmlu_pro_2634", "mmlu_pro_2638", "mmlu_pro_2651", "mmlu_pro_2652", "mmlu_pro_2654", "mmlu_pro_2665", "mmlu_pro_2666", "mmlu_pro_2672", "mmlu_pro_2676", "mmlu_pro_2680", "mmlu_pro_2694", "mmlu_pro_2692", "mmlu_pro_2701", "mmlu_pro_2706", "mmlu_pro_2709", "mmlu_pro_2718", "mmlu_pro_2727", "mmlu_pro_2720", "mmlu_pro_2732", "mmlu_pro_2734", "mmlu_pro_2747", "mmlu_pro_2753", "mmlu_pro_2757", "mmlu_pro_2761", "mmlu_pro_2770", "mmlu_pro_2786", "mmlu_pro_2790", "mmlu_pro_2802", "mmlu_pro_2803", "mmlu_pro_2811", "mmlu_pro_2815", "mmlu_pro_2831", "mmlu_pro_2833", "mmlu_pro_2832", "mmlu_pro_2837", "mmlu_pro_2841", "mmlu_pro_2845", "mmlu_pro_2835", "mmlu_pro_2858", "mmlu_pro_2848", "mmlu_pro_2860", "mmlu_pro_2861", "mmlu_pro_2864", "mmlu_pro_2876", "mmlu_pro_2898", "mmlu_pro_2906", "mmlu_pro_2914", "mmlu_pro_2929", "mmlu_pro_2918", "mmlu_pro_2939", "mmlu_pro_2930", "mmlu_pro_2965", "mmlu_pro_2953", "mmlu_pro_2970", "mmlu_pro_2980", "mmlu_pro_2985", "mmlu_pro_2986", "mmlu_pro_2987", "mmlu_pro_2988", "mmlu_pro_3022", "mmlu_pro_3019", "mmlu_pro_3028", "mmlu_pro_3031", "mmlu_pro_3037", "mmlu_pro_3041", "mmlu_pro_3043", "mmlu_pro_3049", "mmlu_pro_3046", "mmlu_pro_3069", "mmlu_pro_3070", "mmlu_pro_3072", "mmlu_pro_3074", "mmlu_pro_3086", "mmlu_pro_3089", "mmlu_pro_3087", "mmlu_pro_3091", "mmlu_pro_3101", "mmlu_pro_3092", "mmlu_pro_3108", "mmlu_pro_3118", "mmlu_pro_3114", "mmlu_pro_3120", "mmlu_pro_3126", "mmlu_pro_3144", "mmlu_pro_3159", "mmlu_pro_3161", "mmlu_pro_3162", "mmlu_pro_3171", "mmlu_pro_3196", "mmlu_pro_3211", "mmlu_pro_3177", "mmlu_pro_3218", "mmlu_pro_3219", "mmlu_pro_3230", "mmlu_pro_3259", "mmlu_pro_3253", "mmlu_pro_3265", "mmlu_pro_3269", "mmlu_pro_3270", "mmlu_pro_3273", "mmlu_pro_3275", "mmlu_pro_3289", "mmlu_pro_3281", "mmlu_pro_3299", "mmlu_pro_3303", "mmlu_pro_3305", "mmlu_pro_3310", "mmlu_pro_3306", "mmlu_pro_3315", "mmlu_pro_3311", "mmlu_pro_3320", "mmlu_pro_3319", "mmlu_pro_3331", "mmlu_pro_3338", "mmlu_pro_3336", "mmlu_pro_3339", "mmlu_pro_3340", "mmlu_pro_3352", "mmlu_pro_3354", "mmlu_pro_3358", "mmlu_pro_3362", "mmlu_pro_3364", "mmlu_pro_3363", "mmlu_pro_3369", "mmlu_pro_3370", "mmlu_pro_3378", "mmlu_pro_3375", "mmlu_pro_3404", "mmlu_pro_3402", "mmlu_pro_3410", "mmlu_pro_3409", "mmlu_pro_3414", "mmlu_pro_3419", "mmlu_pro_3417", "mmlu_pro_3433", "mmlu_pro_3421", "mmlu_pro_3444", "mmlu_pro_3434", "mmlu_pro_3446", "mmlu_pro_3453", "mmlu_pro_3459", "mmlu_pro_3460", "mmlu_pro_3461", "mmlu_pro_3465", "mmlu_pro_3470", "mmlu_pro_3479", "mmlu_pro_3489", "mmlu_pro_3511", "mmlu_pro_3499", "mmlu_pro_3522", "mmlu_pro_3523", "mmlu_pro_3534", "mmlu_pro_3539", "mmlu_pro_3543", "mmlu_pro_3538", "mmlu_pro_3562", "mmlu_pro_3558", "mmlu_pro_3572", "mmlu_pro_3570", "mmlu_pro_3580", "mmlu_pro_3579", "mmlu_pro_3584", "mmlu_pro_3587", "mmlu_pro_3598", "mmlu_pro_3599", "mmlu_pro_3600", "mmlu_pro_3610", "mmlu_pro_3612", "mmlu_pro_3626", "mmlu_pro_3630", "mmlu_pro_3644", "mmlu_pro_3637", "mmlu_pro_3646", "mmlu_pro_3647", "mmlu_pro_3670", "mmlu_pro_3678", "mmlu_pro_3681", "mmlu_pro_3687", "mmlu_pro_3680", "mmlu_pro_3695", "mmlu_pro_3696", "mmlu_pro_3703", "mmlu_pro_3711", "mmlu_pro_3720", "mmlu_pro_3722", "mmlu_pro_3729", "mmlu_pro_3736", "mmlu_pro_3725", "mmlu_pro_3737", "mmlu_pro_3739", "mmlu_pro_3740", "mmlu_pro_3757", "mmlu_pro_3759", "mmlu_pro_3758", "mmlu_pro_3764", "mmlu_pro_3768", "mmlu_pro_3786", "mmlu_pro_3777", "mmlu_pro_3791", "mmlu_pro_3805", "mmlu_pro_3807", "mmlu_pro_3817", "mmlu_pro_3815", "mmlu_pro_3822", "mmlu_pro_3819", "mmlu_pro_3826", "mmlu_pro_3828", "mmlu_pro_3832", "mmlu_pro_3835", "mmlu_pro_3842", "mmlu_pro_3855", "mmlu_pro_3863", "mmlu_pro_3856", "mmlu_pro_3864", "mmlu_pro_3871", "mmlu_pro_3867", "mmlu_pro_3895", "mmlu_pro_3903", "mmlu_pro_3902", "mmlu_pro_3909", "mmlu_pro_3910", "mmlu_pro_3912", "mmlu_pro_3924", "mmlu_pro_3913", "mmlu_pro_3936", "mmlu_pro_3954", "mmlu_pro_3955", "mmlu_pro_3976", "mmlu_pro_3964", "mmlu_pro_3979", "mmlu_pro_3981", "mmlu_pro_3985", "mmlu_pro_3991", "mmlu_pro_3994", "mmlu_pro_4004", "mmlu_pro_4006", "mmlu_pro_4009", "mmlu_pro_4018", "mmlu_pro_4029", "mmlu_pro_4032", "mmlu_pro_4038", "mmlu_pro_4037", "mmlu_pro_4040", "mmlu_pro_4057", "mmlu_pro_4045", "mmlu_pro_4064", "mmlu_pro_4089", "mmlu_pro_4063", "mmlu_pro_4105", "mmlu_pro_4117", "mmlu_pro_4109", "mmlu_pro_4139", "mmlu_pro_4143", "mmlu_pro_4144", "mmlu_pro_4146", "mmlu_pro_4158", "mmlu_pro_4148", "mmlu_pro_4163", "mmlu_pro_4165", "mmlu_pro_4181", "mmlu_pro_4198", "mmlu_pro_4204", "mmlu_pro_4207", "mmlu_pro_4205", "mmlu_pro_4210", "mmlu_pro_4217", "mmlu_pro_4215", "mmlu_pro_4226", "mmlu_pro_4228", "mmlu_pro_4232", "mmlu_pro_4234", "mmlu_pro_4242", "mmlu_pro_4243", "mmlu_pro_4246", "mmlu_pro_4253", "mmlu_pro_4258", "mmlu_pro_4261", "mmlu_pro_4270", "mmlu_pro_4269", "mmlu_pro_4279", "mmlu_pro_4287", "mmlu_pro_4295", "mmlu_pro_4296", "mmlu_pro_4297", "mmlu_pro_4302", "mmlu_pro_4333", "mmlu_pro_4340", "mmlu_pro_4343", "mmlu_pro_4350", "mmlu_pro_4371", "mmlu_pro_4392", "mmlu_pro_4399", "mmlu_pro_4406", "mmlu_pro_4411", "mmlu_pro_4410", "mmlu_pro_4434", "mmlu_pro_4433", "mmlu_pro_4439", "mmlu_pro_4459", "mmlu_pro_4468", "mmlu_pro_4479", "mmlu_pro_4489", "mmlu_pro_4508", "mmlu_pro_4507", "mmlu_pro_4509", "mmlu_pro_4512", "mmlu_pro_4518", "mmlu_pro_4522", "mmlu_pro_4519", "mmlu_pro_4524", "mmlu_pro_4531", "mmlu_pro_4528", "mmlu_pro_4534", "mmlu_pro_4537", "mmlu_pro_4536", "mmlu_pro_4544", "mmlu_pro_4551", "mmlu_pro_4562", "mmlu_pro_4557", "mmlu_pro_4574", "mmlu_pro_4583", "mmlu_pro_4586", "mmlu_pro_4589", "mmlu_pro_4597", "mmlu_pro_4598", "mmlu_pro_4606", "mmlu_pro_4616", "mmlu_pro_4617", "mmlu_pro_4630", "mmlu_pro_4621", "mmlu_pro_4637", "mmlu_pro_4636", "mmlu_pro_4639", "mmlu_pro_4640", "mmlu_pro_4643", "mmlu_pro_4660", "mmlu_pro_4661", "mmlu_pro_4670", "mmlu_pro_4674", "mmlu_pro_4684", "mmlu_pro_4681", "mmlu_pro_4696", "mmlu_pro_4698", "mmlu_pro_4704", "mmlu_pro_4703", "mmlu_pro_4720", "mmlu_pro_4709", "mmlu_pro_4724", "mmlu_pro_4723", "mmlu_pro_4736", "mmlu_pro_4744", "mmlu_pro_4760", "mmlu_pro_4753", "mmlu_pro_4763", "mmlu_pro_4771", "mmlu_pro_4776", "mmlu_pro_4786", "mmlu_pro_4787", "mmlu_pro_4792", "mmlu_pro_4794", "mmlu_pro_4795", "mmlu_pro_4807", "mmlu_pro_4800", "mmlu_pro_4814", "mmlu_pro_4820", "mmlu_pro_4829", "mmlu_pro_4836", "mmlu_pro_4877", "mmlu_pro_4842", "mmlu_pro_4878", "mmlu_pro_4884", "mmlu_pro_4893", "mmlu_pro_4897", "mmlu_pro_4904", "mmlu_pro_4899", "mmlu_pro_4916", "mmlu_pro_4924", "mmlu_pro_4938", "mmlu_pro_4939", "mmlu_pro_4955", "mmlu_pro_4949", "mmlu_pro_4957", "mmlu_pro_4956", "mmlu_pro_4961", "mmlu_pro_4960", "mmlu_pro_4986", "mmlu_pro_4969", "mmlu_pro_4992", "mmlu_pro_5023", "mmlu_pro_5016", "mmlu_pro_5046", "mmlu_pro_5038", "mmlu_pro_5058", "mmlu_pro_5050", "mmlu_pro_5061", "mmlu_pro_5072", "mmlu_pro_5074", "mmlu_pro_5078", "mmlu_pro_5080", "mmlu_pro_5085", "mmlu_pro_5086", "mmlu_pro_5088", "mmlu_pro_5091", "mmlu_pro_5093", "mmlu_pro_5103", "mmlu_pro_5095", "mmlu_pro_5114", "mmlu_pro_5108", "mmlu_pro_5128", "mmlu_pro_5121", "mmlu_pro_5134", "mmlu_pro_5130", "mmlu_pro_5140", "mmlu_pro_5149", "mmlu_pro_5153", "mmlu_pro_5157", "mmlu_pro_5158", "mmlu_pro_5159", "mmlu_pro_5167", "mmlu_pro_5172", "mmlu_pro_5191", "mmlu_pro_5193", "mmlu_pro_5212", "mmlu_pro_5214", "mmlu_pro_5225", "mmlu_pro_5226", "mmlu_pro_5229", "mmlu_pro_5231", "mmlu_pro_5237", "mmlu_pro_5245", "mmlu_pro_5248", "mmlu_pro_5246", "mmlu_pro_5256", "mmlu_pro_5262", "mmlu_pro_5263", "mmlu_pro_5264", "mmlu_pro_5267", "mmlu_pro_5270", "mmlu_pro_5279", "mmlu_pro_5281", "mmlu_pro_5294", "mmlu_pro_5300", "mmlu_pro_5305", "mmlu_pro_5311", "mmlu_pro_5319", "mmlu_pro_5328", "mmlu_pro_5324", "mmlu_pro_5331", "mmlu_pro_5333", "mmlu_pro_5348", "mmlu_pro_5339", "mmlu_pro_5369", "mmlu_pro_5374", "mmlu_pro_5381", "mmlu_pro_5387", "mmlu_pro_5396", "mmlu_pro_5393", "mmlu_pro_5421", "mmlu_pro_5423", "mmlu_pro_5431", "mmlu_pro_5443", "mmlu_pro_5445", "mmlu_pro_5447", "mmlu_pro_5450", "mmlu_pro_5459", "mmlu_pro_5456", "mmlu_pro_5473", "mmlu_pro_5465", "mmlu_pro_5482", "mmlu_pro_5500", "mmlu_pro_5501", "mmlu_pro_5502", "mmlu_pro_5507", "mmlu_pro_5512", "mmlu_pro_5519", "mmlu_pro_5523", "mmlu_pro_5522", "mmlu_pro_5540", "mmlu_pro_5541", "mmlu_pro_5542", "mmlu_pro_5557", "mmlu_pro_5554", "mmlu_pro_5561", "mmlu_pro_5563", "mmlu_pro_5571", "mmlu_pro_5565", "mmlu_pro_5588", "mmlu_pro_5598", "mmlu_pro_5607", "mmlu_pro_5600", "mmlu_pro_5610", "mmlu_pro_5613", "mmlu_pro_5616", "mmlu_pro_5626", "mmlu_pro_5623", "mmlu_pro_5630", "mmlu_pro_5638", "mmlu_pro_5639", "mmlu_pro_5647", "mmlu_pro_5661", "mmlu_pro_5662", "mmlu_pro_5669", "mmlu_pro_5672", "mmlu_pro_5677", "mmlu_pro_5695", "mmlu_pro_5696", "mmlu_pro_5705", "mmlu_pro_5711", "mmlu_pro_5709", "mmlu_pro_5714", "mmlu_pro_5716", "mmlu_pro_5721", "mmlu_pro_5737", "mmlu_pro_5730", "mmlu_pro_5738", "mmlu_pro_5746", "mmlu_pro_5760", "mmlu_pro_5763", "mmlu_pro_5780", "mmlu_pro_5778", "mmlu_pro_5784", "mmlu_pro_5791", "mmlu_pro_5812", "mmlu_pro_5815", "mmlu_pro_5817", "mmlu_pro_5818", "mmlu_pro_5823", "mmlu_pro_5829", "mmlu_pro_5836", "mmlu_pro_5854", "mmlu_pro_5853", "mmlu_pro_5855", "mmlu_pro_5860", "mmlu_pro_5876", "mmlu_pro_5879", "mmlu_pro_5888", "mmlu_pro_5891", "mmlu_pro_5899", "mmlu_pro_5902", "mmlu_pro_5911", "mmlu_pro_5905", "mmlu_pro_5927", "mmlu_pro_5912", "mmlu_pro_5931", "mmlu_pro_5932", "mmlu_pro_5947", "mmlu_pro_5955", "mmlu_pro_5959", "mmlu_pro_5965", "mmlu_pro_5960", "mmlu_pro_5967", "mmlu_pro_5966"]

abc_exp/results/gate_rerun/qids_phi-4-mini_s1.json

@@ -0,0 +1 @@
+["mmlu_pro_5971", "mmlu_pro_5981", "mmlu_pro_5982", "mmlu_pro_5985", "mmlu_pro_5998", "mmlu_pro_5997", "mmlu_pro_6007", "mmlu_pro_6009", "mmlu_pro_6012", "mmlu_pro_6020", "mmlu_pro_6027", "mmlu_pro_6021", "mmlu_pro_6032", "mmlu_pro_6029", "mmlu_pro_6034", "mmlu_pro_6033", "mmlu_pro_6038", "mmlu_pro_6052", "mmlu_pro_6055", "mmlu_pro_6054", "mmlu_pro_6062", "mmlu_pro_6073", "mmlu_pro_6074", "mmlu_pro_6084", "mmlu_pro_6079", "mmlu_pro_6094", "mmlu_pro_6088", "mmlu_pro_6095", "mmlu_pro_6114", "mmlu_pro_6120", "mmlu_pro_6134", "mmlu_pro_6130", "mmlu_pro_6140", "mmlu_pro_6144", "mmlu_pro_6152", "mmlu_pro_6160", "mmlu_pro_6162", "mmlu_pro_6171", "mmlu_pro_6176", "mmlu_pro_6188", "mmlu_pro_6177", "mmlu_pro_6194", "mmlu_pro_6204", "mmlu_pro_6206", "mmlu_pro_6212", "mmlu_pro_6215", "mmlu_pro_6217", "mmlu_pro_6219", "mmlu_pro_6221", "mmlu_pro_6229", "mmlu_pro_6237", "mmlu_pro_6249", "mmlu_pro_6252", "mmlu_pro_6263", "mmlu_pro_6269", "mmlu_pro_6291", "mmlu_pro_6299", "mmlu_pro_6302", "mmlu_pro_6310", "mmlu_pro_6306", "mmlu_pro_6314", "mmlu_pro_6313", "mmlu_pro_6317", "mmlu_pro_6332", "mmlu_pro_6318", "mmlu_pro_6341", "mmlu_pro_6336", "mmlu_pro_6347", "mmlu_pro_6344", "mmlu_pro_6354", "mmlu_pro_6367", "mmlu_pro_6372", "mmlu_pro_6379", "mmlu_pro_6382", "mmlu_pro_6389", "mmlu_pro_6413", "mmlu_pro_6435", "mmlu_pro_6443", "mmlu_pro_6452", "mmlu_pro_6464", "mmlu_pro_6471", "mmlu_pro_6485", "mmlu_pro_6482", "mmlu_pro_6488", "mmlu_pro_6487", "mmlu_pro_6491", "mmlu_pro_6490", "mmlu_pro_6494", "mmlu_pro_6495", "mmlu_pro_6499", "mmlu_pro_6504", "mmlu_pro_6523", "mmlu_pro_6524", "mmlu_pro_6537", "mmlu_pro_6539", "mmlu_pro_6542", "mmlu_pro_6546", "mmlu_pro_6547", "mmlu_pro_6554", "mmlu_pro_6565", "mmlu_pro_6567", "mmlu_pro_6576", "mmlu_pro_6586", "mmlu_pro_6592", "mmlu_pro_6596", "mmlu_pro_6609", "mmlu_pro_6624", "mmlu_pro_6625", "mmlu_pro_6646", "mmlu_pro_6658", "mmlu_pro_6674", "mmlu_pro_6679", "mmlu_pro_6682", "mmlu_pro_6694", "mmlu_pro_6700", "mmlu_pro_6720", "mmlu_pro_6722", "mmlu_pro_6726", "mmlu_pro_6729", "mmlu_pro_6733", "mmlu_pro_6752", "mmlu_pro_6736", "mmlu_pro_6760", "mmlu_pro_6763", "mmlu_pro_6765", "mmlu_pro_6769", "mmlu_pro_6770", "mmlu_pro_6771", "mmlu_pro_6775", "mmlu_pro_6780", "mmlu_pro_6786", "mmlu_pro_6802", "mmlu_pro_6804", "mmlu_pro_6807", "mmlu_pro_6817", "mmlu_pro_6820", "mmlu_pro_6829", "mmlu_pro_6849", "mmlu_pro_6845", "mmlu_pro_6852", "mmlu_pro_6861", "mmlu_pro_6872", "mmlu_pro_6865", "mmlu_pro_6879", "mmlu_pro_6881", "mmlu_pro_6886", "mmlu_pro_6885", "mmlu_pro_6888", "mmlu_pro_6891", "mmlu_pro_6894", "mmlu_pro_6892", "mmlu_pro_6920", "mmlu_pro_6907", "mmlu_pro_6921", "mmlu_pro_6923", "mmlu_pro_6928", "mmlu_pro_6933", "mmlu_pro_6939", "mmlu_pro_6942", "mmlu_pro_6964", "mmlu_pro_6970", "mmlu_pro_6975", "mmlu_pro_6973", "mmlu_pro_6995", "mmlu_pro_6994", "mmlu_pro_7001", "mmlu_pro_7002", "mmlu_pro_7007", "mmlu_pro_7008", "mmlu_pro_7011", "mmlu_pro_7043", "mmlu_pro_7029", "mmlu_pro_7049", "mmlu_pro_7050", "mmlu_pro_7051", "mmlu_pro_7058", "mmlu_pro_7076", "mmlu_pro_7085", "mmlu_pro_7095", "mmlu_pro_7097", "mmlu_pro_7100", "mmlu_pro_7110", "mmlu_pro_7090", "mmlu_pro_7113", "mmlu_pro_7111", "mmlu_pro_7118", "mmlu_pro_7127", "mmlu_pro_7126", "mmlu_pro_7131", "mmlu_pro_7132", "mmlu_pro_7154", "mmlu_pro_7142", "mmlu_pro_7163", "mmlu_pro_7171", "mmlu_pro_7175", "mmlu_pro_7176", "mmlu_pro_7181", "mmlu_pro_7183", "mmlu_pro_7190", "mmlu_pro_7194", "mmlu_pro_7209", "mmlu_pro_7232", "mmlu_pro_7249", "mmlu_pro_7236", "mmlu_pro_7252", "mmlu_pro_7269", "mmlu_pro_7279", "mmlu_pro_7253", "mmlu_pro_7283", "mmlu_pro_7285", "mmlu_pro_7287", "mmlu_pro_7284", "mmlu_pro_7294", "mmlu_pro_7305", "mmlu_pro_7306", "mmlu_pro_7327", "mmlu_pro_7331", "mmlu_pro_7345", "mmlu_pro_7344", "mmlu_pro_7359", "mmlu_pro_7361", "mmlu_pro_7365", "mmlu_pro_7367", "mmlu_pro_7369", "mmlu_pro_7371", "mmlu_pro_7372", "mmlu_pro_7376", "mmlu_pro_7380", "mmlu_pro_7386", "mmlu_pro_7388", "mmlu_pro_7391", "mmlu_pro_7395", "mmlu_pro_7397", "mmlu_pro_7406", "mmlu_pro_7402", "mmlu_pro_7415", "mmlu_pro_7413", "mmlu_pro_7423", "mmlu_pro_7427", "mmlu_pro_7422", "mmlu_pro_7444", "mmlu_pro_7438", "mmlu_pro_7468", "mmlu_pro_7470", "mmlu_pro_7473", "mmlu_pro_7482", "mmlu_pro_7501", "mmlu_pro_7506", "mmlu_pro_7503", "mmlu_pro_7516", "mmlu_pro_7524", "mmlu_pro_7523", "mmlu_pro_7532", "mmlu_pro_7528", "mmlu_pro_7538", "mmlu_pro_7540", "mmlu_pro_7544", "mmlu_pro_7564", "mmlu_pro_7572", "mmlu_pro_7589", "mmlu_pro_7595", "mmlu_pro_7600", "mmlu_pro_7601", "mmlu_pro_7596", "mmlu_pro_7612", "mmlu_pro_7619", "mmlu_pro_7620", "mmlu_pro_7621", "mmlu_pro_7626", "mmlu_pro_7629", "mmlu_pro_7635", "mmlu_pro_7639", "mmlu_pro_7633", "mmlu_pro_7660", "mmlu_pro_7656", "mmlu_pro_7675", "mmlu_pro_7661", "mmlu_pro_7694", "mmlu_pro_7682", "mmlu_pro_7708", "mmlu_pro_7709", "mmlu_pro_7715", "mmlu_pro_7716", "mmlu_pro_7723", "mmlu_pro_7717", "mmlu_pro_7737", "mmlu_pro_7727", "mmlu_pro_7750", "mmlu_pro_7755", "mmlu_pro_7764", "mmlu_pro_7770", "mmlu_pro_7765", "mmlu_pro_7777", "mmlu_pro_7779", "mmlu_pro_7789", "mmlu_pro_7790", "mmlu_pro_7792", "mmlu_pro_7798", "mmlu_pro_7812", "mmlu_pro_7808", "mmlu_pro_7817", "mmlu_pro_7828", "mmlu_pro_7826", "mmlu_pro_7836", "mmlu_pro_7835", "mmlu_pro_7848", "mmlu_pro_7837", "mmlu_pro_7851", "mmlu_pro_7852", "mmlu_pro_7858", "mmlu_pro_7871", "mmlu_pro_7878", "mmlu_pro_7880", "mmlu_pro_7893", "mmlu_pro_7887", "mmlu_pro_7900", "mmlu_pro_7909", "mmlu_pro_7917", "mmlu_pro_7918", "mmlu_pro_7902", "mmlu_pro_7925", "mmlu_pro_7934", "mmlu_pro_7936", "mmlu_pro_7939", "mmlu_pro_7940", "mmlu_pro_7945", "mmlu_pro_7953", "mmlu_pro_7946", "mmlu_pro_7957", "mmlu_pro_7962", "mmlu_pro_7975", "mmlu_pro_7981", "mmlu_pro_7972", "mmlu_pro_7983", "mmlu_pro_7989", "mmlu_pro_7993", "mmlu_pro_7995", "mmlu_pro_8003", "mmlu_pro_7996", "mmlu_pro_8004", "mmlu_pro_8021", "mmlu_pro_8023", "mmlu_pro_8048", "mmlu_pro_8050", "mmlu_pro_8034", "mmlu_pro_8071", "mmlu_pro_8076", "mmlu_pro_8080", "mmlu_pro_8084", "mmlu_pro_8101", "mmlu_pro_8095", "mmlu_pro_8108", "mmlu_pro_8109", "mmlu_pro_8123", "mmlu_pro_8138", "mmlu_pro_8115", "mmlu_pro_8152", "mmlu_pro_8140", "mmlu_pro_8154", "mmlu_pro_8156", "mmlu_pro_8158", "mmlu_pro_8166", "mmlu_pro_8159", "mmlu_pro_8176", "mmlu_pro_8181", "mmlu_pro_8178", "mmlu_pro_8195", "mmlu_pro_8184", "mmlu_pro_8196", "mmlu_pro_8218", "mmlu_pro_8224", "mmlu_pro_8222", "mmlu_pro_8226", "mmlu_pro_8235", "mmlu_pro_8236", "mmlu_pro_8242", "mmlu_pro_8247", "mmlu_pro_8241", "mmlu_pro_8259", "mmlu_pro_8249", "mmlu_pro_8262", "mmlu_pro_8273", "mmlu_pro_8275", "mmlu_pro_8277", "mmlu_pro_8284", "mmlu_pro_8283", "mmlu_pro_8291", "mmlu_pro_8294", "mmlu_pro_8301", "mmlu_pro_8295", "mmlu_pro_8302", "mmlu_pro_8320", "mmlu_pro_8333", "mmlu_pro_8337", "mmlu_pro_8341", "mmlu_pro_8342", "mmlu_pro_8354", "mmlu_pro_8348", "mmlu_pro_8365", "mmlu_pro_8361", "mmlu_pro_8371", "mmlu_pro_8379", "mmlu_pro_8389", "mmlu_pro_8397", "mmlu_pro_8391", "mmlu_pro_8417", "mmlu_pro_8411", "mmlu_pro_8419", "mmlu_pro_8441", "mmlu_pro_8449", "mmlu_pro_8458", "mmlu_pro_8474", "mmlu_pro_8473", "mmlu_pro_8478", "mmlu_pro_8475", "mmlu_pro_8481", "mmlu_pro_8484", "mmlu_pro_8508", "mmlu_pro_8515", "mmlu_pro_8486", "mmlu_pro_8516", "mmlu_pro_8530", "mmlu_pro_8562", "mmlu_pro_8563", "mmlu_pro_8569", "mmlu_pro_8567", "mmlu_pro_8580", "mmlu_pro_8582", "mmlu_pro_8584", "mmlu_pro_8586", "mmlu_pro_8601", "mmlu_pro_8594", "mmlu_pro_8611", "mmlu_pro_8614", "mmlu_pro_8629", "mmlu_pro_8641", "mmlu_pro_8644", "mmlu_pro_8647", "mmlu_pro_8654", "mmlu_pro_8658", "mmlu_pro_8669", "mmlu_pro_8670", "mmlu_pro_8690", "mmlu_pro_8699", "mmlu_pro_8704", "mmlu_pro_8709", "mmlu_pro_8710", "mmlu_pro_8724", "mmlu_pro_8725", "mmlu_pro_8729", "mmlu_pro_8733", "mmlu_pro_8745", "mmlu_pro_8747", "mmlu_pro_8760", "mmlu_pro_8761", "mmlu_pro_8766", "mmlu_pro_8781", "mmlu_pro_8796", "mmlu_pro_8803", "mmlu_pro_8805", "mmlu_pro_8816", "mmlu_pro_8826", "mmlu_pro_8820", "mmlu_pro_8831", "mmlu_pro_8833", "mmlu_pro_8839", "mmlu_pro_8844", "mmlu_pro_8843", "mmlu_pro_8848", "mmlu_pro_8855", "mmlu_pro_8856", "mmlu_pro_8850", "mmlu_pro_8865", "mmlu_pro_8875", "mmlu_pro_8872", "mmlu_pro_8903", "mmlu_pro_8878", "mmlu_pro_8912", "mmlu_pro_8932", "mmlu_pro_8934", "mmlu_pro_8919", "mmlu_pro_8936", "mmlu_pro_8941", "mmlu_pro_8955", "mmlu_pro_8956", "mmlu_pro_8950", "mmlu_pro_8958", "mmlu_pro_8975", "mmlu_pro_8983", "mmlu_pro_8971", "mmlu_pro_9004", "mmlu_pro_9002", "mmlu_pro_9005", "mmlu_pro_9013", "mmlu_pro_9010", "mmlu_pro_9019", "mmlu_pro_9016", "mmlu_pro_9028", "mmlu_pro_9024", "mmlu_pro_9039", "mmlu_pro_9033", "mmlu_pro_9042", "mmlu_pro_9072", "mmlu_pro_9063", "mmlu_pro_9099", "mmlu_pro_9084", "mmlu_pro_9114", "mmlu_pro_9101", "mmlu_pro_9121", "mmlu_pro_9123", "mmlu_pro_9124", "mmlu_pro_9125", "mmlu_pro_9129", "mmlu_pro_9137", "mmlu_pro_9140", "mmlu_pro_9141", "mmlu_pro_9143", "mmlu_pro_9157", "mmlu_pro_9155", "mmlu_pro_9172", "mmlu_pro_9164", "mmlu_pro_9181", "mmlu_pro_9185", "mmlu_pro_9187", "mmlu_pro_9192", "mmlu_pro_9188", "mmlu_pro_9213", "mmlu_pro_9217", "mmlu_pro_9219", "mmlu_pro_9239", "mmlu_pro_9228", "mmlu_pro_9243", "mmlu_pro_9244", "mmlu_pro_9245", "mmlu_pro_9254", "mmlu_pro_9265", "mmlu_pro_9274", "mmlu_pro_9286", "mmlu_pro_9282", "mmlu_pro_9291", "mmlu_pro_9307", "mmlu_pro_9314", "mmlu_pro_9324", "mmlu_pro_9308", "mmlu_pro_9328", "mmlu_pro_9329", "mmlu_pro_9337", "mmlu_pro_9340", "mmlu_pro_9342", "mmlu_pro_9345", "mmlu_pro_9349", "mmlu_pro_9375", "mmlu_pro_9374", "mmlu_pro_9376", "mmlu_pro_9378", "mmlu_pro_9396", "mmlu_pro_9380", "mmlu_pro_9400", "mmlu_pro_9407", "mmlu_pro_9410", "mmlu_pro_9414", "mmlu_pro_9411", "mmlu_pro_9415", "mmlu_pro_9422", "mmlu_pro_9442", "mmlu_pro_9439", "mmlu_pro_9448", "mmlu_pro_9443", "mmlu_pro_9483", "mmlu_pro_9491", "mmlu_pro_9494", "mmlu_pro_9500", "mmlu_pro_9511", "mmlu_pro_9503", "mmlu_pro_9521", "mmlu_pro_9512", "mmlu_pro_9545", "mmlu_pro_9548", "mmlu_pro_9533", "mmlu_pro_9567", "mmlu_pro_9586", "mmlu_pro_9603", "mmlu_pro_9613", "mmlu_pro_9619", "mmlu_pro_9621", "mmlu_pro_9623", "mmlu_pro_9629", "mmlu_pro_9627", "mmlu_pro_9632", "mmlu_pro_9641", "mmlu_pro_9647", "mmlu_pro_9653", "mmlu_pro_9644", "mmlu_pro_9661", "mmlu_pro_9671", "mmlu_pro_9675", "mmlu_pro_9687", "mmlu_pro_9686", "mmlu_pro_9697", "mmlu_pro_9705", "mmlu_pro_9725", "mmlu_pro_9715", "mmlu_pro_9734", "mmlu_pro_9738", "mmlu_pro_9739", "mmlu_pro_9743", "mmlu_pro_9753", "mmlu_pro_9768", "mmlu_pro_9756", "mmlu_pro_9791", "mmlu_pro_9777", "mmlu_pro_9793", "mmlu_pro_9798", "mmlu_pro_9808", "mmlu_pro_9817", "mmlu_pro_9819", "mmlu_pro_9820", "mmlu_pro_9840", "mmlu_pro_9839", "mmlu_pro_9851", "mmlu_pro_9843", "mmlu_pro_9852", "mmlu_pro_9859", "mmlu_pro_9876", "mmlu_pro_9882", "mmlu_pro_9872", "mmlu_pro_9884", "mmlu_pro_9891", "mmlu_pro_9903", "mmlu_pro_9895", "mmlu_pro_9907", "mmlu_pro_9923", "mmlu_pro_9926", "mmlu_pro_9925", "mmlu_pro_9933", "mmlu_pro_9942", "mmlu_pro_9968", "mmlu_pro_9930", "mmlu_pro_9969", "mmlu_pro_9971", "mmlu_pro_9975", "mmlu_pro_9977", "mmlu_pro_9976", "mmlu_pro_9989", "mmlu_pro_9987", "mmlu_pro_9991", "mmlu_pro_10005", "mmlu_pro_10011", "mmlu_pro_9993", "mmlu_pro_10026", "mmlu_pro_10034", "mmlu_pro_10035", "mmlu_pro_10036", "mmlu_pro_10037", "mmlu_pro_10038", "mmlu_pro_10051", "mmlu_pro_10060", "mmlu_pro_10044", "mmlu_pro_10063", "mmlu_pro_10065", "mmlu_pro_10074", "mmlu_pro_10080", "mmlu_pro_10079", "mmlu_pro_10087", "mmlu_pro_10085", "mmlu_pro_10095", "mmlu_pro_10090", "mmlu_pro_10098", "mmlu_pro_10112", "mmlu_pro_10127", "mmlu_pro_10131", "mmlu_pro_10134", "mmlu_pro_10139", "mmlu_pro_10142", "mmlu_pro_10141", "mmlu_pro_10150", "mmlu_pro_10147", "mmlu_pro_10156", "mmlu_pro_10159", "mmlu_pro_10162", "mmlu_pro_10170", "mmlu_pro_10185", "mmlu_pro_10179", "mmlu_pro_10186", "mmlu_pro_10222", "mmlu_pro_10241", "mmlu_pro_10230", "mmlu_pro_10245", "mmlu_pro_10242", "mmlu_pro_10247", "mmlu_pro_10270", "mmlu_pro_10271", "mmlu_pro_10283", "mmlu_pro_10285", "mmlu_pro_10288", "mmlu_pro_10299", "mmlu_pro_10290", "mmlu_pro_10302", "mmlu_pro_10300", "mmlu_pro_10308", "mmlu_pro_10314", "mmlu_pro_10315", "mmlu_pro_10316", "mmlu_pro_10321", "mmlu_pro_10318", "mmlu_pro_10324", "mmlu_pro_10341", "mmlu_pro_10333", "mmlu_pro_10343", "mmlu_pro_10344", "mmlu_pro_10349", "mmlu_pro_10346", "mmlu_pro_10360", "mmlu_pro_10372", "mmlu_pro_10367", "mmlu_pro_10382", "mmlu_pro_10387", "mmlu_pro_10411", "mmlu_pro_10408", "mmlu_pro_10417", "mmlu_pro_10422", "mmlu_pro_10432", "mmlu_pro_10437", "mmlu_pro_10441", "mmlu_pro_10467", "mmlu_pro_10476", "mmlu_pro_10454", "mmlu_pro_10481", "mmlu_pro_10478", "mmlu_pro_10484", "mmlu_pro_10487", "mmlu_pro_10489", "mmlu_pro_10492", "mmlu_pro_10502", "mmlu_pro_10503", "mmlu_pro_10515", "mmlu_pro_10505", "mmlu_pro_10520", "mmlu_pro_10526", "mmlu_pro_10527", "mmlu_pro_10532", "mmlu_pro_10538", "mmlu_pro_10543", "mmlu_pro_10544", "mmlu_pro_10568", "mmlu_pro_10571", "mmlu_pro_10586", "mmlu_pro_10589", "mmlu_pro_10578", "mmlu_pro_10600", "mmlu_pro_10602", "mmlu_pro_10609", "mmlu_pro_10621", "mmlu_pro_10627", "mmlu_pro_10641", "mmlu_pro_10651", "mmlu_pro_10656", "mmlu_pro_10648", "mmlu_pro_10661", "mmlu_pro_10663", "mmlu_pro_10669", "mmlu_pro_10692", "mmlu_pro_10689", "mmlu_pro_10694", "mmlu_pro_10703", "mmlu_pro_10717", "mmlu_pro_10713", "mmlu_pro_10728", "mmlu_pro_10737", "mmlu_pro_10752", "mmlu_pro_10741", "mmlu_pro_10753", "mmlu_pro_10758", "mmlu_pro_10762", "mmlu_pro_10765", "mmlu_pro_10769", "mmlu_pro_10783", "mmlu_pro_10792", "mmlu_pro_10797", "mmlu_pro_10770", "mmlu_pro_10802", "mmlu_pro_10822", "mmlu_pro_10823", "mmlu_pro_10830", "mmlu_pro_10816", "mmlu_pro_10842", "mmlu_pro_10854", "mmlu_pro_10851", "mmlu_pro_10868", "mmlu_pro_10863", "mmlu_pro_10873", "mmlu_pro_10878", "mmlu_pro_10882", "mmlu_pro_10883", "mmlu_pro_10893", "mmlu_pro_10895", "mmlu_pro_10903", "mmlu_pro_10905", "mmlu_pro_10909", "mmlu_pro_10910", "mmlu_pro_10922", "mmlu_pro_10928", "mmlu_pro_10925", "mmlu_pro_10932", "mmlu_pro_10936", "mmlu_pro_10937", "mmlu_pro_10971", "mmlu_pro_10969", "mmlu_pro_10982", "mmlu_pro_10979", "mmlu_pro_10986", "mmlu_pro_10990", "mmlu_pro_10993", "mmlu_pro_11009", "mmlu_pro_10995", "mmlu_pro_11022", "mmlu_pro_11014", "mmlu_pro_11024", "mmlu_pro_11032", "mmlu_pro_11034", "mmlu_pro_11040", "mmlu_pro_11045", "mmlu_pro_11059", "mmlu_pro_11053", "mmlu_pro_11062", "mmlu_pro_11100", "mmlu_pro_11091", "mmlu_pro_11101", "mmlu_pro_11109", "mmlu_pro_11110", "mmlu_pro_11115", "mmlu_pro_11114", "mmlu_pro_11117", "mmlu_pro_11127", "mmlu_pro_11149", "mmlu_pro_11153", "mmlu_pro_11150", "mmlu_pro_11170", "mmlu_pro_11157", "mmlu_pro_11173", "mmlu_pro_11177", "mmlu_pro_11186", "mmlu_pro_11183", "mmlu_pro_11188", "mmlu_pro_11197", "mmlu_pro_11189", "mmlu_pro_11201", "mmlu_pro_11211", "mmlu_pro_11203", "mmlu_pro_11217", "mmlu_pro_11224", "mmlu_pro_11221", "mmlu_pro_11243", "mmlu_pro_11226", "mmlu_pro_11246", "mmlu_pro_11254", "mmlu_pro_11247", "mmlu_pro_11265", "mmlu_pro_11284", "mmlu_pro_11288", "mmlu_pro_11305", "mmlu_pro_11300", "mmlu_pro_11314", "mmlu_pro_11331", "mmlu_pro_11336", "mmlu_pro_11343", "mmlu_pro_11347", "mmlu_pro_11345", "mmlu_pro_11368", "mmlu_pro_11381", "mmlu_pro_11372", "mmlu_pro_11382", "mmlu_pro_11391", "mmlu_pro_11395", "mmlu_pro_11396", "mmlu_pro_11407", "mmlu_pro_11402", "mmlu_pro_11422", "mmlu_pro_11413", "mmlu_pro_11423", "mmlu_pro_11425", "mmlu_pro_11444", "mmlu_pro_11437", "mmlu_pro_11453", "mmlu_pro_11445", "mmlu_pro_11467", "mmlu_pro_11466", "mmlu_pro_11468", "mmlu_pro_11474", "mmlu_pro_11477", "mmlu_pro_11484", "mmlu_pro_11490", "mmlu_pro_11493", "mmlu_pro_11504", "mmlu_pro_11503", "mmlu_pro_11514", "mmlu_pro_11539", "mmlu_pro_11543", "mmlu_pro_11548", "mmlu_pro_11560", "mmlu_pro_11586", "mmlu_pro_11587", "mmlu_pro_11588", "mmlu_pro_11610", "mmlu_pro_11627", "mmlu_pro_11629", "mmlu_pro_11630", "mmlu_pro_11633", "mmlu_pro_11637", "mmlu_pro_11650", "mmlu_pro_11653", "mmlu_pro_11654", "mmlu_pro_11658", "mmlu_pro_11664", "mmlu_pro_11660", "mmlu_pro_11682", "mmlu_pro_11689", "mmlu_pro_11688", "mmlu_pro_11701", "mmlu_pro_11710", "mmlu_pro_11714", "mmlu_pro_11730", "mmlu_pro_11725", "mmlu_pro_11739", "mmlu_pro_11746", "mmlu_pro_11732", "mmlu_pro_11762", "mmlu_pro_11750", "mmlu_pro_11770", "mmlu_pro_11765", "mmlu_pro_11774", "mmlu_pro_11783", "mmlu_pro_11778", "mmlu_pro_11795", "mmlu_pro_11793", "mmlu_pro_11801", "mmlu_pro_11810", "mmlu_pro_11817", "mmlu_pro_11823", "mmlu_pro_11825", "mmlu_pro_11826", "mmlu_pro_11828", "mmlu_pro_11832", "mmlu_pro_11835", "mmlu_pro_11842", "mmlu_pro_11844", "mmlu_pro_11846", "mmlu_pro_11852", "mmlu_pro_11848", "mmlu_pro_11864", "mmlu_pro_11865", "mmlu_pro_11868", "mmlu_pro_11870", "mmlu_pro_11872", "mmlu_pro_11903", "mmlu_pro_11874", "mmlu_pro_11904", "mmlu_pro_11914", "mmlu_pro_11917", "mmlu_pro_11920", "mmlu_pro_11922", "mmlu_pro_11924", "mmlu_pro_11936", "mmlu_pro_11948", "mmlu_pro_11963", "mmlu_pro_11977", "mmlu_pro_11965", "mmlu_pro_11988", "mmlu_pro_11980", "mmlu_pro_11993", "mmlu_pro_12003", "mmlu_pro_12007", "mmlu_pro_12006", "mmlu_pro_12010", "mmlu_pro_12015", "mmlu_pro_12017", "mmlu_pro_12016", "mmlu_pro_12025", "mmlu_pro_12023", "mmlu_pro_12030"]

abc_exp/results/gate_rerun/qids_qwen3-4b_s0.json

@@ -0,0 +1 @@
+["mmlu_pro_19", "mmlu_pro_10", "mmlu_pro_27", "mmlu_pro_20", "mmlu_pro_29", "mmlu_pro_39", "mmlu_pro_48", "mmlu_pro_49", "mmlu_pro_60", "mmlu_pro_52", "mmlu_pro_73", "mmlu_pro_88", "mmlu_pro_102", "mmlu_pro_107", "mmlu_pro_112", "mmlu_pro_123", "mmlu_pro_117", "mmlu_pro_154", "mmlu_pro_142", "mmlu_pro_157", "mmlu_pro_167", "mmlu_pro_165", "mmlu_pro_169", "mmlu_pro_171", "mmlu_pro_177", "mmlu_pro_180", "mmlu_pro_201", "mmlu_pro_181", "mmlu_pro_211", "mmlu_pro_213", "mmlu_pro_220", "mmlu_pro_222", "mmlu_pro_233", "mmlu_pro_234", "mmlu_pro_238", "mmlu_pro_240", "mmlu_pro_245", "mmlu_pro_256", "mmlu_pro_254", "mmlu_pro_257", "mmlu_pro_261", "mmlu_pro_264", "mmlu_pro_267", "mmlu_pro_269", "mmlu_pro_275", "mmlu_pro_273", "mmlu_pro_279", "mmlu_pro_291", "mmlu_pro_293", "mmlu_pro_295", "mmlu_pro_296", "mmlu_pro_303", "mmlu_pro_299", "mmlu_pro_308", "mmlu_pro_318", "mmlu_pro_309", "mmlu_pro_321", "mmlu_pro_320", "mmlu_pro_325", "mmlu_pro_328", "mmlu_pro_334", "mmlu_pro_344", "mmlu_pro_364", "mmlu_pro_369", "mmlu_pro_370", "mmlu_pro_372", "mmlu_pro_377", "mmlu_pro_382", "mmlu_pro_384", "mmlu_pro_413", "mmlu_pro_412", "mmlu_pro_416", "mmlu_pro_420", "mmlu_pro_425", "mmlu_pro_430", "mmlu_pro_431", "mmlu_pro_436", "mmlu_pro_440", "mmlu_pro_458", "mmlu_pro_451", "mmlu_pro_459", "mmlu_pro_476", "mmlu_pro_480", "mmlu_pro_485", "mmlu_pro_491", "mmlu_pro_498", "mmlu_pro_521", "mmlu_pro_517", "mmlu_pro_526", "mmlu_pro_522", "mmlu_pro_539", "mmlu_pro_530", "mmlu_pro_542", "mmlu_pro_552", "mmlu_pro_556", "mmlu_pro_554", "mmlu_pro_558", "mmlu_pro_577", "mmlu_pro_590", "mmlu_pro_596", "mmlu_pro_602", "mmlu_pro_604", "mmlu_pro_610", "mmlu_pro_612", "mmlu_pro_615", "mmlu_pro_627", "mmlu_pro_631", "mmlu_pro_633", "mmlu_pro_646", "mmlu_pro_657", "mmlu_pro_658", "mmlu_pro_673", "mmlu_pro_661", "mmlu_pro_678", "mmlu_pro_683", "mmlu_pro_686", "mmlu_pro_693", "mmlu_pro_703", "mmlu_pro_716", "mmlu_pro_712", "mmlu_pro_743", "mmlu_pro_726", "mmlu_pro_753", "mmlu_pro_754", "mmlu_pro_755", "mmlu_pro_756", "mmlu_pro_758", "mmlu_pro_766", "mmlu_pro_765", "mmlu_pro_768", "mmlu_pro_791", "mmlu_pro_796", "mmlu_pro_822", "mmlu_pro_797", "mmlu_pro_830", "mmlu_pro_833", "mmlu_pro_835", "mmlu_pro_841", "mmlu_pro_852", "mmlu_pro_859", "mmlu_pro_863", "mmlu_pro_866", "mmlu_pro_880", "mmlu_pro_883", "mmlu_pro_886", "mmlu_pro_891", "mmlu_pro_890", "mmlu_pro_895", "mmlu_pro_896", "mmlu_pro_903", "mmlu_pro_910", "mmlu_pro_913", "mmlu_pro_932", "mmlu_pro_919", "mmlu_pro_943", "mmlu_pro_954", "mmlu_pro_968", "mmlu_pro_972", "mmlu_pro_978", "mmlu_pro_982", "mmlu_pro_996", "mmlu_pro_1001", "mmlu_pro_1011", "mmlu_pro_1014", "mmlu_pro_1027", "mmlu_pro_1029", "mmlu_pro_1038", "mmlu_pro_1034", "mmlu_pro_1055", "mmlu_pro_1040", "mmlu_pro_1070", "mmlu_pro_1061", "mmlu_pro_1081", "mmlu_pro_1082", "mmlu_pro_1092", "mmlu_pro_1105", "mmlu_pro_1113", "mmlu_pro_1112", "mmlu_pro_1117", "mmlu_pro_1123", "mmlu_pro_1173", "mmlu_pro_1156", "mmlu_pro_1179", "mmlu_pro_1174", "mmlu_pro_1191", "mmlu_pro_1181", "mmlu_pro_1199", "mmlu_pro_1198", "mmlu_pro_1209", "mmlu_pro_1212", "mmlu_pro_1217", "mmlu_pro_1221", "mmlu_pro_1233", "mmlu_pro_1227", "mmlu_pro_1241", "mmlu_pro_1243", "mmlu_pro_1247", "mmlu_pro_1252", "mmlu_pro_1255", "mmlu_pro_1261", "mmlu_pro_1263", "mmlu_pro_1265", "mmlu_pro_1268", "mmlu_pro_1278", "mmlu_pro_1299", "mmlu_pro_1305", "mmlu_pro_1306", "mmlu_pro_1308", "mmlu_pro_1326", "mmlu_pro_1327", "mmlu_pro_1341", "mmlu_pro_1339", "mmlu_pro_1351", "mmlu_pro_1353", "mmlu_pro_1355", "mmlu_pro_1358", "mmlu_pro_1360", "mmlu_pro_1362", "mmlu_pro_1367", "mmlu_pro_1365", "mmlu_pro_1378", "mmlu_pro_1380", "mmlu_pro_1383", "mmlu_pro_1388", "mmlu_pro_1396", "mmlu_pro_1390", "mmlu_pro_1399", "mmlu_pro_1402", "mmlu_pro_1410", "mmlu_pro_1413", "mmlu_pro_1414", "mmlu_pro_1416", "mmlu_pro_1418", "mmlu_pro_1429", "mmlu_pro_1430", "mmlu_pro_1433", "mmlu_pro_1436", "mmlu_pro_1442", "mmlu_pro_1451", "mmlu_pro_1453", "mmlu_pro_1462", "mmlu_pro_1479", "mmlu_pro_1469", "mmlu_pro_1481", "mmlu_pro_1482", "mmlu_pro_1485", "mmlu_pro_1500", "mmlu_pro_1502", "mmlu_pro_1509", "mmlu_pro_1511", "mmlu_pro_1515", "mmlu_pro_1519", "mmlu_pro_1522", "mmlu_pro_1534", "mmlu_pro_1541", "mmlu_pro_1543", "mmlu_pro_1544", "mmlu_pro_1547", "mmlu_pro_1548", "mmlu_pro_1563", "mmlu_pro_1567", "mmlu_pro_1568", "mmlu_pro_1576", "mmlu_pro_1589", "mmlu_pro_1590", "mmlu_pro_1606", "mmlu_pro_1612", "mmlu_pro_1615", "mmlu_pro_1618", "mmlu_pro_1626", "mmlu_pro_1628", "mmlu_pro_1646", "mmlu_pro_1639", "mmlu_pro_1648", "mmlu_pro_1649", "mmlu_pro_1661", "mmlu_pro_1652", "mmlu_pro_1663", "mmlu_pro_1664", "mmlu_pro_1676", "mmlu_pro_1681", "mmlu_pro_1689", "mmlu_pro_1694", "mmlu_pro_1710", "mmlu_pro_1715", "mmlu_pro_1721", "mmlu_pro_1716", "mmlu_pro_1727", "mmlu_pro_1729", "mmlu_pro_1730", "mmlu_pro_1733", "mmlu_pro_1766", "mmlu_pro_1750", "mmlu_pro_1774", "mmlu_pro_1772", "mmlu_pro_1776", "mmlu_pro_1793", "mmlu_pro_1801", "mmlu_pro_1815", "mmlu_pro_1838", "mmlu_pro_1847", "mmlu_pro_1855", "mmlu_pro_1863", "mmlu_pro_1865", "mmlu_pro_1869", "mmlu_pro_1891", "mmlu_pro_1879", "mmlu_pro_1905", "mmlu_pro_1897", "mmlu_pro_1907", "mmlu_pro_1911", "mmlu_pro_1917", "mmlu_pro_1916", "mmlu_pro_1927", "mmlu_pro_1918", "mmlu_pro_1938", "mmlu_pro_1949", "mmlu_pro_1939", "mmlu_pro_1954", "mmlu_pro_1959", "mmlu_pro_1969", "mmlu_pro_1971", "mmlu_pro_1980", "mmlu_pro_1983", "mmlu_pro_1991", "mmlu_pro_1994", "mmlu_pro_1999", "mmlu_pro_2006", "mmlu_pro_2011", "mmlu_pro_2014", "mmlu_pro_2017", "mmlu_pro_2022", "mmlu_pro_2023", "mmlu_pro_2026", "mmlu_pro_2051", "mmlu_pro_2054", "mmlu_pro_2057", "mmlu_pro_2066", "mmlu_pro_2068", "mmlu_pro_2073", "mmlu_pro_2076", "mmlu_pro_2078", "mmlu_pro_2081", "mmlu_pro_2105", "mmlu_pro_2107", "mmlu_pro_2114", "mmlu_pro_2117", "mmlu_pro_2119", "mmlu_pro_2144", "mmlu_pro_2156", "mmlu_pro_2165", "mmlu_pro_2160", "mmlu_pro_2167", "mmlu_pro_2177", "mmlu_pro_2178", "mmlu_pro_2181", "mmlu_pro_2184", "mmlu_pro_2191", "mmlu_pro_2193", "mmlu_pro_2199", "mmlu_pro_2208", "mmlu_pro_2238", "mmlu_pro_2227", "mmlu_pro_2240", "mmlu_pro_2241", "mmlu_pro_2242", "mmlu_pro_2260", "mmlu_pro_2274", "mmlu_pro_2276", "mmlu_pro_2303", "mmlu_pro_2306", "mmlu_pro_2313", "mmlu_pro_2315", "mmlu_pro_2337", "mmlu_pro_2334", "mmlu_pro_2338", "mmlu_pro_2340", "mmlu_pro_2352", "mmlu_pro_2362", "mmlu_pro_2360", "mmlu_pro_2365", "mmlu_pro_2366", "mmlu_pro_2381", "mmlu_pro_2388", "mmlu_pro_2390", "mmlu_pro_2395", "mmlu_pro_2407", "mmlu_pro_2423", "mmlu_pro_2429", "mmlu_pro_2424", "mmlu_pro_2458", "mmlu_pro_2440", "mmlu_pro_2468", "mmlu_pro_2472", "mmlu_pro_2480", "mmlu_pro_2487", "mmlu_pro_2481", "mmlu_pro_2490", "mmlu_pro_2492", "mmlu_pro_2496", "mmlu_pro_2497", "mmlu_pro_2498", "mmlu_pro_2511", "mmlu_pro_2516", "mmlu_pro_2521", "mmlu_pro_2522", "mmlu_pro_2539", "mmlu_pro_2535", "mmlu_pro_2540", "mmlu_pro_2544", "mmlu_pro_2553", "mmlu_pro_2545", "mmlu_pro_2556", "mmlu_pro_2567", "mmlu_pro_2574", "mmlu_pro_2577", "mmlu_pro_2582", "mmlu_pro_2579", "mmlu_pro_2588", "mmlu_pro_2586", "mmlu_pro_2596", "mmlu_pro_2594", "mmlu_pro_2606", "mmlu_pro_2598", "mmlu_pro_2621", "mmlu_pro_2610", "mmlu_pro_2627", "mmlu_pro_2630", "mmlu_pro_2633", "mmlu_pro_2638", "mmlu_pro_2634", "mmlu_pro_2652", "mmlu_pro_2651", "mmlu_pro_2654", "mmlu_pro_2665", "mmlu_pro_2672", "mmlu_pro_2666", "mmlu_pro_2676", "mmlu_pro_2680", "mmlu_pro_2692", "mmlu_pro_2694", "mmlu_pro_2701", "mmlu_pro_2706", "mmlu_pro_2709", "mmlu_pro_2718", "mmlu_pro_2720", "mmlu_pro_2727", "mmlu_pro_2732", "mmlu_pro_2734", "mmlu_pro_2747", "mmlu_pro_2753", "mmlu_pro_2757", "mmlu_pro_2761", "mmlu_pro_2770", "mmlu_pro_2786", "mmlu_pro_2790", "mmlu_pro_2802", "mmlu_pro_2803", "mmlu_pro_2811", "mmlu_pro_2815", "mmlu_pro_2831", "mmlu_pro_2832", "mmlu_pro_2833", "mmlu_pro_2837", "mmlu_pro_2835", "mmlu_pro_2841", "mmlu_pro_2845", "mmlu_pro_2848", "mmlu_pro_2858", "mmlu_pro_2860", "mmlu_pro_2864", "mmlu_pro_2861", "mmlu_pro_2898", "mmlu_pro_2876", "mmlu_pro_2914", "mmlu_pro_2906", "mmlu_pro_2918", "mmlu_pro_2929", "mmlu_pro_2930", "mmlu_pro_2939", "mmlu_pro_2953", "mmlu_pro_2965", "mmlu_pro_2970", "mmlu_pro_2980", "mmlu_pro_2985", "mmlu_pro_2986", "mmlu_pro_2988", "mmlu_pro_2987", "mmlu_pro_3022", "mmlu_pro_3019", "mmlu_pro_3028", "mmlu_pro_3031", "mmlu_pro_3037", "mmlu_pro_3041", "mmlu_pro_3043", "mmlu_pro_3046", "mmlu_pro_3049", "mmlu_pro_3069", "mmlu_pro_3072", "mmlu_pro_3070", "mmlu_pro_3086", "mmlu_pro_3074", "mmlu_pro_3087", "mmlu_pro_3089", "mmlu_pro_3091", "mmlu_pro_3092", "mmlu_pro_3101", "mmlu_pro_3108", "mmlu_pro_3114", "mmlu_pro_3118", "mmlu_pro_3126", "mmlu_pro_3120", "mmlu_pro_3144", "mmlu_pro_3159", "mmlu_pro_3161", "mmlu_pro_3162", "mmlu_pro_3171", "mmlu_pro_3196", "mmlu_pro_3177", "mmlu_pro_3218", "mmlu_pro_3211", "mmlu_pro_3219", "mmlu_pro_3230", "mmlu_pro_3259", "mmlu_pro_3253", "mmlu_pro_3265", "mmlu_pro_3269", "mmlu_pro_3270", "mmlu_pro_3273", "mmlu_pro_3275", "mmlu_pro_3289", "mmlu_pro_3281", "mmlu_pro_3299", "mmlu_pro_3303", "mmlu_pro_3305", "mmlu_pro_3306", "mmlu_pro_3310", "mmlu_pro_3315", "mmlu_pro_3311", "mmlu_pro_3319", "mmlu_pro_3320", "mmlu_pro_3331", "mmlu_pro_3336", "mmlu_pro_3338", "mmlu_pro_3339", "mmlu_pro_3340", "mmlu_pro_3354", "mmlu_pro_3352", "mmlu_pro_3358", "mmlu_pro_3362", "mmlu_pro_3363", "mmlu_pro_3364", "mmlu_pro_3369", "mmlu_pro_3370", "mmlu_pro_3375", "mmlu_pro_3378", "mmlu_pro_3402", "mmlu_pro_3404", "mmlu_pro_3409", "mmlu_pro_3410", "mmlu_pro_3414", "mmlu_pro_3417", "mmlu_pro_3419", "mmlu_pro_3433", "mmlu_pro_3421", "mmlu_pro_3444", "mmlu_pro_3434", "mmlu_pro_3453", "mmlu_pro_3446", "mmlu_pro_3460", "mmlu_pro_3459", "mmlu_pro_3465", "mmlu_pro_3461", "mmlu_pro_3470", "mmlu_pro_3479", "mmlu_pro_3489", "mmlu_pro_3499", "mmlu_pro_3511", "mmlu_pro_3522", "mmlu_pro_3523", "mmlu_pro_3534", "mmlu_pro_3538", "mmlu_pro_3539", "mmlu_pro_3543", "mmlu_pro_3562", "mmlu_pro_3558", "mmlu_pro_3570", "mmlu_pro_3572", "mmlu_pro_3579", "mmlu_pro_3580", "mmlu_pro_3584", "mmlu_pro_3587", "mmlu_pro_3598", "mmlu_pro_3599", "mmlu_pro_3600", "mmlu_pro_3610", "mmlu_pro_3612", "mmlu_pro_3626", "mmlu_pro_3630", "mmlu_pro_3637", "mmlu_pro_3644", "mmlu_pro_3646", "mmlu_pro_3647", "mmlu_pro_3670", "mmlu_pro_3678", "mmlu_pro_3681", "mmlu_pro_3680", "mmlu_pro_3687", "mmlu_pro_3695", "mmlu_pro_3696", "mmlu_pro_3703", "mmlu_pro_3711", "mmlu_pro_3720", "mmlu_pro_3722", "mmlu_pro_3725", "mmlu_pro_3729", "mmlu_pro_3737", "mmlu_pro_3736", "mmlu_pro_3739", "mmlu_pro_3740", "mmlu_pro_3757", "mmlu_pro_3758", "mmlu_pro_3759", "mmlu_pro_3764", "mmlu_pro_3768", "mmlu_pro_3777", "mmlu_pro_3786", "mmlu_pro_3805", "mmlu_pro_3791", "mmlu_pro_3815", "mmlu_pro_3807", "mmlu_pro_3819", "mmlu_pro_3817", "mmlu_pro_3822", "mmlu_pro_3826", "mmlu_pro_3828", "mmlu_pro_3832", "mmlu_pro_3842", "mmlu_pro_3835", "mmlu_pro_3855", "mmlu_pro_3856", "mmlu_pro_3863", "mmlu_pro_3864", "mmlu_pro_3867", "mmlu_pro_3871", "mmlu_pro_3895", "mmlu_pro_3902", "mmlu_pro_3903", "mmlu_pro_3909", "mmlu_pro_3910", "mmlu_pro_3913", "mmlu_pro_3912", "mmlu_pro_3924", "mmlu_pro_3936", "mmlu_pro_3954", "mmlu_pro_3955", "mmlu_pro_3964", "mmlu_pro_3976", "mmlu_pro_3979", "mmlu_pro_3981", "mmlu_pro_3985", "mmlu_pro_3991", "mmlu_pro_3994", "mmlu_pro_4004", "mmlu_pro_4006", "mmlu_pro_4009", "mmlu_pro_4029", "mmlu_pro_4018", "mmlu_pro_4032", "mmlu_pro_4037", "mmlu_pro_4038", "mmlu_pro_4040", "mmlu_pro_4057", "mmlu_pro_4045", "mmlu_pro_4064", "mmlu_pro_4063", "mmlu_pro_4089", "mmlu_pro_4105", "mmlu_pro_4109", "mmlu_pro_4117", "mmlu_pro_4139", "mmlu_pro_4143", "mmlu_pro_4146", "mmlu_pro_4144", "mmlu_pro_4158", "mmlu_pro_4148", "mmlu_pro_4163", "mmlu_pro_4165", "mmlu_pro_4181", "mmlu_pro_4198", "mmlu_pro_4204", "mmlu_pro_4205", "mmlu_pro_4207", "mmlu_pro_4210", "mmlu_pro_4215", "mmlu_pro_4217", "mmlu_pro_4228", "mmlu_pro_4226", "mmlu_pro_4234", "mmlu_pro_4232", "mmlu_pro_4243", "mmlu_pro_4242", "mmlu_pro_4246", "mmlu_pro_4253", "mmlu_pro_4258", "mmlu_pro_4261", "mmlu_pro_4269", "mmlu_pro_4270", "mmlu_pro_4279", "mmlu_pro_4287", "mmlu_pro_4295", "mmlu_pro_4296", "mmlu_pro_4297", "mmlu_pro_4302", "mmlu_pro_4333", "mmlu_pro_4340", "mmlu_pro_4343", "mmlu_pro_4371", "mmlu_pro_4350", "mmlu_pro_4399", "mmlu_pro_4392", "mmlu_pro_4406", "mmlu_pro_4410", "mmlu_pro_4411", "mmlu_pro_4433", "mmlu_pro_4439", "mmlu_pro_4434", "mmlu_pro_4459", "mmlu_pro_4468", "mmlu_pro_4489", "mmlu_pro_4479", "mmlu_pro_4507", "mmlu_pro_4508", "mmlu_pro_4509", "mmlu_pro_4512", "mmlu_pro_4518", "mmlu_pro_4519", "mmlu_pro_4522", "mmlu_pro_4524", "mmlu_pro_4528", "mmlu_pro_4531", "mmlu_pro_4534", "mmlu_pro_4537", "mmlu_pro_4536", "mmlu_pro_4544", "mmlu_pro_4551", "mmlu_pro_4562", "mmlu_pro_4557", "mmlu_pro_4583", "mmlu_pro_4574", "mmlu_pro_4589", "mmlu_pro_4586", "mmlu_pro_4598", "mmlu_pro_4597", "mmlu_pro_4606", "mmlu_pro_4616", "mmlu_pro_4617", "mmlu_pro_4621", "mmlu_pro_4630", "mmlu_pro_4636", "mmlu_pro_4637", "mmlu_pro_4639", "mmlu_pro_4640", "mmlu_pro_4643", "mmlu_pro_4660", "mmlu_pro_4661", "mmlu_pro_4670", "mmlu_pro_4674", "mmlu_pro_4681", "mmlu_pro_4684", "mmlu_pro_4696", "mmlu_pro_4698", "mmlu_pro_4704", "mmlu_pro_4703", "mmlu_pro_4709", "mmlu_pro_4720", "mmlu_pro_4723", "mmlu_pro_4724", "mmlu_pro_4736", "mmlu_pro_4744", "mmlu_pro_4760", "mmlu_pro_4753", "mmlu_pro_4771", "mmlu_pro_4763", "mmlu_pro_4786", "mmlu_pro_4776", "mmlu_pro_4787", "mmlu_pro_4792", "mmlu_pro_4794", "mmlu_pro_4795", "mmlu_pro_4800", "mmlu_pro_4814", "mmlu_pro_4807", "mmlu_pro_4820", "mmlu_pro_4829", "mmlu_pro_4836", "mmlu_pro_4842", "mmlu_pro_4877", "mmlu_pro_4878", "mmlu_pro_4884", "mmlu_pro_4893", "mmlu_pro_4897", "mmlu_pro_4899", "mmlu_pro_4904", "mmlu_pro_4916", "mmlu_pro_4924", "mmlu_pro_4938", "mmlu_pro_4939", "mmlu_pro_4955", "mmlu_pro_4949", "mmlu_pro_4956", "mmlu_pro_4957", "mmlu_pro_4961", "mmlu_pro_4960", "mmlu_pro_4986", "mmlu_pro_4969", "mmlu_pro_4992", "mmlu_pro_5016", "mmlu_pro_5023", "mmlu_pro_5038", "mmlu_pro_5046", "mmlu_pro_5050", "mmlu_pro_5058", "mmlu_pro_5061", "mmlu_pro_5072", "mmlu_pro_5074", "mmlu_pro_5080", "mmlu_pro_5078", "mmlu_pro_5085", "mmlu_pro_5086", "mmlu_pro_5091", "mmlu_pro_5088", "mmlu_pro_5093", "mmlu_pro_5095", "mmlu_pro_5103", "mmlu_pro_5108", "mmlu_pro_5121", "mmlu_pro_5114", "mmlu_pro_5128", "mmlu_pro_5130", "mmlu_pro_5134", "mmlu_pro_5140", "mmlu_pro_5153", "mmlu_pro_5149", "mmlu_pro_5157", "mmlu_pro_5158", "mmlu_pro_5159", "mmlu_pro_5167", "mmlu_pro_5172", "mmlu_pro_5191", "mmlu_pro_5193", "mmlu_pro_5212", "mmlu_pro_5214", "mmlu_pro_5225", "mmlu_pro_5226", "mmlu_pro_5231", "mmlu_pro_5229", "mmlu_pro_5245", "mmlu_pro_5237", "mmlu_pro_5246", "mmlu_pro_5248", "mmlu_pro_5262", "mmlu_pro_5256", "mmlu_pro_5264", "mmlu_pro_5263", "mmlu_pro_5267", "mmlu_pro_5279", "mmlu_pro_5270", "mmlu_pro_5281", "mmlu_pro_5294", "mmlu_pro_5300", "mmlu_pro_5305", "mmlu_pro_5319", "mmlu_pro_5311", "mmlu_pro_5324", "mmlu_pro_5328", "mmlu_pro_5331", "mmlu_pro_5333", "mmlu_pro_5339", "mmlu_pro_5348", "mmlu_pro_5369", "mmlu_pro_5374", "mmlu_pro_5381", "mmlu_pro_5387", "mmlu_pro_5393", "mmlu_pro_5396", "mmlu_pro_5421", "mmlu_pro_5431", "mmlu_pro_5423", "mmlu_pro_5443", "mmlu_pro_5445", "mmlu_pro_5450", "mmlu_pro_5447", "mmlu_pro_5456", "mmlu_pro_5459", "mmlu_pro_5465", "mmlu_pro_5473", "mmlu_pro_5482", "mmlu_pro_5500", "mmlu_pro_5501", "mmlu_pro_5507", "mmlu_pro_5502", "mmlu_pro_5512", "mmlu_pro_5519", "mmlu_pro_5522", "mmlu_pro_5523", "mmlu_pro_5541", "mmlu_pro_5540", "mmlu_pro_5542", "mmlu_pro_5554", "mmlu_pro_5557", "mmlu_pro_5563", "mmlu_pro_5561", "mmlu_pro_5571", "mmlu_pro_5565", "mmlu_pro_5588", "mmlu_pro_5598", "mmlu_pro_5607", "mmlu_pro_5600", "mmlu_pro_5610", "mmlu_pro_5616", "mmlu_pro_5613", "mmlu_pro_5626", "mmlu_pro_5623", "mmlu_pro_5630", "mmlu_pro_5638", "mmlu_pro_5647", "mmlu_pro_5639", "mmlu_pro_5662", "mmlu_pro_5661", "mmlu_pro_5672", "mmlu_pro_5669", "mmlu_pro_5695", "mmlu_pro_5677", "mmlu_pro_5696", "mmlu_pro_5705", "mmlu_pro_5709", "mmlu_pro_5711", "mmlu_pro_5714", "mmlu_pro_5716", "mmlu_pro_5721", "mmlu_pro_5730", "mmlu_pro_5737", "mmlu_pro_5738", "mmlu_pro_5746", "mmlu_pro_5763", "mmlu_pro_5760", "mmlu_pro_5778", "mmlu_pro_5780", "mmlu_pro_5784", "mmlu_pro_5791", "mmlu_pro_5815", "mmlu_pro_5812", "mmlu_pro_5817", "mmlu_pro_5818", "mmlu_pro_5823", "mmlu_pro_5829", "mmlu_pro_5836", "mmlu_pro_5853", "mmlu_pro_5854", "mmlu_pro_5855", "mmlu_pro_5860", "mmlu_pro_5876", "mmlu_pro_5879", "mmlu_pro_5888", "mmlu_pro_5891", "mmlu_pro_5899", "mmlu_pro_5902", "mmlu_pro_5905", "mmlu_pro_5911", "mmlu_pro_5927", "mmlu_pro_5912", "mmlu_pro_5931", "mmlu_pro_5932", "mmlu_pro_5947", "mmlu_pro_5959", "mmlu_pro_5955", "mmlu_pro_5960", "mmlu_pro_5965", "mmlu_pro_5966", "mmlu_pro_5967"]

abc_exp/results/gate_rerun/qids_qwen3-4b_s1.json

@@ -0,0 +1 @@
+["mmlu_pro_5971", "mmlu_pro_5981", "mmlu_pro_5982", "mmlu_pro_5985", "mmlu_pro_5997", "mmlu_pro_5998", "mmlu_pro_6007", "mmlu_pro_6009", "mmlu_pro_6012", "mmlu_pro_6020", "mmlu_pro_6021", "mmlu_pro_6027", "mmlu_pro_6029", "mmlu_pro_6033", "mmlu_pro_6032", "mmlu_pro_6034", "mmlu_pro_6038", "mmlu_pro_6052", "mmlu_pro_6054", "mmlu_pro_6055", "mmlu_pro_6062", "mmlu_pro_6073", "mmlu_pro_6079", "mmlu_pro_6074", "mmlu_pro_6084", "mmlu_pro_6088", "mmlu_pro_6094", "mmlu_pro_6095", "mmlu_pro_6120", "mmlu_pro_6114", "mmlu_pro_6134", "mmlu_pro_6130", "mmlu_pro_6140", "mmlu_pro_6144", "mmlu_pro_6152", "mmlu_pro_6160", "mmlu_pro_6162", "mmlu_pro_6171", "mmlu_pro_6177", "mmlu_pro_6176", "mmlu_pro_6188", "mmlu_pro_6194", "mmlu_pro_6204", "mmlu_pro_6206", "mmlu_pro_6215", "mmlu_pro_6212", "mmlu_pro_6217", "mmlu_pro_6219", "mmlu_pro_6221", "mmlu_pro_6229", "mmlu_pro_6237", "mmlu_pro_6249", "mmlu_pro_6252", "mmlu_pro_6263", "mmlu_pro_6269", "mmlu_pro_6299", "mmlu_pro_6291", "mmlu_pro_6302", "mmlu_pro_6306", "mmlu_pro_6313", "mmlu_pro_6310", "mmlu_pro_6317", "mmlu_pro_6314", "mmlu_pro_6332", "mmlu_pro_6318", "mmlu_pro_6341", "mmlu_pro_6336", "mmlu_pro_6344", "mmlu_pro_6347", "mmlu_pro_6354", "mmlu_pro_6367", "mmlu_pro_6372", "mmlu_pro_6382", "mmlu_pro_6379", "mmlu_pro_6413", "mmlu_pro_6389", "mmlu_pro_6435", "mmlu_pro_6443", "mmlu_pro_6452", "mmlu_pro_6464", "mmlu_pro_6471", "mmlu_pro_6482", "mmlu_pro_6485", "mmlu_pro_6487", "mmlu_pro_6488", "mmlu_pro_6490", "mmlu_pro_6491", "mmlu_pro_6494", "mmlu_pro_6495", "mmlu_pro_6499", "mmlu_pro_6504", "mmlu_pro_6523", "mmlu_pro_6524", "mmlu_pro_6539", "mmlu_pro_6537", "mmlu_pro_6542", "mmlu_pro_6546", "mmlu_pro_6547", "mmlu_pro_6554", "mmlu_pro_6567", "mmlu_pro_6565", "mmlu_pro_6586", "mmlu_pro_6576", "mmlu_pro_6592", "mmlu_pro_6609", "mmlu_pro_6596", "mmlu_pro_6624", "mmlu_pro_6625", "mmlu_pro_6646", "mmlu_pro_6658", "mmlu_pro_6674", "mmlu_pro_6679", "mmlu_pro_6694", "mmlu_pro_6682", "mmlu_pro_6700", "mmlu_pro_6722", "mmlu_pro_6720", "mmlu_pro_6726", "mmlu_pro_6729", "mmlu_pro_6733", "mmlu_pro_6736", "mmlu_pro_6752", "mmlu_pro_6760", "mmlu_pro_6763", "mmlu_pro_6765", "mmlu_pro_6769", "mmlu_pro_6770", "mmlu_pro_6771", "mmlu_pro_6780", "mmlu_pro_6775", "mmlu_pro_6786", "mmlu_pro_6802", "mmlu_pro_6804", "mmlu_pro_6807", "mmlu_pro_6820", "mmlu_pro_6817", "mmlu_pro_6829", "mmlu_pro_6845", "mmlu_pro_6849", "mmlu_pro_6852", "mmlu_pro_6861", "mmlu_pro_6865", "mmlu_pro_6872", "mmlu_pro_6881", "mmlu_pro_6879", "mmlu_pro_6886", "mmlu_pro_6885", "mmlu_pro_6888", "mmlu_pro_6891", "mmlu_pro_6894", "mmlu_pro_6892", "mmlu_pro_6920", "mmlu_pro_6907", "mmlu_pro_6921", "mmlu_pro_6928", "mmlu_pro_6923", "mmlu_pro_6933", "mmlu_pro_6939", "mmlu_pro_6942", "mmlu_pro_6964", "mmlu_pro_6970", "mmlu_pro_6973", "mmlu_pro_6975", "mmlu_pro_6994", "mmlu_pro_6995", "mmlu_pro_7001", "mmlu_pro_7002", "mmlu_pro_7007", "mmlu_pro_7008", "mmlu_pro_7011", "mmlu_pro_7029", "mmlu_pro_7043", "mmlu_pro_7049", "mmlu_pro_7050", "mmlu_pro_7058", "mmlu_pro_7051", "mmlu_pro_7076", "mmlu_pro_7085", "mmlu_pro_7095", "mmlu_pro_7090", "mmlu_pro_7097", "mmlu_pro_7100", "mmlu_pro_7110", "mmlu_pro_7111", "mmlu_pro_7113", "mmlu_pro_7118", "mmlu_pro_7126", "mmlu_pro_7131", "mmlu_pro_7127", "mmlu_pro_7132", "mmlu_pro_7142", "mmlu_pro_7154", "mmlu_pro_7171", "mmlu_pro_7163", "mmlu_pro_7176", "mmlu_pro_7175", "mmlu_pro_7181", "mmlu_pro_7183", "mmlu_pro_7194", "mmlu_pro_7190", "mmlu_pro_7209", "mmlu_pro_7232", "mmlu_pro_7249", "mmlu_pro_7236", "mmlu_pro_7252", "mmlu_pro_7253", "mmlu_pro_7269", "mmlu_pro_7279", "mmlu_pro_7283", "mmlu_pro_7284", "mmlu_pro_7285", "mmlu_pro_7287", "mmlu_pro_7294", "mmlu_pro_7305", "mmlu_pro_7306", "mmlu_pro_7331", "mmlu_pro_7327", "mmlu_pro_7344", "mmlu_pro_7345", "mmlu_pro_7359", "mmlu_pro_7365", "mmlu_pro_7361", "mmlu_pro_7367", "mmlu_pro_7369", "mmlu_pro_7372", "mmlu_pro_7371", "mmlu_pro_7380", "mmlu_pro_7376", "mmlu_pro_7386", "mmlu_pro_7388", "mmlu_pro_7395", "mmlu_pro_7397", "mmlu_pro_7391", "mmlu_pro_7406", "mmlu_pro_7402", "mmlu_pro_7413", "mmlu_pro_7415", "mmlu_pro_7423", "mmlu_pro_7422", "mmlu_pro_7438", "mmlu_pro_7427", "mmlu_pro_7444", "mmlu_pro_7468", "mmlu_pro_7470", "mmlu_pro_7473", "mmlu_pro_7482", "mmlu_pro_7501", "mmlu_pro_7503", "mmlu_pro_7506", "mmlu_pro_7516", "mmlu_pro_7523", "mmlu_pro_7524", "mmlu_pro_7528", "mmlu_pro_7532", "mmlu_pro_7538", "mmlu_pro_7540", "mmlu_pro_7564", "mmlu_pro_7544", "mmlu_pro_7572", "mmlu_pro_7589", "mmlu_pro_7595", "mmlu_pro_7596", "mmlu_pro_7600", "mmlu_pro_7601", "mmlu_pro_7612", "mmlu_pro_7619", "mmlu_pro_7620", "mmlu_pro_7621", "mmlu_pro_7626", "mmlu_pro_7629", "mmlu_pro_7633", "mmlu_pro_7635", "mmlu_pro_7639", "mmlu_pro_7656", "mmlu_pro_7660", "mmlu_pro_7661", "mmlu_pro_7675", "mmlu_pro_7694", "mmlu_pro_7682", "mmlu_pro_7708", "mmlu_pro_7709", "mmlu_pro_7715", "mmlu_pro_7716", "mmlu_pro_7723", "mmlu_pro_7717", "mmlu_pro_7727", "mmlu_pro_7737", "mmlu_pro_7750", "mmlu_pro_7755", "mmlu_pro_7764", "mmlu_pro_7770", "mmlu_pro_7765", "mmlu_pro_7777", "mmlu_pro_7779", "mmlu_pro_7789", "mmlu_pro_7790", "mmlu_pro_7792", "mmlu_pro_7798", "mmlu_pro_7808", "mmlu_pro_7812", "mmlu_pro_7817", "mmlu_pro_7826", "mmlu_pro_7828", "mmlu_pro_7836", "mmlu_pro_7835", "mmlu_pro_7837", "mmlu_pro_7848", "mmlu_pro_7851", "mmlu_pro_7852", "mmlu_pro_7871", "mmlu_pro_7858", "mmlu_pro_7878", "mmlu_pro_7880", "mmlu_pro_7887", "mmlu_pro_7893", "mmlu_pro_7900", "mmlu_pro_7902", "mmlu_pro_7909", "mmlu_pro_7917", "mmlu_pro_7918", "mmlu_pro_7925", "mmlu_pro_7934", "mmlu_pro_7936", "mmlu_pro_7939", "mmlu_pro_7940", "mmlu_pro_7945", "mmlu_pro_7953", "mmlu_pro_7946", "mmlu_pro_7957", "mmlu_pro_7962", "mmlu_pro_7972", "mmlu_pro_7975", "mmlu_pro_7981", "mmlu_pro_7983", "mmlu_pro_7989", "mmlu_pro_7993", "mmlu_pro_7995", "mmlu_pro_7996", "mmlu_pro_8003", "mmlu_pro_8004", "mmlu_pro_8021", "mmlu_pro_8023", "mmlu_pro_8034", "mmlu_pro_8048", "mmlu_pro_8050", "mmlu_pro_8071", "mmlu_pro_8076", "mmlu_pro_8080", "mmlu_pro_8084", "mmlu_pro_8095", "mmlu_pro_8101", "mmlu_pro_8108", "mmlu_pro_8109", "mmlu_pro_8115", "mmlu_pro_8123", "mmlu_pro_8138", "mmlu_pro_8140", "mmlu_pro_8152", "mmlu_pro_8154", "mmlu_pro_8158", "mmlu_pro_8156", "mmlu_pro_8166", "mmlu_pro_8176", "mmlu_pro_8159", "mmlu_pro_8178", "mmlu_pro_8181", "mmlu_pro_8184", "mmlu_pro_8195", "mmlu_pro_8196", "mmlu_pro_8218", "mmlu_pro_8224", "mmlu_pro_8222", "mmlu_pro_8226", "mmlu_pro_8235", "mmlu_pro_8236", "mmlu_pro_8241", "mmlu_pro_8242", "mmlu_pro_8247", "mmlu_pro_8249", "mmlu_pro_8259", "mmlu_pro_8262", "mmlu_pro_8273", "mmlu_pro_8275", "mmlu_pro_8277", "mmlu_pro_8284", "mmlu_pro_8283", "mmlu_pro_8294", "mmlu_pro_8291", "mmlu_pro_8295", "mmlu_pro_8301", "mmlu_pro_8302", "mmlu_pro_8320", "mmlu_pro_8333", "mmlu_pro_8337", "mmlu_pro_8342", "mmlu_pro_8341", "mmlu_pro_8354", "mmlu_pro_8348", "mmlu_pro_8365", "mmlu_pro_8371", "mmlu_pro_8361", "mmlu_pro_8379", "mmlu_pro_8389", "mmlu_pro_8397", "mmlu_pro_8391", "mmlu_pro_8411", "mmlu_pro_8417", "mmlu_pro_8419", "mmlu_pro_8441", "mmlu_pro_8449", "mmlu_pro_8458", "mmlu_pro_8473", "mmlu_pro_8474", "mmlu_pro_8475", "mmlu_pro_8478", "mmlu_pro_8481", "mmlu_pro_8484", "mmlu_pro_8508", "mmlu_pro_8486", "mmlu_pro_8516", "mmlu_pro_8515", "mmlu_pro_8562", "mmlu_pro_8530", "mmlu_pro_8567", "mmlu_pro_8563", "mmlu_pro_8569", "mmlu_pro_8580", "mmlu_pro_8582", "mmlu_pro_8584", "mmlu_pro_8586", "mmlu_pro_8594", "mmlu_pro_8601", "mmlu_pro_8611", "mmlu_pro_8614", "mmlu_pro_8629", "mmlu_pro_8641", "mmlu_pro_8644", "mmlu_pro_8647", "mmlu_pro_8654", "mmlu_pro_8658", "mmlu_pro_8669", "mmlu_pro_8690", "mmlu_pro_8670", "mmlu_pro_8699", "mmlu_pro_8704", "mmlu_pro_8710", "mmlu_pro_8709", "mmlu_pro_8724", "mmlu_pro_8729", "mmlu_pro_8725", "mmlu_pro_8733", "mmlu_pro_8745", "mmlu_pro_8747", "mmlu_pro_8760", "mmlu_pro_8761", "mmlu_pro_8766", "mmlu_pro_8781", "mmlu_pro_8796", "mmlu_pro_8803", "mmlu_pro_8805", "mmlu_pro_8816", "mmlu_pro_8826", "mmlu_pro_8820", "mmlu_pro_8831", "mmlu_pro_8839", "mmlu_pro_8833", "mmlu_pro_8843", "mmlu_pro_8844", "mmlu_pro_8848", "mmlu_pro_8850", "mmlu_pro_8855", "mmlu_pro_8856", "mmlu_pro_8865", "mmlu_pro_8875", "mmlu_pro_8872", "mmlu_pro_8878", "mmlu_pro_8903", "mmlu_pro_8912", "mmlu_pro_8919", "mmlu_pro_8932", "mmlu_pro_8934", "mmlu_pro_8936", "mmlu_pro_8941", "mmlu_pro_8950", "mmlu_pro_8955", "mmlu_pro_8956", "mmlu_pro_8958", "mmlu_pro_8975", "mmlu_pro_8971", "mmlu_pro_8983", "mmlu_pro_9004", "mmlu_pro_9002", "mmlu_pro_9010", "mmlu_pro_9005", "mmlu_pro_9013", "mmlu_pro_9016", "mmlu_pro_9019", "mmlu_pro_9024", "mmlu_pro_9028", "mmlu_pro_9033", "mmlu_pro_9039", "mmlu_pro_9042", "mmlu_pro_9072", "mmlu_pro_9063", "mmlu_pro_9099", "mmlu_pro_9084", "mmlu_pro_9101", "mmlu_pro_9114", "mmlu_pro_9121", "mmlu_pro_9123", "mmlu_pro_9124", "mmlu_pro_9125", "mmlu_pro_9129", "mmlu_pro_9137", "mmlu_pro_9140", "mmlu_pro_9141", "mmlu_pro_9143", "mmlu_pro_9155", "mmlu_pro_9157", "mmlu_pro_9164", "mmlu_pro_9172", "mmlu_pro_9181", "mmlu_pro_9185", "mmlu_pro_9187", "mmlu_pro_9188", "mmlu_pro_9192", "mmlu_pro_9213", "mmlu_pro_9217", "mmlu_pro_9219", "mmlu_pro_9239", "mmlu_pro_9228", "mmlu_pro_9244", "mmlu_pro_9243", "mmlu_pro_9245", "mmlu_pro_9254", "mmlu_pro_9265", "mmlu_pro_9274", "mmlu_pro_9282", "mmlu_pro_9291", "mmlu_pro_9307", "mmlu_pro_9286", "mmlu_pro_9314", "mmlu_pro_9324", "mmlu_pro_9308", "mmlu_pro_9328", "mmlu_pro_9329", "mmlu_pro_9340", "mmlu_pro_9337", "mmlu_pro_9342", "mmlu_pro_9345", "mmlu_pro_9349", "mmlu_pro_9374", "mmlu_pro_9375", "mmlu_pro_9376", "mmlu_pro_9378", "mmlu_pro_9396", "mmlu_pro_9380", "mmlu_pro_9400", "mmlu_pro_9407", "mmlu_pro_9410", "mmlu_pro_9411", "mmlu_pro_9414", "mmlu_pro_9415", "mmlu_pro_9422", "mmlu_pro_9439", "mmlu_pro_9442", "mmlu_pro_9443", "mmlu_pro_9448", "mmlu_pro_9483", "mmlu_pro_9491", "mmlu_pro_9494", "mmlu_pro_9500", "mmlu_pro_9503", "mmlu_pro_9511", "mmlu_pro_9521", "mmlu_pro_9512", "mmlu_pro_9545", "mmlu_pro_9533", "mmlu_pro_9548", "mmlu_pro_9567", "mmlu_pro_9586", "mmlu_pro_9603", "mmlu_pro_9613", "mmlu_pro_9621", "mmlu_pro_9619", "mmlu_pro_9623", "mmlu_pro_9627", "mmlu_pro_9629", "mmlu_pro_9632", "mmlu_pro_9641", "mmlu_pro_9644", "mmlu_pro_9653", "mmlu_pro_9647", "mmlu_pro_9661", "mmlu_pro_9671", "mmlu_pro_9675", "mmlu_pro_9686", "mmlu_pro_9687", "mmlu_pro_9697", "mmlu_pro_9705", "mmlu_pro_9715", "mmlu_pro_9725", "mmlu_pro_9734", "mmlu_pro_9738", "mmlu_pro_9739", "mmlu_pro_9743", "mmlu_pro_9753", "mmlu_pro_9756", "mmlu_pro_9768", "mmlu_pro_9777", "mmlu_pro_9791", "mmlu_pro_9793", "mmlu_pro_9798", "mmlu_pro_9808", "mmlu_pro_9817", "mmlu_pro_9819", "mmlu_pro_9820", "mmlu_pro_9839", "mmlu_pro_9840", "mmlu_pro_9851", "mmlu_pro_9843", "mmlu_pro_9852", "mmlu_pro_9859", "mmlu_pro_9872", "mmlu_pro_9876", "mmlu_pro_9882", "mmlu_pro_9884", "mmlu_pro_9891", "mmlu_pro_9895", "mmlu_pro_9903", "mmlu_pro_9907", "mmlu_pro_9925", "mmlu_pro_9923", "mmlu_pro_9926", "mmlu_pro_9930", "mmlu_pro_9942", "mmlu_pro_9933", "mmlu_pro_9969", "mmlu_pro_9968", "mmlu_pro_9975", "mmlu_pro_9971", "mmlu_pro_9977", "mmlu_pro_9976", "mmlu_pro_9989", "mmlu_pro_9987", "mmlu_pro_9991", "mmlu_pro_9993", "mmlu_pro_10005", "mmlu_pro_10011", "mmlu_pro_10026", "mmlu_pro_10034", "mmlu_pro_10035", "mmlu_pro_10036", "mmlu_pro_10037", "mmlu_pro_10038", "mmlu_pro_10051", "mmlu_pro_10044", "mmlu_pro_10060", "mmlu_pro_10063", "mmlu_pro_10065", "mmlu_pro_10074", "mmlu_pro_10079", "mmlu_pro_10080", "mmlu_pro_10085", "mmlu_pro_10087", "mmlu_pro_10090", "mmlu_pro_10095", "mmlu_pro_10098", "mmlu_pro_10112", "mmlu_pro_10127", "mmlu_pro_10131", "mmlu_pro_10134", "mmlu_pro_10139", "mmlu_pro_10142", "mmlu_pro_10141", "mmlu_pro_10150", "mmlu_pro_10147", "mmlu_pro_10156", "mmlu_pro_10159", "mmlu_pro_10162", "mmlu_pro_10179", "mmlu_pro_10170", "mmlu_pro_10185", "mmlu_pro_10186", "mmlu_pro_10222", "mmlu_pro_10241", "mmlu_pro_10230", "mmlu_pro_10242", "mmlu_pro_10245", "mmlu_pro_10247", "mmlu_pro_10270", "mmlu_pro_10271", "mmlu_pro_10283", "mmlu_pro_10285", "mmlu_pro_10288", "mmlu_pro_10290", "mmlu_pro_10299", "mmlu_pro_10302", "mmlu_pro_10300", "mmlu_pro_10308", "mmlu_pro_10315", "mmlu_pro_10314", "mmlu_pro_10316", "mmlu_pro_10318", "mmlu_pro_10321", "mmlu_pro_10324", "mmlu_pro_10333", "mmlu_pro_10343", "mmlu_pro_10341", "mmlu_pro_10344", "mmlu_pro_10346", "mmlu_pro_10349", "mmlu_pro_10360", "mmlu_pro_10367", "mmlu_pro_10372", "mmlu_pro_10382", "mmlu_pro_10387", "mmlu_pro_10408", "mmlu_pro_10411", "mmlu_pro_10417", "mmlu_pro_10422", "mmlu_pro_10437", "mmlu_pro_10432", "mmlu_pro_10441", "mmlu_pro_10454", "mmlu_pro_10476", "mmlu_pro_10467", "mmlu_pro_10481", "mmlu_pro_10478", "mmlu_pro_10484", "mmlu_pro_10489", "mmlu_pro_10487", "mmlu_pro_10502", "mmlu_pro_10492", "mmlu_pro_10503", "mmlu_pro_10505", "mmlu_pro_10515", "mmlu_pro_10520", "mmlu_pro_10526", "mmlu_pro_10527", "mmlu_pro_10532", "mmlu_pro_10538", "mmlu_pro_10543", "mmlu_pro_10544", "mmlu_pro_10568", "mmlu_pro_10571", "mmlu_pro_10578", "mmlu_pro_10586", "mmlu_pro_10589", "mmlu_pro_10600", "mmlu_pro_10602", "mmlu_pro_10609", "mmlu_pro_10621", "mmlu_pro_10627", "mmlu_pro_10641", "mmlu_pro_10651", "mmlu_pro_10648", "mmlu_pro_10661", "mmlu_pro_10656", "mmlu_pro_10669", "mmlu_pro_10663", "mmlu_pro_10692", "mmlu_pro_10689", "mmlu_pro_10703", "mmlu_pro_10694", "mmlu_pro_10713", "mmlu_pro_10717", "mmlu_pro_10728", "mmlu_pro_10737", "mmlu_pro_10752", "mmlu_pro_10741", "mmlu_pro_10753", "mmlu_pro_10758", "mmlu_pro_10762", "mmlu_pro_10765", "mmlu_pro_10769", "mmlu_pro_10783", "mmlu_pro_10770", "mmlu_pro_10792", "mmlu_pro_10797", "mmlu_pro_10802", "mmlu_pro_10816", "mmlu_pro_10822", "mmlu_pro_10823", "mmlu_pro_10830", "mmlu_pro_10842", "mmlu_pro_10851", "mmlu_pro_10854", "mmlu_pro_10863", "mmlu_pro_10868", "mmlu_pro_10873", "mmlu_pro_10878", "mmlu_pro_10883", "mmlu_pro_10882", "mmlu_pro_10893", "mmlu_pro_10895", "mmlu_pro_10903", "mmlu_pro_10905", "mmlu_pro_10909", "mmlu_pro_10910", "mmlu_pro_10925", "mmlu_pro_10922", "mmlu_pro_10928", "mmlu_pro_10932", "mmlu_pro_10937", "mmlu_pro_10936", "mmlu_pro_10969", "mmlu_pro_10971", "mmlu_pro_10979", "mmlu_pro_10986", "mmlu_pro_10982", "mmlu_pro_10990", "mmlu_pro_10993", "mmlu_pro_11009", "mmlu_pro_10995", "mmlu_pro_11022", "mmlu_pro_11014", "mmlu_pro_11024", "mmlu_pro_11032", "mmlu_pro_11040", "mmlu_pro_11034", "mmlu_pro_11045", "mmlu_pro_11053", "mmlu_pro_11059", "mmlu_pro_11062", "mmlu_pro_11100", "mmlu_pro_11091", "mmlu_pro_11101", "mmlu_pro_11109", "mmlu_pro_11110", "mmlu_pro_11114", "mmlu_pro_11117", "mmlu_pro_11115", "mmlu_pro_11127", "mmlu_pro_11149", "mmlu_pro_11153", "mmlu_pro_11150", "mmlu_pro_11170", "mmlu_pro_11157", "mmlu_pro_11177", "mmlu_pro_11173", "mmlu_pro_11183", "mmlu_pro_11186", "mmlu_pro_11188", "mmlu_pro_11189", "mmlu_pro_11197", "mmlu_pro_11201", "mmlu_pro_11211", "mmlu_pro_11203", "mmlu_pro_11217", "mmlu_pro_11221", "mmlu_pro_11224", "mmlu_pro_11243", "mmlu_pro_11226", "mmlu_pro_11246", "mmlu_pro_11247", "mmlu_pro_11254", "mmlu_pro_11265", "mmlu_pro_11284", "mmlu_pro_11288", "mmlu_pro_11305", "mmlu_pro_11300", "mmlu_pro_11314", "mmlu_pro_11331", "mmlu_pro_11336", "mmlu_pro_11343", "mmlu_pro_11345", "mmlu_pro_11347", "mmlu_pro_11368", "mmlu_pro_11372", "mmlu_pro_11381", "mmlu_pro_11382", "mmlu_pro_11391", "mmlu_pro_11395", "mmlu_pro_11396", "mmlu_pro_11402", "mmlu_pro_11407", "mmlu_pro_11413", "mmlu_pro_11423", "mmlu_pro_11422", "mmlu_pro_11425", "mmlu_pro_11437", "mmlu_pro_11444", "mmlu_pro_11445", "mmlu_pro_11453", "mmlu_pro_11466", "mmlu_pro_11467", "mmlu_pro_11468", "mmlu_pro_11477", "mmlu_pro_11474", "mmlu_pro_11484", "mmlu_pro_11490", "mmlu_pro_11503", "mmlu_pro_11493", "mmlu_pro_11514", "mmlu_pro_11504", "mmlu_pro_11539", "mmlu_pro_11543", "mmlu_pro_11560", "mmlu_pro_11548", "mmlu_pro_11586", "mmlu_pro_11587", "mmlu_pro_11588", "mmlu_pro_11610", "mmlu_pro_11627", "mmlu_pro_11629", "mmlu_pro_11630", "mmlu_pro_11633", "mmlu_pro_11637", "mmlu_pro_11650", "mmlu_pro_11653", "mmlu_pro_11654", "mmlu_pro_11658", "mmlu_pro_11660", "mmlu_pro_11664", "mmlu_pro_11682", "mmlu_pro_11688", "mmlu_pro_11689", "mmlu_pro_11701", "mmlu_pro_11714", "mmlu_pro_11710", "mmlu_pro_11725", "mmlu_pro_11730", "mmlu_pro_11739", "mmlu_pro_11732", "mmlu_pro_11746", "mmlu_pro_11750", "mmlu_pro_11762", "mmlu_pro_11765", "mmlu_pro_11770", "mmlu_pro_11774", "mmlu_pro_11778", "mmlu_pro_11783", "mmlu_pro_11793", "mmlu_pro_11795", "mmlu_pro_11801", "mmlu_pro_11810", "mmlu_pro_11823", "mmlu_pro_11817", "mmlu_pro_11825", "mmlu_pro_11826", "mmlu_pro_11828", "mmlu_pro_11832", "mmlu_pro_11835", "mmlu_pro_11842", "mmlu_pro_11846", "mmlu_pro_11844", "mmlu_pro_11852", "mmlu_pro_11848", "mmlu_pro_11864", "mmlu_pro_11865", "mmlu_pro_11868", "mmlu_pro_11870", "mmlu_pro_11872", "mmlu_pro_11874", "mmlu_pro_11903", "mmlu_pro_11904", "mmlu_pro_11914", "mmlu_pro_11917", "mmlu_pro_11920", "mmlu_pro_11922", "mmlu_pro_11924", "mmlu_pro_11936", "mmlu_pro_11948", "mmlu_pro_11963", "mmlu_pro_11965", "mmlu_pro_11977", "mmlu_pro_11980", "mmlu_pro_11988", "mmlu_pro_11993", "mmlu_pro_12003", "mmlu_pro_12006", "mmlu_pro_12007", "mmlu_pro_12010", "mmlu_pro_12015", "mmlu_pro_12016", "mmlu_pro_12017", "mmlu_pro_12023", "mmlu_pro_12025", "mmlu_pro_12030"]

abc_exp/results/n9_partial_holm.json

@@ -0,0 +1,54 @@
+{
+ "n_models": 9,
+ "models": [
+  "Sonnet 4.5",
+  "GPT-4o-mini",
+  "GPT-5.4-mini",
+  "Gemini 3-flash",
+  "Phi-4-mini",
+  "Qwen3-4B",
+  "Llama-3.1-8B",
+  "Qwen3-8B",
+  "Qwen3-32B"
+ ],
+ "holm_family_size": 7,
+ "secondary_spearman": {
+  "alpha_adv": {
+   "rho": 0.6166666666666666,
+   "p_raw": 0.07692880993776312,
+   "p_holm": 0.4615728596265787
+  },
+  "alpha_cor": {
+   "rho": -0.19999999999999998,
+   "p_raw": 0.6059012741795534,
+   "p_holm": 1.0
+  },
+  "debate_FR": {
+   "rho": 0.326362689414732,
+   "p_raw": 0.3913604492997979,
+   "p_holm": 1.0
+  },
+  "init_acc": {
+   "rho": -0.7333333333333334,
+   "p_raw": 0.024554150071469705,
+   "p_holm": 0.17187905050028793
+  }
+ },
+ "partial_given_init_acc": {
+  "alpha_total": {
+   "partial_rho": 0.5685642641633971,
+   "p_raw": 0.14140768041442225,
+   "p_holm": 0.7070384020721112
+  },
+  "alpha_adv": {
+   "partial_rho": 0.21282857659767226,
+   "p_raw": 0.6128330238165054,
+   "p_holm": 1.0
+  },
+  "alpha_cor": {
+   "partial_rho": -0.061940498376286415,
+   "p_raw": 0.8841582767470036,
+   "p_holm": 1.0
+  }
+ }
+}