{"what":"d1-3B: round 4's five red arms (a word changed, two grammatical 'not's, two state swaps): precheck = every arm against the provider's fp32 oracle (an arm the oracle does not move by the bar is replaced; round 4 replaced the word arm's base tv4_000 by tv4_001); mac_fp16 / iphone_int8mlp = the arms on the two shipped decoders at S = 64 (round 6a, the graph's |dp| against the oracle's on the same rows); fp16_S16 = the same on the fp16 S = 16 decoder (round 4). The arms' changed sentences are held back where the base record is a reference-only transfer-v4 row.","sources":[{"file":"results/r4_red_precheck.json","bytes":7381,"sha256":"2d67b1e402ccad047b387361fe3f0cf2a9a5909b958ae7dd3e32cb1feef28be6"},{"file":"results/r6a_readout_fp16_pf64.json","bytes":570025,"sha256":"a65331d3218918ff8c196830c196798c82038b847a6c841a53cceaf56781094d"},{"file":"results/r6a_readout_int8mlp_pf64.json","bytes":572833,"sha256":"16f28d47362ccdc5533fa3851c1113a6a4c261638e63d37b6bd2cd2b8f526dc5"},{"file":"results/r4_readout_fp16_pf16_red_r4.json","bytes":23250,"sha256":"37ddbcc23d7d14c415912f72bfbee7ff6a3779f3b0be54270d56b0cc7e7960ea"}],"trimmed":{"removed":["per-call ms lists (call_ms, first_call_of_process_ms; each transcript keeps its medians)","hidden-dump file names and hashes (npz, npz_key, npz_sha256, hidden_npz) and the hidden rows' lowest positions","the oracle's probabilities inside the transcripts (probs_oracle: they are in fixtures-d1-3b.json)","token ids and rendered prompt text (row_ids, ids, text, instructions, state, criteria)","other sessions' processes seen by the drivers (other_gpu_processes*, other_gpu_jobs*, busiest, top)","the arms' from / to sentences of reference-only bases (red_arms_r4.json's change text)"],"serialization":"compact json (no indentation)","full_copies":"the source files above, in the lane directory $ZOO_WORK_ROOT/_d1_3b/ (not in the repository)","paths":"local paths rewritten: the lane directory as $ZOO_WORK_ROOT/_d1_3b, the repository relative, the home directory as ~, the checkpoint snapshot as <snapshot>, the overlay venv as <overlay venv>, temporary directories as $TMPDIR, Xcode as $DEVELOPER_DIR, the phone's app container as <app container>, the phone's ids as <udid> / <coredevice id>","process_ids":"removed (keys pid / pids; 'pid <n>' in text)"},"precheck":{"what":"Round 4 step 2 (red arms): the gate's red rule on the provider's own probabilities, before any graph — every arm's","rule":"red = the perturbed rows against their base rows fail the gate's own bar (FACTS §7): (a) an argmax moves on a question whose oracle top-2 margin is above 0.02, or (b) max |dp| > 0.02, or (c) the mean over the arm's rows of the row's mean |dp| > 0.002","red_arms_sha256":"54fb9ad6063464e1f30f4d7b45da5e782b82e73b1c8ba8854fb7c055c4353655","host_rows_equal_oracle_rows":"7/7","base_rows":[["tv4_000",0],["tv4x_qnli_00",0],["tv4x_paws_00",0],["tv4_001",0],["semif_a3f18f3a63d45345942b",0],["semif_0b43ea8e24e74d621f1b",0]],"precheck":{"available":true,"path":"$ZOO_WORK_ROOT/_d1_3b/oracle/red/records_oracle.json","sha256":"0ef94ca3db8e27d3ecdc4679200d91d51b3a8ed74693d1bdbf52df943641cb27","red_on_oracle":["red_word_tv4_000","red_not_qnli_00","red_not_paws_00","state_swap_tv4_000__tv4_001","state_swap_semif_a3f18f3a63d45345942b__semif_0b43ea8e24e74d621f1b"],"replace":[]},"arms":[{"id":"red_word_tv4_000","kind":"word","rows":1,"argmax_moved":0,"argmax_moved_non_near_tie":0,"max_abs_dp_vs_base":0.006901481412935784,"mean_of_run_mean_abs_dp_vs_base":0.003535534460187008,"red_facts7":{"a_argmax_non_near_tie":false,"b_max_abs_dp":false,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":false,"items":[{"request":"red_word_tv4_000","question":"answer","base":"tv4_000:answer","tokens":106,"oracle_top2_margin":0.9175063266514067,"near_tie":false,"argmax_moved":false,"max_abs_dp_vs_base":0.006901481412935784,"mean_abs_dp_vs_base":0.003535534460187008,"base_probs":[0.012302076516283622,0.030425813328589246,0.947932139979996,0.00933997017513106],"perturbed_probs":[0.012652991743084051,0.03714596702216278,0.9410306585670603,0.009170382667692776]}]},{"id":"red_not_qnli_00","kind":"not","rows":1,"argmax_moved":1,"argmax_moved_non_near_tie":1,"max_abs_dp_vs_base":0.9579308844100081,"mean_of_run_mean_abs_dp_vs_base":0.9579308844100081,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true,"items":[{"request":"red_not_qnli_00","question":"answers","base":"tv4x_qnli_00:answers","tokens":106,"oracle_top2_margin":0.967676077803126,"near_tie":false,"argmax_moved":true,"max_abs_dp_vs_base":0.9579308844100081,"mean_abs_dp_vs_base":0.9579308844100081,"base_probs":[0.9838380389015631,0.016161961098437056],"perturbed_probs":[0.025907154491554996,0.974092845508445]}]},{"id":"red_not_paws_00","kind":"not","rows":1,"argmax_moved":0,"argmax_moved_non_near_tie":0,"max_abs_dp_vs_base":0.035111090960022726,"mean_of_run_mean_abs_dp_vs_base":0.03511109096002268,"red_facts7":{"a_argmax_non_near_tie":false,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true,"items":[{"request":"red_not_paws_00","question":"paraphrase","base":"tv4x_paws_00:paraphrase","tokens":100,"oracle_top2_margin":0.8207418652737444,"near_tie":false,"argmax_moved":false,"max_abs_dp_vs_base":0.035111090960022726,"mean_abs_dp_vs_base":0.03511109096002268,"base_probs":[0.08962906736312783,0.9103709326368722],"perturbed_probs":[0.12474015832315055,0.8752598416768496]}]},{"id":"state_swap_tv4_000__tv4_001","kind":"state_swap","rows":2,"argmax_moved":1,"argmax_moved_non_near_tie":1,"max_abs_dp_vs_base":0.8397602748609407,"mean_of_run_mean_abs_dp_vs_base":0.2163106998495101,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true,"items":[{"request":"tv4_000_with_state_of_tv4_001","question":"answer","base":"tv4_000:answer","tokens":93,"oracle_top2_margin":0.9175063266514067,"near_tie":false,"argmax_moved":false,"max_abs_dp_vs_base":0.02548252453709965,"mean_abs_dp_vs_base":0.012741262268549835,"base_probs":[0.012302076516283622,0.030425813328589246,0.947932139979996,0.00933997017513106],"perturbed_probs":[0.012533761012188059,0.04713503500216771,0.9224496154428964,0.017881588542747854]},{"request":"tv4_001_with_state_of_tv4_000","question":"answer","base":"tv4_001:answer","tokens":90,"oracle_top2_margin":0.9880834835402893,"near_tie":false,"argmax_moved":true,"max_abs_dp_vs_base":0.8397602748609407,"mean_abs_dp_vs_base":0.41988013743047037,"base_probs":[0.0011603837368078472,0.003069886708698145,0.9919266065473916,0.0038431230071023417],"perturbed_probs":[0.5712300466898866,0.16233342548760524,0.15216633168645083,0.11427019613605725]}]},{"id":"state_swap_semif_a3f18f3a63d45345942b__semif_0b43ea8e24e74d621f1b","kind":"state_swap","rows":2,"argmax_moved":2,"argmax_moved_non_near_tie":2,"max_abs_dp_vs_base":0.9509600418191249,"mean_of_run_mean_abs_dp_vs_base":0.6319476860721054,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true,"items":[{"request":"semif_a3f18f3a63d45345942b_with_state_of_semif_0b43ea8e24e74d621f1b","question":"answer","base":"semif_a3f18f3a63d45345942b:answer","tokens":68,"oracle_top2_margin":0.9357222628286423,"near_tie":false,"argmax_moved":true,"max_abs_dp_vs_base":0.9509600418191249,"mean_abs_dp_vs_base":0.6339733612127499,"base_probs":[0.012133949356806195,0.026071893907275722,0.9617941567359181],"perturbed_probs":[0.9478882168461089,0.041277668237097956,0.010834114916793225]},{"request":"semif_0b43ea8e24e74d621f1b_with_state_of_semif_a3f18f3a63d45345942b","question":"answer","base":"semif_0b43ea8e24e74d621f1b:answer","tokens":76,"oracle_top2_margin":0.9216656021196827,"near_tie":false,"argmax_moved":true,"max_abs_dp_vs_base":0.9448830163971914,"mean_abs_dp_vs_base":0.6299220109314609,"base_probs":[0.03331417547630556,0.9549797775959883,0.011706046927706232],"perturbed_probs":[0.02978639345379686,0.013624543221305536,0.9565890633248977]}]}]},"mac_fp16":{"rule":"red = the perturbed rows against their base rows fail the gate's own bar (FACTS §7): (a) an argmax moves on a question whose oracle top-2 margin is above 0.02, or (b) max |dp| > 0.02, or (c) the mean over the arm's rows of the row's mean |dp| > 0.002","rows_file":"$ZOO_WORK_ROOT/_d1_3b/fixtures/red_arms_r4.json","rows_file_sha256":"a1b8533efc0cea5812e1b8bb25c1d7d8ae5841b75a57359b7907873465610443","folded":null,"oracle_precheck":{"available":true,"path":"$ZOO_WORK_ROOT/_d1_3b/oracle/red_r4/records_oracle.json","sha256":"f621feaa3b4ece84b20a876a22a1aa24d8449f2bff7ea2cb57b03cd1f3bbb8af","red_on_oracle":["red_word_tv4_001","red_not_qnli_00","red_not_paws_00","state_swap_tv4_000__tv4_001","state_swap_semif_a3f18f3a63d45345942b__semif_0b43ea8e24e74d621f1b"],"replace":[],"arms":[{"id":"red_word_tv4_001","kind":"word","rows":1,"argmax_moved":1,"argmax_moved_non_near_tie":1,"max_abs_dp_vs_base":0.9764008806742758,"mean_of_run_mean_abs_dp_vs_base":0.48820044033713794,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true},{"id":"red_not_qnli_00","kind":"not","rows":1,"argmax_moved":1,"argmax_moved_non_near_tie":1,"max_abs_dp_vs_base":0.9579308844100081,"mean_of_run_mean_abs_dp_vs_base":0.9579308844100081,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true},{"id":"red_not_paws_00","kind":"not","rows":1,"argmax_moved":0,"argmax_moved_non_near_tie":0,"max_abs_dp_vs_base":0.035111090960022726,"mean_of_run_mean_abs_dp_vs_base":0.03511109096002268,"red_facts7":{"a_argmax_non_near_tie":false,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true},{"id":"state_swap_tv4_000__tv4_001","kind":"state_swap","rows":2,"argmax_moved":1,"argmax_moved_non_near_tie":1,"max_abs_dp_vs_base":0.8397602748609407,"mean_of_run_mean_abs_dp_vs_base":0.2163106998495101,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true},{"id":"state_swap_semif_a3f18f3a63d45345942b__semif_0b43ea8e24e74d621f1b","kind":"state_swap","rows":2,"argmax_moved":2,"argmax_moved_non_near_tie":2,"max_abs_dp_vs_base":0.9509600418191249,"mean_of_run_mean_abs_dp_vs_base":0.6319476860721054,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true}]},"arms":[{"id":"red_word_tv4_001","kind":"word","rows":1,"argmax_moved":1,"argmax_moved_non_near_tie":1,"max_abs_dp_vs_base":0.9764122874425372,"mean_of_run_mean_abs_dp_vs_base":0.4882061437212686,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true,"items":[{"request":"red_word_tv4_001","question":"answer","base":"tv4_001:answer","tokens":77,"oracle_top2_margin":0.9880834835402893,"near_tie":false,"argmax_moved":true,"max_abs_dp_vs_base":0.9764122874425372,"mean_abs_dp_vs_base":0.4882061437212686,"base_probs":[0.0011564291320507315,0.0030649005979227193,0.9919467820819102,0.0038318881881162898],"perturbed_probs":[0.8538234530252322,0.06819957891440163,0.015534494639373037,0.06244247342099304]}],"red_on_oracle":true,"oracle_max_abs_dp_vs_base":0.9764008806742758,"oracle_mean_of_run_mean_abs_dp_vs_base":0.48820044033713794,"graph_dp_minus_oracle_dp_max_abs":3.85183572760156e-05},{"id":"red_not_qnli_00","kind":"not","rows":1,"argmax_moved":1,"argmax_moved_non_near_tie":1,"max_abs_dp_vs_base":0.9579861952827271,"mean_of_run_mean_abs_dp_vs_base":0.9579861952827271,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true,"items":[{"request":"red_not_qnli_00","question":"answers","base":"tv4x_qnli_00:answers","tokens":106,"oracle_top2_margin":0.967676077803126,"near_tie":false,"argmax_moved":true,"max_abs_dp_vs_base":0.9579861952827271,"mean_abs_dp_vs_base":0.9579861952827271,"base_probs":[0.9838642172222498,0.016135782777750053],"perturbed_probs":[0.025878021939522825,0.9741219780604772]}],"red_on_oracle":true,"oracle_max_abs_dp_vs_base":0.9579308844100081,"oracle_mean_of_run_mean_abs_dp_vs_base":0.9579308844100081,"graph_dp_minus_oracle_dp_max_abs":5.5310872719149806e-05},{"id":"red_not_paws_00","kind":"not","rows":1,"argmax_moved":0,"argmax_moved_non_near_tie":0,"max_abs_dp_vs_base":0.034803036003128215,"mean_of_run_mean_abs_dp_vs_base":0.03480303600312814,"red_facts7":{"a_argmax_non_near_tie":false,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true,"items":[{"request":"red_not_paws_00","question":"paraphrase","base":"tv4x_paws_00:paraphrase","tokens":100,"oracle_top2_margin":0.8207418652737444,"near_tie":false,"argmax_moved":false,"max_abs_dp_vs_base":0.034803036003128215,"mean_abs_dp_vs_base":0.03480303600312814,"base_probs":[0.08982889024633645,0.9101711097536634],"perturbed_probs":[0.12463192624946466,0.8753680737505354]}],"red_on_oracle":true,"oracle_max_abs_dp_vs_base":0.035111090960022726,"oracle_mean_of_run_mean_abs_dp_vs_base":0.03511109096002268,"graph_dp_minus_oracle_dp_max_abs":0.00030805495689456563},{"id":"state_swap_tv4_000__tv4_001","kind":"state_swap","rows":2,"argmax_moved":1,"argmax_moved_non_near_tie":1,"max_abs_dp_vs_base":0.8398019974945592,"mean_of_run_mean_abs_dp_vs_base":0.21638542545499434,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true,"items":[{"request":"tv4_000_with_state_of_tv4_001","question":"answer","base":"tv4_000:answer","tokens":93,"oracle_top2_margin":0.9175063266514067,"near_tie":false,"argmax_moved":false,"max_abs_dp_vs_base":0.025739704325418167,"mean_abs_dp_vs_base":0.012869852162709073,"base_probs":[0.012278654065537022,0.030389674486733683,0.9479918105347979,0.009339860912931474],"perturbed_probs":[0.012537452804739072,0.04730357045596628,0.9222521062093797,0.017906870529914954]},{"request":"tv4_001_with_state_of_tv4_000","question":"answer","base":"tv4_001:answer","tokens":90,"oracle_top2_margin":0.9880834835402893,"near_tie":false,"argmax_moved":true,"max_abs_dp_vs_base":0.8398019974945592,"mean_abs_dp_vs_base":0.4199009987472796,"base_probs":[0.0011564291320507315,0.0030649005979227193,0.9919467820819102,0.0038318881881162898],"perturbed_probs":[0.5715075874597876,0.16206027322597968,0.15214478458735106,0.11428735472688162]}],"red_on_oracle":true,"oracle_max_abs_dp_vs_base":0.8397602748609407,"oracle_mean_of_run_mean_abs_dp_vs_base":0.2163106998495101,"graph_dp_minus_oracle_dp_max_abs":0.0002814953746580473},{"id":"state_swap_semif_a3f18f3a63d45345942b__semif_0b43ea8e24e74d621f1b","kind":"state_swap","rows":2,"argmax_moved":2,"argmax_moved_non_near_tie":2,"max_abs_dp_vs_base":0.9509847480641443,"mean_of_run_mean_abs_dp_vs_base":0.6319563667449624,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true,"items":[{"request":"semif_a3f18f3a63d45345942b_with_state_of_semif_0b43ea8e24e74d621f1b","question":"answer","base":"semif_a3f18f3a63d45345942b:answer","tokens":68,"oracle_top2_margin":0.9357222628286423,"near_tie":false,"argmax_moved":true,"max_abs_dp_vs_base":0.9509847480641443,"mean_abs_dp_vs_base":0.6339898320427628,"base_probs":[0.012136207799222254,0.026050424635484536,0.9618133675652933],"perturbed_probs":[0.9478128182244693,0.041358562274381665,0.010828619501148993]},{"request":"semif_0b43ea8e24e74d621f1b_with_state_of_semif_a3f18f3a63d45345942b","question":"answer","base":"semif_0b43ea8e24e74d621f1b:answer","tokens":76,"oracle_top2_margin":0.9216656021196827,"near_tie":false,"argmax_moved":true,"max_abs_dp_vs_base":0.9448843521707431,"mean_abs_dp_vs_base":0.629922901447162,"base_probs":[0.03331372854298944,0.9549914482878533,0.011694823169157258],"perturbed_probs":[0.029799389377367456,0.013621435282732366,0.9565791753399003]}],"red_on_oracle":true,"oracle_max_abs_dp_vs_base":0.9509600418191249,"oracle_mean_of_run_mean_abs_dp_vs_base":0.6319476860721054,"graph_dp_minus_oracle_dp_max_abs":0.0001023633090748953}],"verdict":{"required_arms":["red_word_tv4_001","red_not_qnli_00","red_not_paws_00","state_swap_tv4_000__tv4_001","state_swap_semif_a3f18f3a63d45345942b__semif_0b43ea8e24e74d621f1b"],"replace":[],"required_arms_red":true,"graph_dp_equals_oracle_dp":true,"graph_dp_minus_oracle_dp_max_abs":0.00030805495689456563,"all_arms_red":true},"base_vs_gate_run":[{"row":"tv4_001:answer","bit_equal_gate_run":true},{"row":"tv4x_qnli_00:answers","bit_equal_gate_run":true},{"row":"tv4x_paws_00:paraphrase","bit_equal_gate_run":true},{"row":"tv4_000:answer","bit_equal_gate_run":true},{"row":"semif_a3f18f3a63d45345942b:answer","bit_equal_gate_run":true},{"row":"semif_0b43ea8e24e74d621f1b:answer","bit_equal_gate_run":true}],"base_bit_equal_all":true,"all_arms_red":true,"process":[{"shard":"red_arms","rows":13,"prompts":14,"load_seconds":0.9008763750316575,"load_split_seconds":{"model":0.012568166013807058,"main":0.8883082090178505},"function_names":["main"],"process_wall_seconds":3.041991875041276,"reset_check":{"run":"tv4_001:0:base","bit_equal":true,"hidden_max_abs_diff":0.0,"wall_seconds":0.07356512499973178},"descriptor":{"function":"main","inputs":{"input_ids":[[1,64],"int32"],"position_ids":[[1,-1],"int32"],"image_embeds":[[2816,2048],"float16"]},"outputs":{"hidden":[[1,64,2048],"float16"]},"states":{"keyCache":[[8,1,8,-1,64],"float16"],"valueCache":[[8,1,8,-1,64],"float16"],"convState":[[22,1,2048,2],"float16"]}},"contract_mismatch":[]}],"runs":[{"id":"tv4_001","k":0,"variant":"base","tokens":77,"calls":2,"padded_tokens":128,"wall_seconds":0.9379728749627247,"hidden_sha256":"2be029505c60d3b2900a078e4561cecb1a19406d7772c2580317cff5822e048d","finite":true,"all_zero":false,"index":0,"shard":"red_arms","first_in_process":true},{"id":"tv4x_qnli_00","k":0,"variant":"base","tokens":105,"calls":2,"padded_tokens":128,"wall_seconds":0.07431445794645697,"hidden_sha256":"60f2ca72b9bcdec5e973ba2ab02bbd5636e55ac0755b647a2e23afd212237882","finite":true,"all_zero":false,"index":1,"shard":"red_arms","first_in_process":false},{"id":"tv4x_paws_00","k":0,"variant":"base","tokens":99,"calls":2,"padded_tokens":128,"wall_seconds":0.0738353329943493,"hidden_sha256":"2742f6f748fdb96d4b256b4bec83873ac91f158fe8374a4b9022a210b8f65fb4","finite":true,"all_zero":false,"index":2,"shard":"red_arms","first_in_process":false},{"id":"tv4_000","k":0,"variant":"base","tokens":106,"calls":2,"padded_tokens":128,"wall_seconds":0.07487074995879084,"hidden_sha256":"920b2985f95e9b42eb7a93b6d347f2959084573074ef6fc621f1ddea0dfd7698","finite":true,"all_zero":false,"index":3,"shard":"red_arms","first_in_process":false},{"id":"semif_a3f18f3a63d45345942b","k":0,"variant":"base","tokens":76,"calls":2,"padded_tokens":128,"wall_seconds":0.08220295794308186,"hidden_sha256":"59690eec4f0c13b6abbc1411ccb9bea698a48bbee6df03325c3fd7bf9f4e5ee1","finite":true,"all_zero":false,"index":4,"shard":"red_arms","first_in_process":false},{"id":"semif_0b43ea8e24e74d621f1b","k":0,"variant":"base","tokens":68,"calls":2,"padded_tokens":128,"wall_seconds":0.07370095897931606,"hidden_sha256":"4e12397edfd1a1dd8a3641ff2e8e21b6430fd5f71f2b1978faf7f0b0a43efb9a","finite":true,"all_zero":false,"index":5,"shard":"red_arms","first_in_process":false},{"id":"tv4_001","k":0,"variant":"red_word_tv4_001#0","tokens":77,"calls":2,"padded_tokens":128,"wall_seconds":0.07432845793664455,"hidden_sha256":"d1d5c1a97faa2cc5da158366477037cd2710b687c4a7b84d28df9000e0bd55fd","finite":true,"all_zero":false,"index":6,"shard":"red_arms","first_in_process":false},{"id":"tv4x_qnli_00","k":0,"variant":"red_not_qnli_00#0","tokens":106,"calls":2,"padded_tokens":128,"wall_seconds":0.07415424997452646,"hidden_sha256":"9b51f45c9e41259150a37b4eb7958889c324cdf7aaf10ed4bfe4bbfdfc0046f1","finite":true,"all_zero":false,"index":7,"shard":"red_arms","first_in_process":false},{"id":"tv4x_paws_00","k":0,"variant":"red_not_paws_00#0","tokens":100,"calls":2,"padded_tokens":128,"wall_seconds":0.07392462494317442,"hidden_sha256":"6e9d176125d7bd620a45519127ce010c578f528182b97e61c332501d12afcd48","finite":true,"all_zero":false,"index":8,"shard":"red_arms","first_in_process":false},{"id":"tv4_000","k":0,"variant":"state_swap_tv4_000__tv4_001#0","tokens":93,"calls":2,"padded_tokens":128,"wall_seconds":0.07321775006130338,"hidden_sha256":"4a4423fca2d78e8809b6096d42dec64cb7c47d0047e970172f0d02dbc4bf79d4","finite":true,"all_zero":false,"index":9,"shard":"red_arms","first_in_process":false},{"id":"tv4_001","k":0,"variant":"state_swap_tv4_000__tv4_001#1","tokens":90,"calls":2,"padded_tokens":128,"wall_seconds":0.07421275007072836,"hidden_sha256":"65c950d60a526e2eaabe50828d6a71f33cf0044bbc109c2e1e66a447c823c8e4","finite":true,"all_zero":false,"index":10,"shard":"red_arms","first_in_process":false},{"id":"semif_a3f18f3a63d45345942b","k":0,"variant":"state_swap_semif_a3f18f3a63d45345942b__semif_0b43ea8e24e74d621f1b#0","tokens":68,"calls":2,"padded_tokens":128,"wall_seconds":0.07346637500450015,"hidden_sha256":"1ed8a33ff8d95207d183a697514e60943bc03621bf7faa7e5454b700af6b3136","finite":true,"all_zero":false,"index":11,"shard":"red_arms","first_in_process":false},{"id":"semif_0b43ea8e24e74d621f1b","k":0,"variant":"state_swap_semif_a3f18f3a63d45345942b__semif_0b43ea8e24e74d621f1b#1","tokens":76,"calls":2,"padded_tokens":128,"wall_seconds":0.07362095906864852,"hidden_sha256":"38e3d0f86b3e6f3029406f0e530ba6c1bfd7e392ca6e5c6bc1c894aa9bd3d869","finite":true,"all_zero":false,"index":12,"shard":"red_arms","first_in_process":false}],"reset_bit_equal":true},"iphone_int8mlp":{"rule":"red = the perturbed rows against their base rows fail the gate's own bar (FACTS §7): (a) an argmax moves on a question whose oracle top-2 margin is above 0.02, or (b) max |dp| > 0.02, or (c) the mean over the arm's rows of the row's mean |dp| > 0.002","rows_file":"$ZOO_WORK_ROOT/_d1_3b/fixtures/red_arms_r4.json","rows_file_sha256":"a1b8533efc0cea5812e1b8bb25c1d7d8ae5841b75a57359b7907873465610443","folded":null,"oracle_precheck":{"available":true,"path":"$ZOO_WORK_ROOT/_d1_3b/oracle/red_r4/records_oracle.json","sha256":"f621feaa3b4ece84b20a876a22a1aa24d8449f2bff7ea2cb57b03cd1f3bbb8af","red_on_oracle":["red_word_tv4_001","red_not_qnli_00","red_not_paws_00","state_swap_tv4_000__tv4_001","state_swap_semif_a3f18f3a63d45345942b__semif_0b43ea8e24e74d621f1b"],"replace":[],"arms":[{"id":"red_word_tv4_001","kind":"word","rows":1,"argmax_moved":1,"argmax_moved_non_near_tie":1,"max_abs_dp_vs_base":0.9764008806742758,"mean_of_run_mean_abs_dp_vs_base":0.48820044033713794,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true},{"id":"red_not_qnli_00","kind":"not","rows":1,"argmax_moved":1,"argmax_moved_non_near_tie":1,"max_abs_dp_vs_base":0.9579308844100081,"mean_of_run_mean_abs_dp_vs_base":0.9579308844100081,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true},{"id":"red_not_paws_00","kind":"not","rows":1,"argmax_moved":0,"argmax_moved_non_near_tie":0,"max_abs_dp_vs_base":0.035111090960022726,"mean_of_run_mean_abs_dp_vs_base":0.03511109096002268,"red_facts7":{"a_argmax_non_near_tie":false,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true},{"id":"state_swap_tv4_000__tv4_001","kind":"state_swap","rows":2,"argmax_moved":1,"argmax_moved_non_near_tie":1,"max_abs_dp_vs_base":0.8397602748609407,"mean_of_run_mean_abs_dp_vs_base":0.2163106998495101,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true},{"id":"state_swap_semif_a3f18f3a63d45345942b__semif_0b43ea8e24e74d621f1b","kind":"state_swap","rows":2,"argmax_moved":2,"argmax_moved_non_near_tie":2,"max_abs_dp_vs_base":0.9509600418191249,"mean_of_run_mean_abs_dp_vs_base":0.6319476860721054,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true}]},"arms":[{"id":"red_word_tv4_001","kind":"word","rows":1,"argmax_moved":1,"argmax_moved_non_near_tie":1,"max_abs_dp_vs_base":0.9765624499703999,"mean_of_run_mean_abs_dp_vs_base":0.48828122498519994,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true,"items":[{"request":"red_word_tv4_001","question":"answer","base":"tv4_001:answer","tokens":77,"oracle_top2_margin":0.9880834835402893,"near_tie":false,"argmax_moved":true,"max_abs_dp_vs_base":0.9765624499703999,"mean_abs_dp_vs_base":0.48828122498519994,"base_probs":[0.0011469111242165457,0.00304382684147775,0.9919810191023674,0.003828242931938439],"perturbed_probs":[0.85528529981968,0.06756202074063145,0.015418569131967454,0.06173411030772101]}],"red_on_oracle":true,"oracle_max_abs_dp_vs_base":0.9764008806742758,"oracle_mean_of_run_mean_abs_dp_vs_base":0.48820044033713794,"graph_dp_minus_oracle_dp_max_abs":0.0015030645735707449},{"id":"red_not_qnli_00","kind":"not","rows":1,"argmax_moved":1,"argmax_moved_non_near_tie":1,"max_abs_dp_vs_base":0.9575841443308328,"mean_of_run_mean_abs_dp_vs_base":0.9575841443308328,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true,"items":[{"request":"red_not_qnli_00","question":"answers","base":"tv4x_qnli_00:answers","tokens":106,"oracle_top2_margin":0.967676077803126,"near_tie":false,"argmax_moved":true,"max_abs_dp_vs_base":0.9575841443308328,"mean_abs_dp_vs_base":0.9575841443308328,"base_probs":[0.9838779251926246,0.016122074807375302],"perturbed_probs":[0.026293780861791797,0.9737062191382081]}],"red_on_oracle":true,"oracle_max_abs_dp_vs_base":0.9579308844100081,"oracle_mean_of_run_mean_abs_dp_vs_base":0.9579308844100081,"graph_dp_minus_oracle_dp_max_abs":0.00034674007917523397},{"id":"red_not_paws_00","kind":"not","rows":1,"argmax_moved":0,"argmax_moved_non_near_tie":0,"max_abs_dp_vs_base":0.035394121322759875,"mean_of_run_mean_abs_dp_vs_base":0.035394121322759854,"red_facts7":{"a_argmax_non_near_tie":false,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true,"items":[{"request":"red_not_paws_00","question":"paraphrase","base":"tv4x_paws_00:paraphrase","tokens":100,"oracle_top2_margin":0.8207418652737444,"near_tie":false,"argmax_moved":false,"max_abs_dp_vs_base":0.035394121322759875,"mean_abs_dp_vs_base":0.035394121322759854,"base_probs":[0.08875337971759,0.9112466202824101],"perturbed_probs":[0.12414750104034983,0.8758524989596502]}],"red_on_oracle":true,"oracle_max_abs_dp_vs_base":0.035111090960022726,"oracle_mean_of_run_mean_abs_dp_vs_base":0.03511109096002268,"graph_dp_minus_oracle_dp_max_abs":0.0002830303627372466},{"id":"state_swap_tv4_000__tv4_001","kind":"state_swap","rows":2,"argmax_moved":1,"argmax_moved_non_near_tie":1,"max_abs_dp_vs_base":0.8401766292879931,"mean_of_run_mean_abs_dp_vs_base":0.21642369586438726,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true,"items":[{"request":"tv4_000_with_state_of_tv4_001","question":"answer","base":"tv4_000:answer","tokens":93,"oracle_top2_margin":0.9175063266514067,"near_tie":false,"argmax_moved":false,"max_abs_dp_vs_base":0.02551815416955594,"mean_abs_dp_vs_base":0.012759077084777975,"base_probs":[0.012317559602245474,0.030557784835101284,0.9478103831813423,0.009314272381311035],"perturbed_probs":[0.012567130905940154,0.04710226475717062,0.9222922290117863,0.018038375325102978]},{"request":"tv4_001_with_state_of_tv4_000","question":"answer","base":"tv4_001:answer","tokens":90,"oracle_top2_margin":0.9880834835402893,"near_tie":false,"argmax_moved":true,"max_abs_dp_vs_base":0.8401766292879931,"mean_abs_dp_vs_base":0.42008831464399654,"base_probs":[0.0011469111242165457,0.00304382684147775,0.9919810191023674,0.003828242931938439],"perturbed_probs":[0.5718904048444216,0.16060704853340757,0.15180438981437427,0.11569815680779665]}],"red_on_oracle":true,"oracle_max_abs_dp_vs_base":0.8397602748609407,"oracle_mean_of_run_mean_abs_dp_vs_base":0.2163106998495101,"graph_dp_minus_oracle_dp_max_abs":0.0017003170869772655},{"id":"state_swap_semif_a3f18f3a63d45345942b__semif_0b43ea8e24e74d621f1b","kind":"state_swap","rows":2,"argmax_moved":2,"argmax_moved_non_near_tie":2,"max_abs_dp_vs_base":0.9511345596731133,"mean_of_run_mean_abs_dp_vs_base":0.6320729910454614,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true,"items":[{"request":"semif_a3f18f3a63d45345942b_with_state_of_semif_0b43ea8e24e74d621f1b","question":"answer","base":"semif_a3f18f3a63d45345942b:answer","tokens":68,"oracle_top2_margin":0.9357222628286423,"near_tie":false,"argmax_moved":true,"max_abs_dp_vs_base":0.9511345596731133,"mean_abs_dp_vs_base":0.6340897064487422,"base_probs":[0.012088734354520582,0.025948674535824997,0.9619625911096544],"perturbed_probs":[0.9483250672076603,0.0408469013557986,0.010828031436541088]},{"request":"semif_0b43ea8e24e74d621f1b_with_state_of_semif_a3f18f3a63d45345942b","question":"answer","base":"semif_0b43ea8e24e74d621f1b:answer","tokens":76,"oracle_top2_margin":0.9216656021196827,"near_tie":false,"argmax_moved":true,"max_abs_dp_vs_base":0.9450844134632711,"mean_abs_dp_vs_base":0.6300562756421808,"base_probs":[0.03325470136169396,0.9550748562447832,0.011670442393522872],"perturbed_probs":[0.029697187132875173,0.013547957010330723,0.956754855856794]}],"red_on_oracle":true,"oracle_max_abs_dp_vs_base":0.9509600418191249,"oracle_mean_of_run_mean_abs_dp_vs_base":0.6319476860721054,"graph_dp_minus_oracle_dp_max_abs":0.00048206536383699206}],"verdict":{"required_arms":["red_word_tv4_001","red_not_qnli_00","red_not_paws_00","state_swap_tv4_000__tv4_001","state_swap_semif_a3f18f3a63d45345942b__semif_0b43ea8e24e74d621f1b"],"replace":[],"required_arms_red":true,"graph_dp_equals_oracle_dp":true,"graph_dp_minus_oracle_dp_max_abs":0.0017003170869772655,"all_arms_red":true},"base_vs_gate_run":[{"row":"tv4_001:answer","bit_equal_gate_run":true},{"row":"tv4x_qnli_00:answers","bit_equal_gate_run":true},{"row":"tv4x_paws_00:paraphrase","bit_equal_gate_run":true},{"row":"tv4_000:answer","bit_equal_gate_run":true},{"row":"semif_a3f18f3a63d45345942b:answer","bit_equal_gate_run":true},{"row":"semif_0b43ea8e24e74d621f1b:answer","bit_equal_gate_run":true}],"base_bit_equal_all":true,"all_arms_red":true,"process":[{"shard":"red_arms","rows":13,"prompts":14,"load_seconds":0.6527015829924494,"load_split_seconds":{"model":0.006853541010059416,"main":0.64584804198239},"function_names":["main"],"process_wall_seconds":2.870171125046909,"reset_check":{"run":"tv4_001:0:base","bit_equal":true,"hidden_max_abs_diff":0.0,"wall_seconds":0.07374400005210191},"descriptor":{"function":"main","inputs":{"input_ids":[[1,64],"int32"],"position_ids":[[1,-1],"int32"],"image_embeds":[[2816,2048],"float16"]},"outputs":{"hidden":[[1,64,2048],"float16"]},"states":{"keyCache":[[8,1,8,-1,64],"float16"],"valueCache":[[8,1,8,-1,64],"float16"],"convState":[[22,1,2048,2],"float16"]}},"contract_mismatch":[]}],"runs":[{"id":"tv4_001","k":0,"variant":"base","tokens":77,"calls":2,"padded_tokens":128,"wall_seconds":1.013376416056417,"hidden_sha256":"8029f6a9b7a06011fd9a512bbdf101e1972cab492a2182f5ac5db15a5cd51973","finite":true,"all_zero":false,"index":0,"shard":"red_arms","first_in_process":true},{"id":"tv4x_qnli_00","k":0,"variant":"base","tokens":105,"calls":2,"padded_tokens":128,"wall_seconds":0.07578974997159094,"hidden_sha256":"3474bbba42c377711c5f83a0fd6e3fbd331f74ca79d9f46358c7777d4cf3e4c7","finite":true,"all_zero":false,"index":1,"shard":"red_arms","first_in_process":false},{"id":"tv4x_paws_00","k":0,"variant":"base","tokens":99,"calls":2,"padded_tokens":128,"wall_seconds":0.07801658299285918,"hidden_sha256":"3cfec852f49795996e3748693bcc3e7f10f150c4ebbb41d49c042129bdef39f6","finite":true,"all_zero":false,"index":2,"shard":"red_arms","first_in_process":false},{"id":"tv4_000","k":0,"variant":"base","tokens":106,"calls":2,"padded_tokens":128,"wall_seconds":0.07493424997664988,"hidden_sha256":"94bccfd2a7bbc3fedc2c7d7a6ac3570dbf25b5a9f617eeba4fb5364bf8e8aafd","finite":true,"all_zero":false,"index":3,"shard":"red_arms","first_in_process":false},{"id":"semif_a3f18f3a63d45345942b","k":0,"variant":"base","tokens":76,"calls":2,"padded_tokens":128,"wall_seconds":0.07935554208233953,"hidden_sha256":"e868b90f7ae7e554539fa902dbf37c7727c69cf2117728a97cf5df5a498c0f7f","finite":true,"all_zero":false,"index":4,"shard":"red_arms","first_in_process":false},{"id":"semif_0b43ea8e24e74d621f1b","k":0,"variant":"base","tokens":68,"calls":2,"padded_tokens":128,"wall_seconds":0.07557320792693645,"hidden_sha256":"28c0dcdd954cfd122a34057a525d39b89ff378846431fef6d5a937cd801fa367","finite":true,"all_zero":false,"index":5,"shard":"red_arms","first_in_process":false},{"id":"tv4_001","k":0,"variant":"red_word_tv4_001#0","tokens":77,"calls":2,"padded_tokens":128,"wall_seconds":0.07382299995515496,"hidden_sha256":"5fc40cf5578660b11f8487819581db5d79cae51212b9a016412cdb3d91c4e0ea","finite":true,"all_zero":false,"index":6,"shard":"red_arms","first_in_process":false},{"id":"tv4x_qnli_00","k":0,"variant":"red_not_qnli_00#0","tokens":106,"calls":2,"padded_tokens":128,"wall_seconds":0.07428858405910432,"hidden_sha256":"aaa15159e8885317b8422be711d3027a9cd361bbfab9a2be25450dac41d4dd13","finite":true,"all_zero":false,"index":7,"shard":"red_arms","first_in_process":false},{"id":"tv4x_paws_00","k":0,"variant":"red_not_paws_00#0","tokens":100,"calls":2,"padded_tokens":128,"wall_seconds":0.07482279208488762,"hidden_sha256":"5421ca394834051df1660070e5ae2c00c7798a775940776500330a77c97d2fbc","finite":true,"all_zero":false,"index":8,"shard":"red_arms","first_in_process":false},{"id":"tv4_000","k":0,"variant":"state_swap_tv4_000__tv4_001#0","tokens":93,"calls":2,"padded_tokens":128,"wall_seconds":0.07339220808353275,"hidden_sha256":"357e12a41fc80df414ffc533bd800a9be63c14177f0b1eb1b231958bc61ce324","finite":true,"all_zero":false,"index":9,"shard":"red_arms","first_in_process":false},{"id":"tv4_001","k":0,"variant":"state_swap_tv4_000__tv4_001#1","tokens":90,"calls":2,"padded_tokens":128,"wall_seconds":0.07387891702819616,"hidden_sha256":"7e45badf17da395bfd1ff3f89ba181d20e9e693df5b1172267ff0c1f48f8ec9c","finite":true,"all_zero":false,"index":10,"shard":"red_arms","first_in_process":false},{"id":"semif_a3f18f3a63d45345942b","k":0,"variant":"state_swap_semif_a3f18f3a63d45345942b__semif_0b43ea8e24e74d621f1b#0","tokens":68,"calls":2,"padded_tokens":128,"wall_seconds":0.07579683396033943,"hidden_sha256":"c5eb97289ba95f41343ed003054ab6ff5a8375ea2c39247413832d65e2539291","finite":true,"all_zero":false,"index":11,"shard":"red_arms","first_in_process":false},{"id":"semif_0b43ea8e24e74d621f1b","k":0,"variant":"state_swap_semif_a3f18f3a63d45345942b__semif_0b43ea8e24e74d621f1b#1","tokens":76,"calls":2,"padded_tokens":128,"wall_seconds":0.07368837494868785,"hidden_sha256":"36bf588fb428e86ed5a234aa5cd3a065965b88a118cc8cf2ba07e9b1d8a72c8b","finite":true,"all_zero":false,"index":12,"shard":"red_arms","first_in_process":false}],"reset_bit_equal":true},"fp16_S16":{"schema":"d1-decoder-readout-red-arms/1","red_rule":"red = the perturbed rows against their base rows fail the gate's own bar (FACTS §7): (a) an argmax moves on a question whose oracle top-2 margin is above 0.02, or (b) max |dp| > 0.02, or (c) the mean over the arm's rows of the row's mean |dp| > 0.002","bar":{"max_abs_dp":0.02,"mean_of_run_mean_abs_dp":0.002,"near_tie_top2_margin":0.02,"argmax":"every question with an oracle top-2 margin above 0.02; near-ties listed apart","max_abs_dp_applies_to":"every option of every question, near-ties included","mean_of_run_mean_abs_dp_definition":"mean over runs (one run = one row = one question) of the mean |dp| over that question's options","reset":"every process re-runs its first row last: hidden rows bit-equal","finite":"every hidden value of every row finite, no row all zero","rows":"every expected row present"},"checks":{"red_required_arms_red":true,"red_graph_dp_equals_oracle_dp":true,"red_base_bit_equal_gate_runs":true,"red_reset_bit_equal":true},"result":"PASS","arms_file":"$ZOO_WORK_ROOT/_d1_3b/fixtures/red_arms_r4.json","arms_sha256":"a1b8533efc0cea5812e1b8bb25c1d7d8ae5841b75a57359b7907873465610443","bundle":{"bundle":"$ZOO_WORK_ROOT/_d1_3b/exports/bundles/d1_3b_decode_fp16_pf16","name":"d1_3b_decode_fp16_pf16","kind":"decision-backbone","compression":null,"prefill_chunk":16,"max_ctx":4096,"toy":null,"metadata_sha256":"a99c6207a80059467a7396db0de8b8bfac6e61381308b0531fc1c3a42ee0ac4d","main_mlirb":{"bytes":5562548109,"sha256":"a16fb8b2bdc03edaec500a60ba9fc63a0a6b3c117d62e8e378380f0bf8d8b257"},"tokenizer_sha256":{"chat_template.jinja":"86f4770449a4797c9b4212d110b0cc70fb993c1ce960095bcfb12eea22f61cca","tokenizer.json":"8096ecb9f54599d756c8de728a598a340bc1e43c0deb77ddd62456c38349fcee","tokenizer_config.json":"ef6770d12dd9ac58d334e979bf160c5cdfd41f13d205fd5d5516857645feb0ea"},"head_sha256":{"option_rows.json":"632921e0e32dbc99cc704a08cefc7bc3b0ed03677c074d055b9bb83aff95177c","option_rows.safetensors":"c17bfc6cdee99c8fd9416a692ac2bf3813a5f808dab8a3c8c2754fcff75b4ca9"},"license_sha256":"4d28ca14dedc0b3d0fcc2b3339f0e79931faa33874f3d24f522183a8fc70068c","aimodelc":{"path":"$ZOO_WORK_ROOT/_d1_3b/exports/bundles_aotc/d1_3b_decode_fp16_pf16.h16c.aimodelc","bytes":10600779202,"tree_sha256":"b3dbd18ee502ccdac998ee968a25b339a5a6550ca890e7ebbd682c55172e247d"},"aimodelc_main_hash":"946d898f56a096da8929690af83d54bd435ede35289c234540b1fa0eec48399c"},"gate_transcript":{"path":"$ZOO_WORK_ROOT/_d1_3b/results/r4_readout_fp16_pf16.json","sha256":"80b83663638c12e3319f48bde78f34e05fffde9cc797b7d9c20f7b14dd44ba1f","result":"PASS"},"oracle":{"path":"$ZOO_WORK_ROOT/_d1_3b/oracle/records_oracle.json","sha256":"067a4f5658937da0e20e307deaa15a38deb55d0f12b2f9993b609461d0430af1"},"red_oracle":{"path":"$ZOO_WORK_ROOT/_d1_3b/oracle/red_r4/records_oracle.json","sha256":"f621feaa3b4ece84b20a876a22a1aa24d8449f2bff7ea2cb57b03cd1f3bbb8af"},"script":{"path":"$ZOO_WORK_ROOT/_d1_3b/scripts/r4_red_graph.py","sha256":"4658796e976b2802381d0d9b7a542ed8646ddb9db5f6d3239aade2d13825a5d6","readout_gate_sha256":"83b887e66d0a31df2b47a79605bbab2a687e92a3a4fdca10bcde3034630251a4"},"environment":{"versions":{"python":"3.11.13","numpy":"2.3.5","coreai-core":"1.0.0b2","coreai-torch":"0.4.1","coreai-models":"0.1.0","torch":"2.9.0","safetensors":"0.7.0","tokenizers":"0.23.0rc0"},"platform":"macOS-27.0-arm64-arm-64bit","macos_build":"26A428","chip":"Apple M4 Max","runtime":"coreai python runtime, AOT h16c GPU .aimodelc, SpecializationOptions.default(), no JIT","readout":"host.py: z = h_slot . E[id] in float64 over the group ids (head/option_rows, fp32), group max, softmax over the options","gpu":"shared with other sessions, _GPU_LOCK read only (the ms are contended reference values)"},"gpu_lock":{"start":{"path":"$ZOO_WORK_ROOT/_GPU_LOCK","exists":true,"bytes":0,"content":"","mtime":"2026-10-08T10:47:57+09:00"},"end":{"path":"$ZOO_WORK_ROOT/_GPU_LOCK","exists":true,"bytes":0,"content":"","mtime":"2026-10-08T10:47:57+09:00"}},"red_arm":{"rule":"red = the perturbed rows against their base rows fail the gate's own bar (FACTS §7): (a) an argmax moves on a question whose oracle top-2 margin is above 0.02, or (b) max |dp| > 0.02, or (c) the mean over the arm's rows of the row's mean |dp| > 0.002","rows_file":"$ZOO_WORK_ROOT/_d1_3b/fixtures/red_arms_r4.json","rows_file_sha256":"a1b8533efc0cea5812e1b8bb25c1d7d8ae5841b75a57359b7907873465610443","folded":null,"oracle_precheck":{"available":true,"path":"$ZOO_WORK_ROOT/_d1_3b/oracle/red_r4/records_oracle.json","sha256":"f621feaa3b4ece84b20a876a22a1aa24d8449f2bff7ea2cb57b03cd1f3bbb8af","red_on_oracle":["red_word_tv4_001","red_not_qnli_00","red_not_paws_00","state_swap_tv4_000__tv4_001","state_swap_semif_a3f18f3a63d45345942b__semif_0b43ea8e24e74d621f1b"],"replace":[],"arms":[{"id":"red_word_tv4_001","kind":"word","rows":1,"argmax_moved":1,"argmax_moved_non_near_tie":1,"max_abs_dp_vs_base":0.9764008806742758,"mean_of_run_mean_abs_dp_vs_base":0.48820044033713794,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true},{"id":"red_not_qnli_00","kind":"not","rows":1,"argmax_moved":1,"argmax_moved_non_near_tie":1,"max_abs_dp_vs_base":0.9579308844100081,"mean_of_run_mean_abs_dp_vs_base":0.9579308844100081,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true},{"id":"red_not_paws_00","kind":"not","rows":1,"argmax_moved":0,"argmax_moved_non_near_tie":0,"max_abs_dp_vs_base":0.035111090960022726,"mean_of_run_mean_abs_dp_vs_base":0.03511109096002268,"red_facts7":{"a_argmax_non_near_tie":false,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true},{"id":"state_swap_tv4_000__tv4_001","kind":"state_swap","rows":2,"argmax_moved":1,"argmax_moved_non_near_tie":1,"max_abs_dp_vs_base":0.8397602748609407,"mean_of_run_mean_abs_dp_vs_base":0.2163106998495101,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true},{"id":"state_swap_semif_a3f18f3a63d45345942b__semif_0b43ea8e24e74d621f1b","kind":"state_swap","rows":2,"argmax_moved":2,"argmax_moved_non_near_tie":2,"max_abs_dp_vs_base":0.9509600418191249,"mean_of_run_mean_abs_dp_vs_base":0.6319476860721054,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true}]},"arms":[{"id":"red_word_tv4_001","kind":"word","rows":1,"argmax_moved":1,"argmax_moved_non_near_tie":1,"max_abs_dp_vs_base":0.9763672319939549,"mean_of_run_mean_abs_dp_vs_base":0.4881836159969775,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true,"items":[{"request":"red_word_tv4_001","question":"answer","base":"tv4_001:answer","tokens":77,"oracle_top2_margin":0.9880834835402893,"near_tie":false,"argmax_moved":true,"max_abs_dp_vs_base":0.9763672319939549,"mean_abs_dp_vs_base":0.4881836159969775,"base_probs":[0.0011609716293289674,0.003075012718445317,0.9919193746628899,0.0038446409893356528],"perturbed_probs":[0.8536519406594565,0.0682832025367294,0.01555214266893501,0.06251271413487915]}],"red_on_oracle":true,"oracle_max_abs_dp_vs_base":0.9764008806742758,"oracle_mean_of_run_mean_abs_dp_vs_base":0.48820044033713794,"graph_dp_minus_oracle_dp_max_abs":0.00014435509176524697},{"id":"red_not_qnli_00","kind":"not","rows":1,"argmax_moved":1,"argmax_moved_non_near_tie":1,"max_abs_dp_vs_base":0.9579304094726216,"mean_of_run_mean_abs_dp_vs_base":0.9579304094726216,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true,"items":[{"request":"red_not_qnli_00","question":"answers","base":"tv4x_qnli_00:answers","tokens":106,"oracle_top2_margin":0.967676077803126,"near_tie":false,"argmax_moved":true,"max_abs_dp_vs_base":0.9579304094726216,"mean_abs_dp_vs_base":0.9579304094726216,"base_probs":[0.9838119782631385,0.01618802173686154],"perturbed_probs":[0.025881568790516844,0.9741184312094832]}],"red_on_oracle":true,"oracle_max_abs_dp_vs_base":0.9579308844100081,"oracle_mean_of_run_mean_abs_dp_vs_base":0.9579308844100081,"graph_dp_minus_oracle_dp_max_abs":4.7493738641879446e-07},{"id":"red_not_paws_00","kind":"not","rows":1,"argmax_moved":0,"argmax_moved_non_near_tie":0,"max_abs_dp_vs_base":0.03486987067280267,"mean_of_run_mean_abs_dp_vs_base":0.034869870672802644,"red_facts7":{"a_argmax_non_near_tie":false,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true,"items":[{"request":"red_not_paws_00","question":"paraphrase","base":"tv4x_paws_00:paraphrase","tokens":100,"oracle_top2_margin":0.8207418652737444,"near_tie":false,"argmax_moved":false,"max_abs_dp_vs_base":0.03486987067280267,"mean_abs_dp_vs_base":0.034869870672802644,"base_probs":[0.08997967460709025,0.9100203253929098],"perturbed_probs":[0.12484954527989287,0.8751504547201071]}],"red_on_oracle":true,"oracle_max_abs_dp_vs_base":0.035111090960022726,"oracle_mean_of_run_mean_abs_dp_vs_base":0.03511109096002268,"graph_dp_minus_oracle_dp_max_abs":0.00024122028722010913},{"id":"state_swap_tv4_000__tv4_001","kind":"state_swap","rows":2,"argmax_moved":1,"argmax_moved_non_near_tie":1,"max_abs_dp_vs_base":0.8394665877851488,"mean_of_run_mean_abs_dp_vs_base":0.2162583241602083,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true,"items":[{"request":"tv4_000_with_state_of_tv4_001","question":"answer","base":"tv4_000:answer","tokens":93,"oracle_top2_margin":0.9175063266514067,"near_tie":false,"argmax_moved":false,"max_abs_dp_vs_base":0.02556670885568424,"mean_abs_dp_vs_base":0.012783354427842113,"base_probs":[0.012282638353939869,0.030371554374588595,0.9480028882703195,0.009342919001152066],"perturbed_probs":[0.012534724146168002,0.047175741943751806,0.9224361794146353,0.017853354495444937]},{"request":"tv4_001_with_state_of_tv4_000","question":"answer","base":"tv4_001:answer","tokens":90,"oracle_top2_margin":0.9880834835402893,"near_tie":false,"argmax_moved":true,"max_abs_dp_vs_base":0.8394665877851488,"mean_abs_dp_vs_base":0.41973329389257447,"base_probs":[0.0011609716293289674,0.003075012718445317,0.9919193746628899,0.0038446409893356528],"perturbed_probs":[0.571103034236305,0.1618026810099064,0.15245278687774116,0.1146414978760474]}],"red_on_oracle":true,"oracle_max_abs_dp_vs_base":0.8397602748609407,"oracle_mean_of_run_mean_abs_dp_vs_base":0.2163106998495101,"graph_dp_minus_oracle_dp_max_abs":0.0005358704874459941},{"id":"state_swap_semif_a3f18f3a63d45345942b__semif_0b43ea8e24e74d621f1b","kind":"state_swap","rows":2,"argmax_moved":2,"argmax_moved_non_near_tie":2,"max_abs_dp_vs_base":0.9509367636006592,"mean_of_run_mean_abs_dp_vs_base":0.6319516513056637,"red_facts7":{"a_argmax_non_near_tie":true,"b_max_abs_dp":true,"c_mean_of_run_means":true},"red":true,"red_fixture_rule":true,"items":[{"request":"semif_a3f18f3a63d45345942b_with_state_of_semif_0b43ea8e24e74d621f1b","question":"answer","base":"semif_a3f18f3a63d45345942b:answer","tokens":68,"oracle_top2_margin":0.9357222628286423,"near_tie":false,"argmax_moved":true,"max_abs_dp_vs_base":0.9509367636006592,"mean_abs_dp_vs_base":0.6339578424004394,"base_probs":[0.0121417814891555,0.026097008273222495,0.961761210237622],"perturbed_probs":[0.9478735608387867,0.041301992524250535,0.010824446636962783]},{"request":"semif_0b43ea8e24e74d621f1b_with_state_of_semif_a3f18f3a63d45345942b","question":"answer","base":"semif_0b43ea8e24e74d621f1b:answer","tokens":76,"oracle_top2_margin":0.9216656021196827,"near_tie":false,"argmax_moved":true,"max_abs_dp_vs_base":0.944918190316332,"mean_abs_dp_vs_base":0.6299454602108879,"base_probs":[0.03329006057297575,0.9549991990928817,0.011710740334142596],"perturbed_probs":[0.029775218575117817,0.013595850774407787,0.9566289306504745]}],"red_on_oracle":true,"oracle_max_abs_dp_vs_base":0.9509600418191249,"oracle_mean_of_run_mean_abs_dp_vs_base":0.6319476860721054,"graph_dp_minus_oracle_dp_max_abs":4.811394379111711e-05}],"verdict":{"required_arms":["red_word_tv4_001","red_not_qnli_00","red_not_paws_00","state_swap_tv4_000__tv4_001","state_swap_semif_a3f18f3a63d45345942b__semif_0b43ea8e24e74d621f1b"],"replace":[],"required_arms_red":true,"graph_dp_equals_oracle_dp":true,"graph_dp_minus_oracle_dp_max_abs":0.0005358704874459941,"all_arms_red":true},"base_vs_gate_run":[{"row":"tv4_001:answer","bit_equal_gate_run":true},{"row":"tv4x_qnli_00:answers","bit_equal_gate_run":true},{"row":"tv4x_paws_00:paraphrase","bit_equal_gate_run":true},{"row":"tv4_000:answer","bit_equal_gate_run":true},{"row":"semif_a3f18f3a63d45345942b:answer","bit_equal_gate_run":true},{"row":"semif_0b43ea8e24e74d621f1b:answer","bit_equal_gate_run":true}],"base_bit_equal_all":true,"all_arms_red":true,"process":[{"shard":"red_arms_r4","rows":13,"prompts":14,"load_seconds":1.249408375006169,"load_split_seconds":{"model":0.007713833008892834,"main":1.2416945419972762},"function_names":["main"],"process_wall_seconds":6.77336250001099,"reset_check":{"run":"tv4_001:0:base","bit_equal":true,"hidden_max_abs_diff":0.0,"wall_seconds":0.2700938330963254},"descriptor":{"function":"main","inputs":{"input_ids":[[1,16],"int32"],"position_ids":[[1,-1],"int32"],"image_embeds":[[2816,2048],"float16"]},"outputs":{"hidden":[[1,16,2048],"float16"]},"states":{"keyCache":[[8,1,8,-1,64],"float16"],"valueCache":[[8,1,8,-1,64],"float16"],"convState":[[22,1,2048,2],"float16"]}},"contract_mismatch":[]}],"runs":[{"id":"tv4_001","k":0,"variant":"base","tokens":77,"calls":5,"padded_tokens":80,"wall_seconds":1.4157865840243176,"hidden_sha256":"7522ff38d85cb2771a2cce9f861e17d1621a512f0f634d931c27867350d50e56","finite":true,"all_zero":false,"index":0,"shard":"red_arms_r4","first_in_process":true},{"id":"tv4x_qnli_00","k":0,"variant":"base","tokens":105,"calls":7,"padded_tokens":112,"wall_seconds":0.16924208297859877,"hidden_sha256":"16507748c25d47311d31de7a7e2b7a0d91d1525d30d096174b3cb2adf77eb7b0","finite":true,"all_zero":false,"index":1,"shard":"red_arms_r4","first_in_process":false},{"id":"tv4x_paws_00","k":0,"variant":"base","tokens":99,"calls":7,"padded_tokens":112,"wall_seconds":0.1672036669915542,"hidden_sha256":"c8669faad6f946f6e462be7b99cffa66d3c4ab4750bacebf4c29e45bf99a7b4f","finite":true,"all_zero":false,"index":2,"shard":"red_arms_r4","first_in_process":false},{"id":"tv4_000","k":0,"variant":"base","tokens":106,"calls":7,"padded_tokens":112,"wall_seconds":0.1952197919599712,"hidden_sha256":"0eb1301e87923e46cea0381cd7608b023660630e8a1e9dd07ad5e8ab52b582f9","finite":true,"all_zero":false,"index":3,"shard":"red_arms_r4","first_in_process":false},{"id":"semif_a3f18f3a63d45345942b","k":0,"variant":"base","tokens":76,"calls":5,"padded_tokens":80,"wall_seconds":0.13357179204467684,"hidden_sha256":"8de4a2c56b67fd9bccbddbd5e975b5e5c574d35b8b07e2e1bc7eb802ca8075e2","finite":true,"all_zero":false,"index":4,"shard":"red_arms_r4","first_in_process":false},{"id":"semif_0b43ea8e24e74d621f1b","k":0,"variant":"base","tokens":68,"calls":5,"padded_tokens":80,"wall_seconds":0.1524958749068901,"hidden_sha256":"0ae0c5a64be7eb675e103afec66c5ba66b14c142999e10cda120232c9efde844","finite":true,"all_zero":false,"index":5,"shard":"red_arms_r4","first_in_process":false},{"id":"tv4_001","k":0,"variant":"red_word_tv4_001#0","tokens":77,"calls":5,"padded_tokens":80,"wall_seconds":0.3216776669723913,"hidden_sha256":"b5afb32e4af0f437a04018e86a5395c1d03fadbb244a499add6707fa9f8de226","finite":true,"all_zero":false,"index":6,"shard":"red_arms_r4","first_in_process":false},{"id":"tv4x_qnli_00","k":0,"variant":"red_not_qnli_00#0","tokens":106,"calls":7,"padded_tokens":112,"wall_seconds":0.4903731249505654,"hidden_sha256":"0a6e261510b3c7591d6b081415f09dd28ca90145309b107426e256b88ab75ab6","finite":true,"all_zero":false,"index":7,"shard":"red_arms_r4","first_in_process":false},{"id":"tv4x_paws_00","k":0,"variant":"red_not_paws_00#0","tokens":100,"calls":7,"padded_tokens":112,"wall_seconds":0.5384495829930529,"hidden_sha256":"0c86526fad032b3569d8c0cd72188f1da07b5d3542f53967633c3924a7cde418","finite":true,"all_zero":false,"index":8,"shard":"red_arms_r4","first_in_process":false},{"id":"tv4_000","k":0,"variant":"state_swap_tv4_000__tv4_001#0","tokens":93,"calls":6,"padded_tokens":96,"wall_seconds":0.3549479170469567,"hidden_sha256":"45dadbbc63acf4fe1c34341d80668c4aee4605003a5a6e6a9c6736183cadf269","finite":true,"all_zero":false,"index":9,"shard":"red_arms_r4","first_in_process":false},{"id":"tv4_001","k":0,"variant":"state_swap_tv4_000__tv4_001#1","tokens":90,"calls":6,"padded_tokens":96,"wall_seconds":0.19416533305775374,"hidden_sha256":"56f5591fdc12f2964f2c5ff7561ec08e549dc55b5aa34ecdee39ac4ca6a03a28","finite":true,"all_zero":false,"index":10,"shard":"red_arms_r4","first_in_process":false},{"id":"semif_a3f18f3a63d45345942b","k":0,"variant":"state_swap_semif_a3f18f3a63d45345942b__semif_0b43ea8e24e74d621f1b#0","tokens":68,"calls":5,"padded_tokens":80,"wall_seconds":0.5125874590594321,"hidden_sha256":"e7556537a879149869754a53fbac696bb9a0956251880d85d1abc36346f7d503","finite":true,"all_zero":false,"index":11,"shard":"red_arms_r4","first_in_process":false},{"id":"semif_0b43ea8e24e74d621f1b","k":0,"variant":"state_swap_semif_a3f18f3a63d45345942b__semif_0b43ea8e24e74d621f1b#1","tokens":76,"calls":5,"padded_tokens":80,"wall_seconds":0.187633250025101,"hidden_sha256":"a6c241962cfa26bf9231d089a81c1c53dbcaa188e31785ee74850812a214d03f","finite":true,"all_zero":false,"index":12,"shard":"red_arms_r4","first_in_process":false}],"reset_bit_equal":true},"seconds":13.665897333063185,"generated_at":"2026-10-08T02:09:13+00:00"}}
