{ "schema": "s1-mini-device-gate/1", "device": "iPhone 17 Pro (iPhone18,1), A19 Pro", "bundle": "s1_mini_decode_int8lin (759 MB, JIT .aimodel, no AOT needed)", "app": "ondevice/PipelinedBench, Release, COREAI_CHUNK_THRESHOLD=1, coreai-pipelined", "date": "2026-08-25", "claim": "device == Mac == fp32 HF, below a hard 1024-token ceiling", "numerics": { "nat": { "case": "323-token dictation transcript -> its 276-token rewrite (599 total)", "result": "276/276 PASS, token-identical to the Mac engine, EOS included" }, "oracle": { "case": "currency + date normalization (the case int4 corrupted)", "result": "27/27 PASS, token-identical to the Mac engine, EOS included" } }, "dynamic_kv_g1024_gate": { "required_by": "the zoo protocol for every dynamic-KV model (g<=256 cannot see the bug)", "result": "PASS — 1024-token generation completes, numerics still exact" }, "ceiling": { "finding": "prompt + generated must stay under 1024 tokens on iOS", "measurement": { "case": "611-token transcript whose Mac rewrite runs 603 tokens (1214 total)", "device_generated": 413, "device_tokens_matching_mac": 413, "wrong_tokens": 0, "stopped_at_absolute_position": 1024, "reading": "truncation, NOT corruption — every token produced was correct" }, "cause": { "verified_in": "coreai-models swift/Sources/CoreAILanguageModels/InferenceEngines/CoreAIPipelinedEngine.swift", "mechanism": "#if os(iOS) + GrowingKVCache -> iosDynamicKVCapacityCap = 1024; maxTokens is capped to 1024 - processed - prompt.count, and a prompt that leaves no budget throws contextLengthExceeded", "why_it_exists": "the iOS on-device compiler miscompiles growing-KV specializations at seq >= 2048 (apple/coreai-models#124); truncating is the guard working, and it is shipped engine behaviour, not a bench artifact", "arithmetic_check": "1024 - 0 - 611 = 413 == the 413 tokens the device produced" }, "product_rule": "S1-mini's card recommends inputs up to ~1000 tokens; on iPhone a ~1000-token prompt leaves almost no room for the rewrite, and a >=1024-token prompt throws. Chunk to roughly <=450-500 input tokens so prompt + rewrite clears the cap.", "evidence": { "mac_prompt_ids": [ 151644, 8948, 198, 2610, 525, 264, 1467, 4622, 3135, 369, 8806, 4686, 9345, 60312, 13, 576, 1946, 12033, 448, 264, 2524, 1555, 37838, 279, 41328, 11, 5944, 11, 323, 2266, 5003, 26, 4240, 279, 35715, 311, 2432, 1846, 5003, 323, 2550, 1172, 279, 27722, 1467, 13, 151645, 198, 151644, 872, 198, 58, 623, 98607, 25, 18267, 8460, 278, 60, 508, 22952, 25, 60701, 60, 508, 1972, 25, 4586, 921, 704, 4443, 16910, 1077, 752, 1101, 633, 678, 315, 419, 1495, 1573, 600, 10667, 432, 582, 1030, 279, 50434, 3395, 419, 6556, 323, 43744, 1052, 1033, 13221, 2326, 2513, 429, 3697, 700, 315, 432, 1156, 279, 15301, 5109, 525, 3520, 2664, 1091, 582, 3381, 1075, 582, 7877, 93835, 12506, 304, 2804, 2326, 537, 81397, 279, 78005, 572, 43744, 1378, 12506, 429, 2684, 32970, 389, 18491, 5923, 1156, 4518, 315, 83649, 270, 2106, 72, 769, 2086, 3166, 374, 279, 80869, 4379, 3937, 504, 3040, 1459, 1378, 3266, 1495, 311, 2326, 1459, 825, 3266, 892, 374, 43744, 279, 15457, 432, 594, 1012, 2474, 582, 3855, 14849, 432, 304, 17073, 17073, 3040, 28047, 1466, 15482, 429, 594, 1576, 315, 279, 389, 37569, 4344, 582, 27259, 304, 70619, 323, 600, 1744, 1340, 594, 1290, 4843, 323, 419, 374, 279, 825, 429, 600, 1744, 582, 1184, 311, 3520, 3061, 911, 374, 43744, 23134, 582, 8039, 291, 369, 4743, 24198, 419, 4279, 323, 582, 3003, 21446, 1378, 279, 15301, 374, 15430, 600, 1744, 582, 3003, 2684, 7196, 3040, 11178, 518, 1590, 6430, 323, 1378, 315, 1105, 582, 2299, 4658, 2087, 311, 9052, 311, 11243, 6081, 773, 4443, 847, 1349, 374, 582, 1265, 2987, 4828, 279, 20892, 476, 582, 1265, 4193, 429, 582, 2299, 2087, 311, 387, 518, 3040, 537, 4743, 323, 421, 582, 2299, 518, 3040, 1221, 279, 68817, 3880, 311, 2297, 1576, 279, 6371, 975, 19110, 4236, 1251, 8028, 14019, 1083, 600, 1184, 311, 1795, 705, 448, 5777, 911, 279, 20728, 5116, 279, 825, 429, 27902, 389, 15217, 26127, 1156, 17073, 17073, 4743, 807, 4829, 311, 1414, 421, 582, 2299, 13873, 287, 518, 279, 1852, 25996, 892, 374, 43744, 29235, 16183, 3040, 7739, 11192, 264, 1042, 476, 421, 582, 2299, 2087, 705, 311, 279, 20179, 25996, 892, 374, 26127, 825, 16183, 4236, 7739, 847, 30555, 374, 4717, 1380, 582, 525, 369, 2441, 1042, 323, 64778, 432, 3055, 582, 1414, 279, 1968, 1830, 6534, 4443, 1128, 770, 14019, 21639, 279, 10455, 504, 1537, 259, 7047, 582, 5558, 911, 35398, 4236, 4420, 315, 13914, 389, 279, 6028, 323, 43744, 279, 1736, 92611, 336, 374, 4152, 79313, 600, 1053, 600, 4172, 3270, 279, 24808, 3772, 773, 600, 1184, 311, 3520, 653, 429, 279, 3704, 5240, 572, 264, 2193, 2297, 429, 3937, 700, 2041, 279, 13475, 3395, 323, 279, 5046, 374, 43744, 582, 2299, 7842, 264, 2567, 2984, 423, 389, 429, 15867, 892, 1265, 614, 1012, 1052, 2669, 26044, 1077, 752, 1083, 6099, 311, 3624, 279, 9530, 311, 3594, 56340, 1573, 279, 4479, 6438, 389, 279, 10510, 44607, 566, 4829, 432, 43744, 35398, 8063, 4115, 8305, 773, 429, 3363, 279, 42541, 835, 315, 1899, 323, 279, 5109, 304, 432, 1184, 311, 2432, 1128, 17017, 4652, 389, 279, 30387, 537, 279, 9960, 504, 279, 2086, 600, 1744, 429, 594, 4297, 4443, 14019, 902, 3783, 825, 803, 3166, 279, 6002, 304, 44581, 4588, 911, 821, 52866, 1549, 807, 1366, 19539, 304, 4378, 429, 4302, 10901, 279, 15555, 323, 1290, 1431, 600, 646, 944, 2968, 1105, 429, 1576, 315, 279, 27875, 20728, 773, 2987, 582, 3271, 429, 20728, 311, 458, 15555, 5537, 476, 582, 79637, 1105, 700, 369, 429, 2692, 600, 3278, 2548, 279, 5339, 2083, 1128, 279, 11893, 374, 389, 279, 15555, 5537, 2999, 151645, 198, 151644, 77091, 198, 151667, 271, 151668, 271 ], "mac_gen_ids": [ 32313, 11, 1077, 752, 1101, 633, 678, 315, 419, 1495, 1573, 358, 10667, 432, 13, 1205, 1030, 279, 50434, 3395, 419, 6556, 11, 323, 1052, 1033, 13221, 2326, 2513, 429, 3697, 700, 315, 432, 382, 5338, 11, 279, 15301, 5109, 525, 3520, 2664, 1091, 582, 3381, 13, 1205, 7877, 220, 16, 24, 12506, 304, 1207, 18, 11, 537, 220, 16, 22, 13, 576, 78005, 572, 1378, 12506, 429, 2684, 32970, 389, 6527, 220, 16, 4518, 315, 6122, 220, 18, 15, 382, 15666, 11, 279, 80869, 4379, 3937, 504, 220, 19, 13, 17, 4, 1495, 311, 220, 18, 13, 16, 13384, 892, 374, 279, 15457, 432, 594, 1012, 2474, 582, 3855, 14849, 432, 304, 220, 17, 15, 17, 19, 13, 20445, 15482, 429, 594, 1576, 315, 279, 389, 37569, 4344, 582, 27259, 304, 6156, 11, 323, 358, 1744, 1340, 594, 1290, 382, 36975, 11, 419, 374, 279, 825, 429, 358, 1744, 582, 1184, 311, 3520, 3061, 911, 25, 23134, 13, 1205, 8039, 291, 369, 4743, 24198, 419, 4279, 11, 323, 582, 3003, 21446, 1378, 13, 576, 15301, 374, 15430, 13, 358, 1744, 582, 3003, 2684, 7196, 3040, 11178, 518, 1590, 6430, 11, 323, 1378, 315, 1105, 582, 2299, 4658, 2087, 311, 9052, 311, 11243, 6081, 382, 4416, 847, 1349, 374, 582, 1265, 2987, 4828, 279, 20892, 476, 582, 1265, 4193, 429, 582, 2299, 2087, 311, 387, 518, 220, 19, 11, 537, 220, 21, 13, 1597, 421, 582, 2299, 518, 220, 19, 11, 1221, 279, 68817, 3880, 311, 2297, 11, 1576, 279, 6371, 975, 19110, 220, 20, 1251, 8028, 382, 11908, 11, 1083, 11, 358, 1184, 311, 1795, 705, 448, 5777, 911, 279, 20728, 5116, 11, 279, 825, 429, 27902, 389, 5470, 220, 18, 16, 11, 220, 17, 15, 17, 21, 13, 2379, 4829, 311, 1414, 421, 582, 2299, 13873, 287, 518, 279, 1852, 25996, 11, 892, 374, 400, 16, 17, 11, 19, 15, 15, 264, 1042, 11, 476, 421, 582, 2299, 2087, 705, 311, 279, 20179, 25996, 11, 892, 374, 400, 18, 16, 11, 20, 15, 15, 13, 3017, 30555, 374, 4717, 1380, 582, 525, 369, 2441, 1042, 323, 64778, 432, 3055, 582, 1414, 279, 1968, 1830, 6534, 382, 3838, 770, 30, 8670, 21639, 11, 279, 10455, 504, 1537, 7589, 11, 582, 5558, 911, 220, 19, 20, 4420, 315, 13914, 389, 279, 6028, 11, 323, 279, 1736, 92611, 336, 374, 4152, 6602, 13, 358, 1053, 358, 4172, 3270, 279, 24808, 3772, 11, 773, 358, 1184, 311, 3520, 653, 429, 13, 576, 3704, 5240, 572, 264, 2193, 2297, 429, 3937, 700, 2041, 279, 13475, 3395, 11, 323, 279, 5046, 374, 582, 2299, 7842, 264, 2567, 2984, 423, 389, 429, 15867, 11, 892, 1265, 614, 1012, 1052, 2669, 13, 70947, 11, 1077, 752, 1083, 6099, 311, 3624, 279, 9530, 311, 35683, 1573, 279, 4479, 6438, 389, 279, 220, 16, 16, 339, 13, 1260, 4829, 432, 220, 19, 23, 4115, 8305, 11, 773, 429, 3363, 279, 220, 24, 339, 835, 315, 1899, 11, 323, 279, 5109, 304, 432, 1184, 311, 2432, 1128, 17017, 4652, 389, 279, 220, 22, 339, 11, 537, 279, 9960, 504, 279, 220, 17, 303, 382, 40, 1744, 429, 594, 4297, 13, 8670, 902, 11, 3783, 11, 825, 803, 3166, 13, 576, 6002, 304, 19846, 4588, 911, 821, 52866, 1549, 13, 2379, 1366, 19539, 304, 4378, 429, 4302, 10901, 279, 9812, 11, 323, 1290, 1431, 358, 646, 944, 2968, 1105, 429, 1576, 315, 279, 27875, 20728, 13, 2055, 2987, 582, 3271, 429, 20728, 311, 458, 9812, 5537, 11, 476, 582, 79637, 1105, 700, 369, 429, 2692, 13, 358, 3278, 2548, 279, 5339, 2083, 1128, 279, 11893, 374, 389, 279, 9812, 5537, 2999, 13, 151645 ], "device_gen_ids": [ 32313, 11, 1077, 752, 1101, 633, 678, 315, 419, 1495, 1573, 358, 10667, 432, 13, 1205, 1030, 279, 50434, 3395, 419, 6556, 11, 323, 1052, 1033, 13221, 2326, 2513, 429, 3697, 700, 315, 432, 382, 5338, 11, 279, 15301, 5109, 525, 3520, 2664, 1091, 582, 3381, 13, 1205, 7877, 220, 16, 24, 12506, 304, 1207, 18, 11, 537, 220, 16, 22, 13, 576, 78005, 572, 1378, 12506, 429, 2684, 32970, 389, 6527, 220, 16, 4518, 315, 6122, 220, 18, 15, 382, 15666, 11, 279, 80869, 4379, 3937, 504, 220, 19, 13, 17, 4, 1495, 311, 220, 18, 13, 16, 13384, 892, 374, 279, 15457, 432, 594, 1012, 2474, 582, 3855, 14849, 432, 304, 220, 17, 15, 17, 19, 13, 20445, 15482, 429, 594, 1576, 315, 279, 389, 37569, 4344, 582, 27259, 304, 6156, 11, 323, 358, 1744, 1340, 594, 1290, 382, 36975, 11, 419, 374, 279, 825, 429, 358, 1744, 582, 1184, 311, 3520, 3061, 911, 25, 23134, 13, 1205, 8039, 291, 369, 4743, 24198, 419, 4279, 11, 323, 582, 3003, 21446, 1378, 13, 576, 15301, 374, 15430, 13, 358, 1744, 582, 3003, 2684, 7196, 3040, 11178, 518, 1590, 6430, 11, 323, 1378, 315, 1105, 582, 2299, 4658, 2087, 311, 9052, 311, 11243, 6081, 382, 4416, 847, 1349, 374, 582, 1265, 2987, 4828, 279, 20892, 476, 582, 1265, 4193, 429, 582, 2299, 2087, 311, 387, 518, 220, 19, 11, 537, 220, 21, 13, 1597, 421, 582, 2299, 518, 220, 19, 11, 1221, 279, 68817, 3880, 311, 2297, 11, 1576, 279, 6371, 975, 19110, 220, 20, 1251, 8028, 382, 11908, 11, 1083, 11, 358, 1184, 311, 1795, 705, 448, 5777, 911, 279, 20728, 5116, 11, 279, 825, 429, 27902, 389, 5470, 220, 18, 16, 11, 220, 17, 15, 17, 21, 13, 2379, 4829, 311, 1414, 421, 582, 2299, 13873, 287, 518, 279, 1852, 25996, 11, 892, 374, 400, 16, 17, 11, 19, 15, 15, 264, 1042, 11, 476, 421, 582, 2299, 2087, 705, 311, 279, 20179, 25996, 11, 892, 374, 400, 18, 16, 11, 20, 15, 15, 13, 3017, 30555, 374, 4717, 1380, 582, 525, 369, 2441, 1042, 323, 64778, 432, 3055, 582, 1414, 279, 1968, 1830, 6534, 382, 3838, 770, 30, 8670, 21639, 11, 279, 10455, 504, 1537, 7589, 11, 582, 5558, 911, 220, 19, 20, 4420, 315, 13914, 389, 279, 6028, 11, 323, 279, 1736, 92611, 336, 374, 4152, 6602, 13, 358, 1053, 358, 4172, 3270, 279, 24808, 3772, 11, 773, 358, 1184, 311, 3520, 653, 429, 13, 576, 3704, 5240, 572 ], "device_is_exact_prefix_of_mac": true, "recheck": "device_gen_ids must equal mac_gen_ids[:len(device_gen_ids)] and len(mac_prompt_ids) + len(device_gen_ids) must equal 1024" } }, "speed": { "settled": { "prefill_tok_s": 69.0, "decode_tok_s": 62.4, "protocol": "p=128 g=256, n=3, cold device, PipelinedBench Release", "run": "s1mini_cool.log" }, "reproduced_by": { "prefill_tok_s": 71.3, "decode_tok_s": 64.7, "protocol": "p=128 g=256, n=2, first run of the session", "run": "s1mini_r1.log" }, "thermal_envelope": { "sustained_tok_s": { "prefill": 34.9, "decode": 30.5 }, "note": "after back-to-back 1024-token generations the same bundle measured 34.9/30.5 (s1mini_speed.log) — roughly half. Seven minutes idle restored 69.0/62.4, so it is thermal, not a regression. Published number is the cold one because that is the protocol every other card uses; the sustained number is recorded because a dictation post-processor runs repeatedly and will meet it." } }, "raw_runs": { "note": "the probe logs themselves are *.log and the repo gitignores those, so the lines that carry the verdict are transcribed here instead of linked", "runs": { "s1mini_cool": [ "engine ready in 1.0 s", "NUMERICS nat=276/276 PASS", "NUMERICS oracle=27/27 PASS", "TRIAL 1 (p=128 g=256) prefill=70.9 decode=64.5 tok/s", "TRIAL 2 (p=128 g=256) prefill=68.6 decode=62.2 tok/s", "TRIAL 3 (p=128 g=256) prefill=67.5 decode=60.5 tok/s", "STATS prefill=69.0 decode=62.4 nat=276/276 oracle=27/27" ], "s1mini_final": [ "engine ready in 0.2 s", "NUMERICS nat=276/276 PASS", "NUMERICS oracle=27/27 PASS", "TRIAL 1 (p=128 g=1024) prefill=71.7 decode=61.2 tok/s", "TRIAL 2 (p=128 g=1024) prefill=65.6 decode=49.7 tok/s", "STATS prefill=68.6 decode=55.4 nat=276/276 oracle=27/27" ], "s1mini_g1024": [ "engine ready in 0.2 s", "NUMERICS nat=11/11 PASS", "NUMERICS oracle=27/27 PASS", "TRIAL 1 (p=128 g=1024) prefill=72.9 decode=60.3 tok/s", "STATS prefill=72.9 decode=60.3 nat=11/11 oracle=27/27" ], "s1mini_long": [ "engine ready in 0.8 s", "NUMERICS nat=413/603 MISMATCH", "NUMERICS oracle=27/27 PASS", "TRIAL 1 (p=128 g=1024) prefill=69.9 decode=59.9 tok/s", "STATS prefill=69.9 decode=59.9 nat=413/603 oracle=27/27" ], "s1mini_r1": [ "engine ready in 4.7 s", "NUMERICS nat=11/11 PASS", "NUMERICS oracle=27/27 PASS", "TRIAL 1 (p=128 g=256) prefill=71.2 decode=64.8 tok/s", "TRIAL 2 (p=128 g=256) prefill=71.3 decode=64.6 tok/s", "STATS prefill=71.3 decode=64.7 nat=11/11 oracle=27/27" ], "s1mini_speed": [ "engine ready in 0.2 s", "NUMERICS nat=276/276 PASS", "NUMERICS oracle=27/27 PASS", "TRIAL 1 (p=128 g=256) prefill=39.8 decode=32.3 tok/s", "TRIAL 2 (p=128 g=256) prefill=35.0 decode=29.5 tok/s", "TRIAL 3 (p=128 g=256) prefill=29.9 decode=29.8 tok/s", "STATS prefill=34.9 decode=30.5 nat=276/276 oracle=27/27" ] } } }