{
 "_status": "FROZEN CANDIDATE v1.0 - thresholds approved by the author 2026-10-18; batch sizes are the data-independent memory-rule result of the preflight 20261018T013921Z",
 "schema": "q32-run-config-v0.2",
 "approval": {
  "thresholds_approved_by_author": true,
  "approved_threshold_ids": [
   "E1",
   "A1",
   "A2",
   "A3",
   "A4",
   "A5",
   "A6",
   "B1",
   "B2",
   "B3",
   "B4",
   "G0-G10"
  ],
  "approved_on": "2026-10-18",
  "approved_via": "author reply in session (AskUserQuestion): 'Zatwierdzam wszystkie bez zmian'; MMLU M500 declined",
  "note": "approved exactly as written in PREREG_QWEN3_32B.md v0.2 / KRYTERIA_DO_ZATWIERDZENIA.md; no change after the freeze without a logged deviation"
 },
 "model": {
  "id": "Qwen/Qwen3-32B",
  "revision": "9216db5781bf21249d130ec9da846c4624c16137",
  "n_params": 32762123264,
  "tensor_bytes": 65524246528,
  "file_bytes": 65524328560,
  "n_shards": 17,
  "tokenizer_json_sha256": "aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4",
  "tokenizer_config_sha256": "d5d09f07b48c3086c508b30d1c9114bd1189145b74e982a265350c923acd8101",
  "n_blocks": 64,
  "n_hidden_states": 65,
  "d_model": 5120
 },
 "load": {
  "dtype": "bfloat16",
  "expect_dtype": "bfloat16",
  "device_map": {
   "": 0
  },
  "env": {
   "EA_DEVICE_MAP_SINGLE": "0"
  },
  "forbidden_env": [
   "EA_LOAD_IN_8BIT",
   "EA_LOAD_IN_4BIT",
   "EA_PREQUANTISED",
   "EA_MAX_MEMORY_GB"
  ],
  "local_weights_dir": "/content/qwen3-32b-rev9216db57"
 },
 "data": {
  "prompts_sha1_16": "6ba7456f5cb03427",
  "set_counts": {
   "main": 200,
   "single": 384,
   "leaveout": 384,
   "factorial": 192,
   "control": 300,
   "ood": 30
  },
  "eab_repo": "https://github.com/aisa-group/decomposing-eval-awareness",
  "eab_commit": "2961c17ebcb35fd15d5c614c4874d1f567b85d1b",
  "geometry_sets": "main,single,leaveout,control,ood",
  "factorial_sets": "main,factorial,control,ood",
  "mmlu": {
   "n_items": 600,
   "seed": 0,
   "sha1_first300": "228ae7b83e7d47af7d786f99c0de68abef7c893a",
   "sha1_all600": "ae8e1e4f8362a0d9b193822d0aa6237aa470a029"
  }
 },
 "template": {
  "geometry": {
   "enable_thinking": true,
   "expected_last_token_id": 198,
   "expected_last_token": "\n",
   "max_len_observed": 800
  },
  "mmlu_and_ablation_scan": {
   "enable_thinking": false,
   "assistant_prefix": "Answer:",
   "expected_last_template_token_id": 271,
   "max_len_observed": 804
  },
  "verbalisation": {
   "cot_prefix": "<think>\n"
  },
  "max_len": 1024,
  "letter_token_ids": [
   362,
   425,
   356,
   422
  ],
  "pad_token_id": 151643
 },
 "layer_rule": {
  "n_blocks": 64,
  "n_hidden_states": 65,
  "primary": "Lp = argmax over hidden states 0..63 of paired d_z on the 100 matched main pairs; hidden state 64 (post final RMSNorm) is excluded; ties -> lowest index",
  "analysis": "sorted set {Lp, argmax(probe_auroc, paired_dz), argmax(dom_auroc, paired_dz), round(0.25*64)=16, round(0.5*64)=32, round(0.75*64)=48} over states 0..63",
  "locality": "first three analysis layers != Lp in ascending order",
  "no_layer_numbers_carried_over_from_other_models": true
 },
 "batch_sizes": {
  "collect": 8,
  "mmlu_validate": 8,
  "ablation_scan": 8,
  "single_prompt_scripts": {
   "test_interventions": 1,
   "steering_verify": 1,
   "steering_mmlu": 1,
   "steering_generate": 1,
   "verbalize_grid": 1
  },
  "rule": "largest of {1,2,4,8,16} whose worst-case workload keeps max_memory_reserved <= 0.90 x device memory (preflight, data-independent)",
  "preflight": "20261018T013921Z",
  "change_after_freeze": "forbidden: OOM = gate G4 failure, no retry at another batch size"
 },
 "generation": {
  "mmlu_validate_max_new": 12,
  "ablation_scan_max_new": 64,
  "steering_generate_max_new": 96,
  "decoding": "greedy (do_sample=False)",
  "longest_generation_tokens": 96,
  "memory_test_at_real_limit": "preflight workload generate_96_hooked (max_new 96, hooks, KV cache) at B=1..8; no stage generates more than 96 tokens"
 },
 "sizes": {
  "full": {
   "n_boot": 1000,
   "boot_seed": 0,
   "mmlu_runs": {
    "M300": {
     "n_logprob": 300,
     "n_generate": 100,
     "max_new": 12,
     "dump_n": 20
    },
    "M600": {
     "n_logprob": 600,
     "n_generate": 200,
     "max_new": 12,
     "dump_n": 20
    }
   },
   "ablation_scan": {
    "n_mmlu": 200,
    "n_gen": 24,
    "n_ppl": 40,
    "max_new": 64,
    "strengths": "0,0.25,0.5,0.75,1.0"
   },
   "steering": {
    "mmlu_limit": 200,
    "verify_set": "ood",
    "verify_limit": 1000000000,
    "gen_sets": "ood",
    "gen_conds": "eval_all,deploy"
   },
   "vgrid": {
    "sets": "ood",
    "limit": 1000000000,
    "strengths": [
     "0.25",
     "0.5",
     "0.75",
     "1.0"
    ],
    "kinds": [
     "ablate",
     "ablate_random",
     "ablate_random_matched"
    ]
   },
   "judge_blind_n": 40,
   "max_tasks": 0
  },
  "smoke": {
   "n_boot": 20,
   "boot_seed": 0,
   "mmlu_runs": {
    "M300": {
     "n_logprob": 24,
     "n_generate": 8,
     "max_new": 12,
     "dump_n": 4
    },
    "M600": {
     "n_logprob": 24,
     "n_generate": 8,
     "max_new": 12,
     "dump_n": 4
    }
   },
   "ablation_scan": {
    "n_mmlu": 8,
    "n_gen": 4,
    "n_ppl": 6,
    "max_new": 16,
    "strengths": "0,0.5,1.0"
   },
   "steering": {
    "mmlu_limit": 8,
    "verify_set": "ood",
    "verify_limit": 6,
    "gen_sets": "ood",
    "gen_conds": "eval_all,deploy"
   },
   "vgrid": {
    "sets": "ood",
    "limit": 4,
    "strengths": [
     "0.5",
     "1.0"
    ],
    "kinds": [
     "ablate",
     "ablate_random_matched"
    ]
   },
   "judge_blind_n": 4,
   "max_tasks": 8
  }
 },
 "steering": {
  "conds_mmlu": "none,ablate,ablate_layerwise,ablate_layerwise_random,ablate_random0,add_x1,add_x2,add_random",
  "conds_gen": "none,ablate,ablate_layerwise,add_x2,add_random",
  "alpha": "diff_norm of layers[Lp] (mean eval-deploy difference norm at Lp)"
 },
 "gates": {
  "G0": {
   "min_vram_gib": 75.0,
   "min_compute_capability": 8,
   "n_gpu": 1
  },
  "G2": {
   "max_len": 1024
  },
  "G4": {
   "mem_reserved_frac_max": 0.9
  },
  "G5": {
   "measure": "minimum cosine similarity between last-token hidden state of a batched left-padded forward (B=8) and the single-prompt forward; 8 prompts (length quantiles); gated layers 1,16,32,48,63 (threshold below), layer 64 reported only",
   "cos_min": 0.99,
   "evaluated": "preflight 20261018T013921Z (observed min 0.99816 at layer 63); not repeated in the run"
  },
  "G6": {
   "fp16_abs_max": 60000
  },
  "G7": {
   "shuffle_null_range": [
    0.4,
    0.6
   ],
   "n_main_pairs": 100
  },
  "G8": {
   "ablate_residual_frac_max": 0.1,
   "random_shift_frac_max": 0.1,
   "scale": "max(|mean_proj(none)|, sd_proj(none))",
   "floor": "eps(bf16)*mean_norm(none)/sqrt(d_model)",
   "bf16_floor_mult": 10,
   "bf16_floor_calibration": "calibrate_bf16_floor.py: seeded random direction + fixed non-benchmark sentence, layers 1..63; K = ceil(2 x max ratio of residual/single-rounding-floor over layers and the three checks); record in bf16_floor_calibration.json (hashed in the manifest)"
  },
  "G9": {
   "neutral_logprob_acc_min": 0.5,
   "scorer_agreement_neutral_min": 0.85,
   "unparsable_rate_max": 0.1
  },
  "G10": {
   "baseline_degeneracy_max": 0.3,
   "baseline_special_spam_max": 0.3
  }
 },
 "decision_thresholds": {
  "_status": "APPROVED by the author 2026-10-18, unchanged from the proposal",
  "E1": {
   "ceiling_min": 0.6,
   "separation_R_hi_lt": 0.5,
   "one_axis_R_lo_gt": 0.8
  },
  "A1": {
   "cos_ci_lo_ge": 0.7,
   "cos_ci_hi_lt": 0.7,
   "necessity_drop_ge": 0.25,
   "necessity_ratio_ge": 2.0
  },
  "A2": {
   "lofo_ci_hi_lt": 0.75,
   "own_auroc_ge": 0.65,
   "contradict_ci_lo_ge": 0.85
  },
  "A3": {
   "pc1_lt": 0.5,
   "participation_ratio_gt": 3.0
  },
  "A4": {
   "abs_frac_le": 0.1
  },
  "A5": {
   "auroc_ge": 0.9
  },
  "A6": {
   "auroc_after_length_ge": 0.9
  },
  "B1": {
   "abs_delta_le": 0.03,
   "mcnemar_p_gt": 0.05,
   "boot_n": 10000,
   "boot_seed": 0
  },
  "B2": {
   "degeneracy_margin": 0.1,
   "ppl_ratio_max": 1.5
  }
 },
 "software": {
  "transformers": "==5.19.0",
  "others": "accelerate safetensors numpy scipy scikit-learn pandas matplotlib pyarrow pyyaml huggingface_hub",
  "torch": "as installed on host; recorded in the run manifest"
 },
 "storage": {
  "private_dataset": "Mati83moni/qwen3-32b-evalaware",
  "per_tag_dir": "runs/<tag>/",
  "activations_dtype": "float16 (guard G6)"
 },
 "code": {
  "origin": "Space Mati83moni/evaluation-awareness-pllum-12b revision 7e0ad711f0142891dbfefb4fc486043da62919ac (read-only source)",
  "src_tools_sha256": {
   "src/ablation_scan.py": "fd7066f52755afb6c39adb019f2ed3fbbb50e1bae295375bca14b99f1d1c4cce",
   "src/analyze_mmlu.py": "c1ff9d515c966e830da71905e8551fd373363791e7b3f6049717453494bcf4ef",
   "src/bootstrap_geometry.py": "3dec443ce0e1d20bef7980655b1d55f9f69169ea698ad9114f574e211dd774fa",
   "src/build_dataset.py": "5152daa288ffe1451fbe5c06313239e3afb5f3a0e3500e25f926c8d541c4ea91",
   "src/collect_activations.py": "199b6fe1beaf9b380667e58bc2dc3832e42a7e28385450d590088efb7f949318",
   "src/compare_directions.py": "19d0b41c4c4a04d5b4d76fcdaef05d64ace2c7ded8e7a1cc289369b7a87d4d7c",
   "src/crosslayer.py": "ca4623af78444d079ed8604423dc82763afce34488e6c3f2907c2198a5582b10",
   "src/decorrelate.py": "ff86e728cb548096a2bc2e6d9eb6cb2adda3b1c0798461d6b662adf7183162a9",
   "src/directions.py": "a07b860cf59f8fb5adf725c7293fabe1f9e9f12381a645680736fd784161440b",
   "src/judge.py": "73b962a67dd3949dc68ea2ade6a6c4a8639c8b6145deed0c3571f064471ae4be",
   "src/massive_activations.py": "341de80803e5cd75260d7c28bd036ab6c8d0b9cb9fc8c93ed0ee41a12901a897",
   "src/mmlu_validate.py": "cd3bc37a0abe2050dcd93700dfdcb5fc45efa91d7d9f38f9254604c61710c59f",
   "src/model_utils.py": "61552a3e970325bcc5a56476cb2e6e58427a6cb8253aedc4e633596efb163ef8",
   "src/plots.py": "e0d0bb9482e618f441bf319c38cbe38a935bfc8de751bd2dae6bc1c822b43a14",
   "src/prepare_mmlu.py": "2f899875d3d37edccbd70b7ce7cda435e9def79e9b1f8663d6e0af8faaa52915",
   "src/register_factorial.py": "18f6986ef3b27c9cea264c2eb7c9c11b089e2e538bcdcc7545cf20bee093cdb6",
   "src/steering.py": "88c4718b32374dae07bf59b6f8108a14f5e1b79ebc08b925fa3b83f4ffefe08c",
   "src/summarize.py": "18aaa1917f77d9ab4ab9ec938fcddad4a2d4113952d96c45f8d4384a54389a6b",
   "src/test_interventions.py": "6d023d7393dcf45cf0ca8d9c990d6b22855fcb32816e7066b9869296588fa195",
   "src/verbalize.py": "fbdb8a0e34de99ee8e2e3c941d05680bbf0ff3987c2a9727fc2d92cf8f911dd6",
   "src/verbalize_grid.py": "3321589d37814968b3a0668dd896b09c87fd803a6c517edcdf7086226f92a1d5",
   "tools/check_tokenizer.py": "770d288eee47688fe3ec79104e72cf7707f2d59cad9795624f367e89c457da05",
   "tools/load_probe.py": "43525ec07d501a56093f316f66b685157961a6f7bd0d954f89290f896f3df349"
  },
  "new_files": [
   "colab_q32.py",
   "memwrap.py",
   "evaluate_criteria.py",
   "freeze_manifest.py",
   "selftest_interventions.py",
   "calibrate_bf16_floor.py",
   "bf16_floor_calibration.json"
  ]
 }
}