{
  "reviewed_at_utc": "2026-10-18T22:44:34.776995+00:00",
  "upstream_url": "https://github.com/gittensor-ai-lab/sparkinfer-k3",
  "revision": "7689cc77e0e18d4b4f1937cf2a189dcd5a8d90f3",
  "python": "3.12.3 (main, Aug 31 2026, 10:18:26) [GCC 13.3.0]",
  "platform": "Linux-6.8.0-147-generic-x86_64-with-glibc2.39",
  "compiler": "g++ (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0",
  "scope": "CPU source/contracts only; no GPU execution, model weights or throughput validation",
  "checks": [
    {
      "name": "tp_shard",
      "stage": "compile",
      "command": [
        "g++",
        "-std=c++17",
        "-O2",
        "-I/tmp/course-sparkinfer-k3/runtime/include",
        "-I/tmp/course-sparkinfer-k3/include",
        "/tmp/course-sparkinfer-k3/runtime/tests/tp_shard_cpu_test.cpp",
        "/tmp/course-sparkinfer-k3/runtime/src/tp/shard.cpp",
        "-o",
        "/tmp/course-sparkinfer-cpu-tlx7to72/tp_shard"
      ],
      "exit_code": 0,
      "stdout": "",
      "stderr": ""
    },
    {
      "name": "tp_shard",
      "stage": "execute",
      "command": [
        "/tmp/course-sparkinfer-cpu-tlx7to72/tp_shard"
      ],
      "exit_code": 0,
      "stdout": "tp1_is_the_identity_shard\nqwen36_tp8\nkimi_k3_tp8\nkimi_k3_tp4_for_b300\nshape_must_divide_on_every_axis\nexpert_bands_tile_exactly\nvocab_bands_tile_exactly\nindivisible_shapes_are_rejected_with_a_reason\nexpert_fallback_to_ffn_width\nbad_rank_and_device_list\nrow_shard_slices_are_contiguous_and_tile\ncol_shard_is_strided_and_tiles_per_row\nrow_and_col_shard_compose\nreduce_accounting\neven_band_edges\nkda_tp8_real_dims\nkda_bands_tile_and_hold_the_invariant\nkda_tp1_is_the_identity_shard\nkda_rejects_what_it_cannot_split\nkda_replicate_and_offset_bands\nkda_q8_0_traffic_divides_by_eight\nkda_collective_accounting\n\n5117 checks, 0 failure(s)\n",
      "stderr": ""
    },
    {
      "name": "tp_backend",
      "stage": "compile",
      "command": [
        "g++",
        "-std=c++17",
        "-O2",
        "-I/tmp/course-sparkinfer-k3/runtime/include",
        "-I/tmp/course-sparkinfer-k3/include",
        "/tmp/course-sparkinfer-k3/runtime/tests/tp_backend_cpu_test.cpp",
        "/tmp/course-sparkinfer-k3/runtime/src/tp/backend_select.cpp",
        "-o",
        "/tmp/course-sparkinfer-cpu-tlx7to72/tp_backend"
      ],
      "exit_code": 0,
      "stdout": "",
      "stderr": ""
    },
    {
      "name": "tp_backend",
      "stage": "execute",
      "command": [
        "/tmp/course-sparkinfer-cpu-tlx7to72/tp_backend"
      ],
      "exit_code": 0,
      "stdout": "tp1_never_issues_a_collective\nnccl_is_the_default_on_a_real_node\nvendored_backends_are_selectable_on_capable_hardware\nhardware_gates_are_reported_before_vendoring\ntoo_few_devices_is_a_hard_downgrade\nbackend_none_at_tp8_is_rejected_loudly\nmissing_nccl_does_not_silently_unshard\nenv_parsing\nbackend_names_are_stable\n\n45 checks, 0 failure(s)\n",
      "stderr": ""
    },
    {
      "name": "tp_weight_plan",
      "stage": "compile",
      "command": [
        "g++",
        "-std=c++17",
        "-O2",
        "-I/tmp/course-sparkinfer-k3/runtime/include",
        "-I/tmp/course-sparkinfer-k3/include",
        "/tmp/course-sparkinfer-k3/runtime/tests/tp_weight_plan_cpu_test.cpp",
        "/tmp/course-sparkinfer-k3/runtime/src/tp/weight_plan.cpp",
        "/tmp/course-sparkinfer-k3/runtime/src/tp/shard.cpp",
        "-o",
        "/tmp/course-sparkinfer-cpu-tlx7to72/tp_weight_plan"
      ],
      "exit_code": 0,
      "stdout": "",
      "stderr": ""
    },
    {
      "name": "tp_weight_plan",
      "stage": "execute",
      "command": [
        "/tmp/course-sparkinfer-cpu-tlx7to72/tp_weight_plan"
      ],
      "exit_code": 0,
      "stdout": "the_two_collective_tensors\nattention_rules\nmoe_rules\nmodel_level_rules\nunknown_names_are_refused_not_guessed\nlayer_index_is_ignored\ntp1_collapses_everything_to_replicate\nrow_shard_plan_bands_tile\ncol_shard_plan_bands_tile\nkv_replication_is_applied_and_explained\nkv_shards_when_there_are_enough_heads\nfused_qkv_with_replicated_kv_is_refused\nexpert_shard_plan\nexpert_ffn_width_fallback\nindivisible_shapes_are_refused_at_plan_time\nevery_known_suffix_has_a_real_rule\nrule_names_are_stable\nkda_every_loader_tensor_has_a_rule\nkda_plan_is_internally_consistent\nkda_plan_extents_match_kda_shard_dims\nkda_attn_k_v_are_not_a_kv_group\nkda_table_has_no_shadowed_duplicates\n\n1740 checks, 0 failure(s)\n",
      "stderr": ""
    },
    {
      "name": "tp_weight_residency",
      "stage": "compile",
      "command": [
        "g++",
        "-std=c++17",
        "-O2",
        "-I/tmp/course-sparkinfer-k3/runtime/include",
        "-I/tmp/course-sparkinfer-k3/include",
        "/tmp/course-sparkinfer-k3/runtime/tests/tp_weight_residency_cpu_test.cpp",
        "/tmp/course-sparkinfer-k3/runtime/src/tp/weight_residency.cpp",
        "/tmp/course-sparkinfer-k3/runtime/src/tp/weight_plan.cpp",
        "/tmp/course-sparkinfer-k3/runtime/src/tp/shard.cpp",
        "/tmp/course-sparkinfer-k3/runtime/src/gguf.cpp",
        "-o",
        "/tmp/course-sparkinfer-cpu-tlx7to72/tp_weight_residency"
      ],
      "exit_code": 0,
      "stdout": "",
      "stderr": ""
    },
    {
      "name": "tp_weight_residency",
      "stage": "execute",
      "command": [
        "/tmp/course-sparkinfer-cpu-tlx7to72/tp_weight_residency"
      ],
      "exit_code": 0,
      "stdout": "=== tp weight residency (GGUF bytes -> per-rank buffers) ===\n\n[block_aligned_band]\n\n[K3 leading-dense ffn_down: the mid-block cut]\n\n[K3 routed experts: 3584 x 3072 x 896 IQ2_XS]\n  note: one MoE layer's experts = 7.97 GiB full, 1.00 GiB per rank\n\n[RowShard: whole memory rows, no alignment constraint]\n\n[Replicate and TP=1]\n\n[refusals \u2014 these must NOT silently succeed]\n\n[2-D MoE: expert groups x ffn band]\n\n[model-level residency: replication is the TP memory cost]\n  model 7.991 GiB | rank 1.020 GiB | replicated 0.023979 GiB | all ranks 8.159 GiB\n\nPASS: 293 checks, 0 failures\n",
      "stderr": ""
    },
    {
      "name": "kimi_k3_config",
      "stage": "compile",
      "command": [
        "g++",
        "-std=c++17",
        "-O2",
        "-I/tmp/course-sparkinfer-k3/runtime/include",
        "-I/tmp/course-sparkinfer-k3/include",
        "/tmp/course-sparkinfer-k3/runtime/tests/kimi_k3_config_cpu_test.cpp",
        "/tmp/course-sparkinfer-k3/runtime/src/gguf.cpp",
        "-o",
        "/tmp/course-sparkinfer-cpu-tlx7to72/kimi_k3_config"
      ],
      "exit_code": 0,
      "stdout": "",
      "stderr": ""
    },
    {
      "name": "kimi_k3_config",
      "stage": "execute",
      "command": [
        "/tmp/course-sparkinfer-cpu-tlx7to72/kimi_k3_config"
      ],
      "exit_code": 0,
      "stdout": "kimi_k3_config_cpu_test: OK  (69 KDA + 24 MLA, situ/attn_res/latent required)\n",
      "stderr": "[kimi-k3] missing REQUIRED key kimi-k3.activation.situ_beta\n[kimi-k3] missing REQUIRED array kimi-k3.attention.head_count_kv\n"
    },
    {
      "name": "kimi_k3_manifest",
      "stage": "compile",
      "command": [
        "g++",
        "-std=c++17",
        "-O2",
        "-I/tmp/course-sparkinfer-k3/runtime/include",
        "-I/tmp/course-sparkinfer-k3/include",
        "/tmp/course-sparkinfer-k3/runtime/tests/kimi_k3_manifest_cpu_test.cpp",
        "/tmp/course-sparkinfer-k3/runtime/src/gguf.cpp",
        "-o",
        "/tmp/course-sparkinfer-cpu-tlx7to72/kimi_k3_manifest"
      ],
      "exit_code": 0,
      "stdout": "",
      "stderr": ""
    },
    {
      "name": "kimi_k3_manifest",
      "stage": "execute",
      "command": [
        "/tmp/course-sparkinfer-cpu-tlx7to72/kimi_k3_manifest"
      ],
      "exit_code": 0,
      "stdout": "kimi_k3_manifest_cpu_test: OK  (93 tensors across 2 shards, 93 checks)\n",
      "stderr": "[gguf] split.no=1 but expected shard index 0\n"
    },
    {
      "name": "kimi_k3_decode_plan",
      "stage": "compile",
      "command": [
        "g++",
        "-std=c++17",
        "-O2",
        "-I/tmp/course-sparkinfer-k3/runtime/include",
        "-I/tmp/course-sparkinfer-k3/include",
        "/tmp/course-sparkinfer-k3/runtime/tests/kimi_k3_decode_plan_cpu_test.cpp",
        "/tmp/course-sparkinfer-k3/runtime/src/models/kimi_k3_decode_plan.cpp",
        "/tmp/course-sparkinfer-k3/runtime/src/tp/weight_plan.cpp",
        "/tmp/course-sparkinfer-k3/runtime/src/tp/shard.cpp",
        "/tmp/course-sparkinfer-k3/runtime/src/gguf.cpp",
        "-o",
        "/tmp/course-sparkinfer-cpu-tlx7to72/kimi_k3_decode_plan"
      ],
      "exit_code": 0,
      "stdout": "",
      "stderr": ""
    },
    {
      "name": "kimi_k3_decode_plan",
      "stage": "execute",
      "command": [
        "/tmp/course-sparkinfer-cpu-tlx7to72/kimi_k3_decode_plan"
      ],
      "exit_code": 0,
      "stdout": "[Earlier output omitted; complete local report was /tmp/course-sparkinfer-reviewed.json]\ne)\n  MLA rms_norm [q_a_norm] 1536 -> 1536  blk.0.attn_q_a_norm.weight (replicate)\n  MLA matmul [q_b] 1536 -> 18432  blk.0.attn_q_b.weight (row)\n  MLA matmul [kv_a_mqa] 7168 -> 576  blk.0.attn_kv_a_mqa.weight (replicate)\n  MLA rms_norm [kv_a_norm] 512 -> 512  blk.0.attn_kv_a_norm.weight (replicate)\n  MLA mla_absorb_q [absorb_q] 18432 -> 55296  blk.0.attn_k_b.weight (expert)\n       # fold k_b into q so attention runs in latent space; output is key_length (kv_lora+rope_dim) wide per head, not kv_lora_rank alone\n  MLA mla_decode_attn [decode_attn] 55296 -> 12288  blk.0.attn_v_b.weight (expert)\n       # NoPE-only\n  MLA mla_gate_out [gate_out] 12288 -> 12288\n       # sigmoid gate\n  MLA matmul [attn_out] 12288 -> 7168  blk.0.attn_output.weight (col)  +reduce\n  MLA ALL_REDUCE [reduce_attn] 7168 -> 7168\n  MLA add_residual [replace_attn] 7168 -> 7168\n       # REPLACE, not add: see res_bank_push above\n  MLA attn_res_mix [ffn_res_mix] 7168 -> 7168  blk.0.ffn_res_score.weight (replicate)\n       # runs BEFORE ffn_norm, same reasoning as the attention-side mix\n  MLA rms_norm [ffn_norm] 7168 -> 7168  blk.0.ffn_norm.weight (replicate)\n  MLA matmul [ffn_gate] 7168 -> 33792  blk.0.ffn_gate.weight (row)\n  MLA matmul [ffn_up] 7168 -> 33792  blk.0.ffn_up.weight (row)\n  MLA situ [situ] 33792 -> 33792\n  MLA matmul [ffn_down] 33792 -> 7168  blk.0.ffn_down.weight (col)  +reduce\n  MLA ALL_REDUCE [reduce_ffn] 7168 -> 7168\n  MLA add_residual [add_ffn] 7168 -> 7168\n\n[layer 5 (KDA, MoE) as scheduled]\n  KDA attn_res_mix [attn_res_mix] 7168 -> 7168  blk.5.attn_res_score.weight (replicate)\n       # runs BEFORE attn_norm \u2014 chooses what the attention block sees, not what gets folded in after it. Scores from RMS-normalised checkpoints; the weighted sum uses the RAW ones. score_w is [n_embd], a dot-product scorer, not a projection.\n  KDA rms_norm [attn_norm] 7168 -> 7168  blk.5.attn_norm.weight (replicate)\n  KDA matmul [q_proj] 7168 -> 12288  blk.5.attn_q.weight (row)\n  KDA matmul [k_proj] 7168 -> 12288  blk.5.attn_k.weight (row)\n  KDA matmul [v_proj] 7168 -> 12288  blk.5.attn_v.weight (row)\n  KDA kda_conv_step [conv_q] 12288 -> 12288  blk.5.ssm_conv1d_q.weight (replicate)\n       # depthwise causal, silu after\n  KDA kda_conv_step [conv_k] 12288 -> 12288  blk.5.ssm_conv1d_k.weight (replicate)\n  KDA kda_conv_step [conv_v] 12288 -> 12288  blk.5.ssm_conv1d_v.weight (replicate)\n  KDA l2_norm_heads [l2_qk] 12288 -> 12288\n       # per-head L2 on q and k\n  KDA matmul [f_a] 7168 -> 128  blk.5.ssm_f_a.weight (replicate)\n  KDA matmul [f_b] 128 -> 12288  blk.5.ssm_f_b.weight (row)\n  KDA matmul [dt_bias] 12288 -> 12288  blk.5.ssm_dt.bias (replicate)\n  KDA kda_decay_gate [decay_gate] 12288 -> 12288  blk.5.ssm_a (replicate)\n       # A (ssm_a) is PER-HEAD [n_head]; g_raw (f_b(f_a(x))+dt_bias) is PER-CHANNEL [d_inner]. The gate's output is per-channel because g_raw is, even though A broadcasts one scalar across every channel in its head.\n  KDA matmul [beta] 7168 -> 96  blk.5.ssm_beta.weight (replicate)\n  KDA kda_decode_step [delta_rule] 12288 -> 12288\n       # gated delta rule, recurrent state update\n  KDA matmul [g_proj] 7168 -> 12288  blk.5.ssm_g.weight (row)\n  KDA kda_gate_out [gate_out] 12288 -> 12288  blk.5.ssm_norm.weight (replicate)\n  KDA matmul [attn_out] 12288 -> 7168  blk.5.attn_output.weight (col)  +reduce\n  KDA ALL_REDUCE [reduce_attn] 7168 -> 7168\n  KDA add_residual [add_attn] 7168 -> 7168\n  KDA attn_res_mix [ffn_res_mix] 7168 -> 7168  blk.5.ffn_res_score.weight (replicate)\n       # runs BEFORE ffn_norm, same reasoning as the attention-side mix\n  KDA rms_norm [ffn_norm] 7168 -> 7168  blk.5.ffn_norm.weight (replicate)\n  KDA matmul [router] 7168 -> 896  blk.5.ffn_gate_inp.weight (replicate)\n  KDA moe_router_noaux_tc [route_noaux_tc] 896 -> 16  blk.5.exp_probs_b.bias (replicate)\n       # expert bias biases SELECTION only, never the returned weights\n  KDA matmul [routed_down] 7168 -> 3584  blk.5.ffn_routed_down.weight (replicate)\n  KDA moe_expert_ffn_iq2xs [expert_ffn] 3584 -> 3584  blk.5.ffn_gate_exps.weight (expert)\n       # gate/up -> situ at moe_ffn -> down, all in latent space\n  KDA moe_expert_ffn_iq2xs [expert_ffn_up] 3584 -> 3072  blk.5.ffn_up_exps.weight (expert)\n  KDA moe_expert_ffn_iq2xs [expert_ffn_down] 3072 -> 3584  blk.5.ffn_down_exps.weight (expert)\n  KDA ALL_REDUCE [reduce_moe_latent] 3584 -> 3584\n  KDA matmul [routed_up] 3584 -> 7168  blk.5.ffn_routed_up.weight (replicate)\n  KDA add_residual [add_ffn] 7168 -> 7168\n\nPASS: 3185 checks, 0 failures\n",
      "stderr": ""
    },
    {
      "name": "test_kimi_k3_baseline.py",
      "stage": "python-suite",
      "command": [
        "/usr/bin/python3",
        "/tmp/course-sparkinfer-k3/bench/scripts/test_kimi_k3_baseline.py"
      ],
      "exit_code": 0,
      "stdout": ">> #59: needs-rebase cleared \u2014 this round re-measured it at 99e4bad5 against the current frontier\n>> #59: cleared stale tier (eval:none) \u2014 it was measured against a frontier that has since moved\n>> #59: cleared stale tier (eval:xs) \u2014 it was measured against a frontier that has since moved\n--- dry-run: would clear stale eval:l from #59\n>> #90: merge-conflict (conflicts with main \u2014 not evaluated)\n>> #86: merge-conflict cleared \u2014 no longer conflicting\n>> #7: needs-rebase (frontier moved)\n   #64: mergeability settled to CONFLICTING after ~0s\n",
      "stderr": "[Earlier output omitted; complete local report was /tmp/course-sparkinfer-reviewed.json]\n_an_undersized_node (__main__.FitCheckTest.test_smaller_quant_rescues_an_undersized_node) ... ok\ntest_baselines_start_unmeasured_for_every_node (__main__.ForkPinTest.test_baselines_start_unmeasured_for_every_node) ... ok\ntest_k3_uses_a_separate_checkout (__main__.ForkPinTest.test_k3_uses_a_separate_checkout) ... ok\ntest_lock_prefix_matches_node_and_quant (__main__.ForkPinTest.test_lock_prefix_matches_node_and_quant)\nThe prefix carries BOTH, because a slot names the measurement completely. ... ok\ntest_qwen_evals_keep_the_upstream_pin (__main__.ForkPinTest.test_qwen_evals_keep_the_upstream_pin) ... ok\ntest_reference_lock_pins_the_base_branch_too (__main__.ForkPinTest.test_reference_lock_pins_the_base_branch_too) ... ok\ntest_sourcing_kimi_k3_switches_the_engine (__main__.ForkPinTest.test_sourcing_kimi_k3_switches_the_engine) ... ok\ntest_deterministic_for_a_seed (__main__.GgufTokenizerTest.test_deterministic_for_a_seed) ... ok\ntest_long_prompt_hits_exact_length (__main__.GgufTokenizerTest.test_long_prompt_hits_exact_length) ... ok\ntest_missing_binary_is_a_clear_error (__main__.GgufTokenizerTest.test_missing_binary_is_a_clear_error) ... ok\ntest_missing_gguf_is_a_clear_error (__main__.GgufTokenizerTest.test_missing_gguf_is_a_clear_error) ... ok\ntest_short_prompt_ids (__main__.GgufTokenizerTest.test_short_prompt_ids) ... ok\ntest_overrides_take_effect (__main__.LlamaFlagTest.test_overrides_take_effect) ... ok\ntest_required_flags_present (__main__.LlamaFlagTest.test_required_flags_present) ... ok\ntest_server_flags_cover_the_hybrid_arch (__main__.LlamaFlagTest.test_server_flags_cover_the_hybrid_arch) ... ok\ntest_b300_arch_is_overridable (__main__.NodeProfileTest.test_b300_arch_is_overridable) ... ok\ntest_default_node_is_m1 (__main__.NodeProfileTest.test_default_node_is_m1) ... ok\ntest_each_profile (__main__.NodeProfileTest.test_each_profile) ... ok\ntest_every_target_config_declares_its_node (__main__.NodeProfileTest.test_every_target_config_declares_its_node) ... ok\ntest_m4_is_a_capability_not_a_node (__main__.NodeProfileTest.test_m4_is_a_capability_not_a_node) ... ok\ntest_superseded_iq1s_target_is_gone (__main__.NodeProfileTest.test_superseded_iq1s_target_is_gone) ... ok\ntest_unknown_node_warns_but_does_not_fail (__main__.NodeProfileTest.test_unknown_node_warns_but_does_not_fail) ... ok\ntest_lock_vars_parse (__main__.PinAuditTest.test_lock_vars_parse) ... ok\ntest_offline_mode_makes_no_requests (__main__.PinAuditTest.test_offline_mode_makes_no_requests) ... ok\ntest_shard_pins_parse_from_the_shell_source (__main__.PinAuditTest.test_shard_pins_parse_from_the_shell_source) ... ok\ntest_backed_baseline_is_accepted (__main__.ReferenceLockProvenanceTest.test_backed_baseline_is_accepted) ... ok\ntest_hand_filled_baseline_is_rejected (__main__.ReferenceLockProvenanceTest.test_hand_filled_baseline_is_rejected) ... ok\ntest_mismatched_value_is_rejected (__main__.ReferenceLockProvenanceTest.test_mismatched_value_is_rejected) ... ok\ntest_prefill_and_decode_are_not_interchangeable (__main__.ReferenceLockProvenanceTest.test_prefill_and_decode_are_not_interchangeable) ... ok\ntest_shipped_lock_is_clean (__main__.ReferenceLockProvenanceTest.test_shipped_lock_is_clean)\nEvery pinned slot in the shipped lock traces to a committed measurement. ... ok\ntest_wrong_node_does_not_satisfy (__main__.ReferenceLockProvenanceTest.test_wrong_node_does_not_satisfy) ... ok\ntest_default_quant_is_iq1s (__main__.ShardPathTest.test_default_quant_is_iq1s) ... ok\ntest_every_published_quant_resolves (__main__.ShardPathTest.test_every_published_quant_resolves) ... ok\ntest_gguf_path_joins_models_dir (__main__.ShardPathTest.test_gguf_path_joins_models_dir) ... ok\ntest_manifest_name_matches_quant (__main__.ShardPathTest.test_manifest_name_matches_quant) ... ok\ntest_quant_override_changes_every_derived_path (__main__.ShardPathTest.test_quant_override_changes_every_derived_path) ... ok\ntest_target_quant_is_q2kxl (__main__.ShardPathTest.test_target_quant_is_q2kxl) ... ok\ntest_absent_manifest_is_warn_only (__main__.SplitDownloadTest.test_absent_manifest_is_warn_only) ... ok\ntest_complete_shard_set_passes (__main__.SplitDownloadTest.test_complete_shard_set_passes) ... ok\ntest_incomplete_shard_set_is_rejected (__main__.SplitDownloadTest.test_incomplete_shard_set_is_rejected) ... ok\ntest_manifest_mismatch_is_fatal (__main__.SplitDownloadTest.test_manifest_mismatch_is_fatal) ... ok\n\n----------------------------------------------------------------------\nRan 200 tests in 33.612s\n\nOK\n"
    },
    {
      "name": "test_label.py",
      "stage": "python-suite",
      "command": [
        "/usr/bin/python3",
        "/tmp/course-sparkinfer-k3/bench/scripts/test_label.py"
      ],
      "exit_code": 0,
      "stdout": "",
      "stderr": "test_baseline_label_when_no_frontier_exists (__main__.LabelPolicyTest.test_baseline_label_when_no_frontier_exists) ... ok\ntest_batched_prefill_frontier_scale_scores_XL_at_forty_six_pct (__main__.LabelPolicyTest.test_batched_prefill_frontier_scale_scores_XL_at_forty_six_pct) ... ok\ntest_correctness_gate_rejects_bad_accuracy (__main__.LabelPolicyTest.test_correctness_gate_rejects_bad_accuracy) ... ok\ntest_difficulty_boost_default_cap_is_one_point_five_x (__main__.LabelPolicyTest.test_difficulty_boost_default_cap_is_one_point_five_x) ... ok\ntest_difficulty_boost_high_cap_still_capped_at_the_raw_speedup (__main__.LabelPolicyTest.test_difficulty_boost_high_cap_still_capped_at_the_raw_speedup) ... ok\ntest_llama_anchored_buckets_without_difficulty_boost (__main__.LabelPolicyTest.test_llama_anchored_buckets_without_difficulty_boost) ... ok\ntest_long_context_metadata_is_preserved_in_verdict (__main__.LabelPolicyTest.test_long_context_metadata_is_preserved_in_verdict) ... ok\ntest_prefill_frontier_scale_scores_L_at_eleven_pct (__main__.LabelPolicyTest.test_prefill_frontier_scale_scores_L_at_eleven_pct) ... ok\ntest_significance_gate_is_strictly_above_two_percent (__main__.LabelPolicyTest.test_significance_gate_is_strictly_above_two_percent) ... ok\ntest_strict_cap_limits_per_context_llama_ref_inflation (__main__.LabelPolicyTest.test_strict_cap_limits_per_context_llama_ref_inflation) ... ok\ntest_unoptimized_model_does_not_mint_xl_from_low_hanging_fruit (__main__.LabelPolicyTest.test_unoptimized_model_does_not_mint_xl_from_low_hanging_fruit) ... ok\ntest_absent_guards_are_inert (__main__.NoRegressionGuardTest.test_absent_guards_are_inert) ... ok\ntest_all_guards_holding_leaves_the_tier_alone (__main__.NoRegressionGuardTest.test_all_guards_holding_leaves_the_tier_alone) ... ok\ntest_correctness_outranks_regression_in_the_reason (__main__.NoRegressionGuardTest.test_correctness_outranks_regression_in_the_reason) ... ok\ntest_latency_guards_respect_direction (__main__.NoRegressionGuardTest.test_latency_guards_respect_direction) ... ok\ntest_provenance_cannot_forge_a_passing_guard (__main__.NoRegressionGuardTest.test_provenance_cannot_forge_a_passing_guard) ... ok\ntest_qwen_provenance_guard_fields_are_still_writable (__main__.NoRegressionGuardTest.test_qwen_provenance_guard_fields_are_still_writable) ... ok\ntest_regression_at_one_context_rejects_despite_a_real_gain (__main__.NoRegressionGuardTest.test_regression_at_one_context_rejects_despite_a_real_gain) ... ok\ntest_scored_context_is_recorded (__main__.NoRegressionGuardTest.test_scored_context_is_recorded) ... ok\ntest_tolerance_absorbs_noise_but_not_decay (__main__.NoRegressionGuardTest.test_tolerance_absorbs_noise_but_not_decay) ... ok\ntest_unparseable_guards_refuse_to_score (__main__.NoRegressionGuardTest.test_unparseable_guards_refuse_to_score) ... ok\ntest_unset_baseline_cannot_regress (__main__.NoRegressionGuardTest.test_unset_baseline_cannot_regress) ... ok\n\n----------------------------------------------------------------------\nRan 22 tests in 1.897s\n\nOK\n"
    }
  ],
  "upstream_source_sha256": {
    "bench/scripts/README.md": "0c89c3795ca3c6780ecff063757260c3e618077e4e8de35ea40f464b85981560",
    "bench/scripts/_common.sh": "af9a69a623ddd014fdb2590d9e7b591f94be7ba2ec430fa9f66605b6470b3a06",
    "bench/scripts/_eval_speed.sh": "dc675cbea192310ae792c3e66a96e597a2d6bff68c775668113aa15317e47a98",
    "bench/scripts/_kimi_k3.sh": "a590149dfc3de33acdaab872a09a10ef2a84ded9eee54590d5dec093d4186d20",
    "bench/scripts/_qwythos.sh": "8cc440f0fe6008a50c2b911cdb6fa3ee75c97a5a5c191b189c936cc90ffbe7fb",
    "bench/scripts/accuracy.sh": "f2dbc86363839ef8618be92980d01d457767185ed82f5732652b6a3f32c6a3ae",
    "bench/scripts/accuracy_compare.py": "cb79dcbfcbb4a8f5100196caadd70579f66330d35062447c01ccda35e70c4d89",
    "bench/scripts/audit_baseline_pins.py": "f8bc37cfa52ed8144f81ce3f1cf9af82332ae37330b0585fca7a56db043eddd0",
    "bench/scripts/bench.sh": "ce94cca5ecb2e19087ef7a954e43a77f6ceead76fa130b1b674cfdf6b8615880",
    "bench/scripts/bench_vs_llamacpp.sh": "09d532fb27325c5bf3ca08d57aa50a7f070642bf02bcfdc49a86ada2f6f5422d",
    "bench/scripts/capture_parity_refs.sh": "0bef4bd13b3c1d2d4d1d6e00825354e64b9403ca16c5f0ba315bff172370c367",
    "bench/scripts/check_reference_lock.py": "129d4ff0583833c0d6de818ea5b5987344fbc2b7c0d3c09e561449709f91bb55",
    "bench/scripts/compare_logits.py": "bbada3876d7a8f37d0d8a8c42f37a105ff595a54b783c138be962f4f0aa48833",
    "bench/scripts/dflash_accuracy.sh": "4d2406a216764a9efe51742416809b9c79d15ce2e1730b00dcb6a1dc49e413ea",
    "bench/scripts/dump_ref_logits.cpp": "5082dba0f6050b060f3ad7fbaf1e1f54c7ac5045492ef480a95d1b8ab48bfa23",
    "bench/scripts/dump_ref_logits.sh": "caffce03349c9ae012e1a87a312ad4793c790451fe2c6d2bd4456ec7b9755ddf",
    "bench/scripts/eval_corpus.txt": "509f3e14b6c4936ea0a103ad3daee3c3211e94b3c1f7c9ec42d13da6b1b27fcd",
    "bench/scripts/eval_quant_quality.sh": "6879d134d9fdf68892165a9f68cde44e18f672d76d5eb7feb709afdd948527a8",
    "bench/scripts/eval_text.txt": "25a1956ff87cf83dc7410a1c245bc99dba6e48602ad4c58a430e6016a08960db",
    "bench/scripts/evaluate.sh": "e6cb15760b60f23d110c88e06821fa330c42e031a54a13a990722e0b3efc0200",
    "bench/scripts/evaluate_bidir.sh": "623bdae535c22a1e20204a65cb7c2f9b81fd3213e86ed96cb9820e9bc965a1b1",
    "bench/scripts/evaluate_dual.sh": "8b147b629e4745400840d612a98aebcd92410d0afd2e9f5f53cb600a0fc2367c",
    "bench/scripts/evaluate_triple.sh": "b338d7087325b22335871dfcb2915e40e54248e60e9de95cac8fdef82131f77d",
    "bench/scripts/gen_eval_prompt.py": "8aa9f24958a85e177d238b8bd90dead75687697043eaa67202c839eb4cb131ef",
    "bench/scripts/kimi_k3_attest.py": "51691e0df1f657fe50f29a353459b7972d5a7cfd23db496f130cad44b899cf48",
    "bench/scripts/kimi_k3_baseline.sh": "fdeaad9aa1abe7c7c95a6050759c2e4421fca7a5c171e50fb00e4626c00bfb1a",
    "bench/scripts/kimi_k3_eval.sh": "a174e4a8d8eba180705d846697fefefa96cfa79b4d6eb36f279b2f2975cfafc7",
    "bench/scripts/kimi_k3_reference_server.sh": "84b74a62e2ef2234c4bdd5ad9fc2d15d44712d4dee56feec1d6ab547913fa931",
    "bench/scripts/kimi_k3_run.sh": "f3484ddd4b9f155fa51644452f5afe8b0969f177ef4d3e1e43a615ec6d9da546",
    "bench/scripts/label.py": "19ab4d3ef782358bb5a936b101bcb8e509a07f4ff23cb2be53e0e0057bc46c99",
    "bench/scripts/probe_all_quants.sh": "1a3b2911df318725b445ee66719610d6ec9b9080f80e93a077ec2c660e88d756",
    "bench/scripts/probe_max_ctx.sh": "5b2d6e8be763d5c976aefa5f1bbfe1ed262a43c98f19ae8ba801f96534f0dd42",
    "bench/scripts/probe_max_ctx_llama.sh": "c57dcd632ef3a6e156b6fc68bd0b9fa6a19a1bb0f007cfeebac7c11bbf5ce55d",
    "bench/scripts/qwen36_compare_128_512.sh": "510ea8eec6ee8902aacc6fa65fb44efcb9f6840330941072f1107b339025c25f",
    "bench/scripts/reference.lock": "9aff6c8f356995eab68cd97aaee87810819f840d39d6340f0986a43f1655094c",
    "bench/scripts/run_all.sh": "5a02b57a30f2702d30864c97f61c9c6c7902bbd0cf89000fb7fc88f3394ab5e4",
    "bench/scripts/self_consistency.py": "38c1c59134c8c47a2d50c5b12dd35387b48b5a3a2cb435db502e3633e85204e8",
    "bench/scripts/test_accuracy_cache.py": "77fd13a74e5f24a7771baf15eb82af74c958b6c0aa973eeb3ff2aef43f4920af",
    "bench/scripts/test_bench_sweep.py": "b4d40d3e050d62f470192ccdd67c231275b2c6544d2365ee66ab809788cfa7ac",
    "bench/scripts/test_bidir_headline.py": "e9a096f2c2d0f3e228020d92be2f0775a7ff95df4d0a616a37fb2ea3999afd6f",
    "bench/scripts/test_cb_prefill_score.py": "6f147e3d410dfa8a5f532e4469746f298e61483ba8639b03983b96c59b30a451",
    "bench/scripts/test_guard_reject.py": "e7413fd1cd8a01cba18ea64535dd641c1e3e4fc27871e7d29ac90e2762b96890",
    "bench/scripts/test_kimi_k3_baseline.py": "0473c43eae6dc9484ec084528a462e2e6405a35609ac33e8e0c7798b901ebb99",
    "bench/scripts/test_label.py": "3a8f511a2c381c87519c4d26ca2bb56bf65887622399805993006178b7cbeb58",
    "bench/scripts/test_label_ttft.py": "4d0b571a2b8ed964a1f66211fdd46806a29bf739f3ac20f641caf78d99e28b20",
    "bench/scripts/test_prefill_label_merge.py": "d0fecab3d02275cfe015bbe67264bd494bd4f252d6ecc530d42a75e26c0fd963",
    "bench/scripts/warm_llamacpp.sh": "91ed5b6a9c290c377daabc3132c04a8e1cf016bb5c9292526491a66d07231bc7",
    "eval/README.md": "7935bf40f2cd16a21cb3132183b2c779d975bbd3f303b0659afd2fc9b74eec02",
    "eval/backfill_dashboard.py": "f1bdc054db45295827e489e0bb715828816827bb90793b683afa0daf110ae30b",
    "eval/copycat_guard.py": "fe07561366dca792ae44acf0ef894beb487c1845908382ddbd5153a017231c84",
    "eval/copycat_policy.py": "e8c53091e4ff6925d5d0649f0078bb3ee7953d045cde8e4f03d29124b2dc28d7",
    "eval/copycat_sweep.py": "718efdd214f27029910fd85329e03466cee1e6be4280d65060483dcfff82a301",
    "eval/k3_eval_bot.py": "3067ea50366b850db6bd55204e448b50520af8894ffbcf0c05bbb14061dced37",
    "eval/polaris/__init__.py": "38023fefe7aab517f8aabd3f21c83a430e3c8344acc1c389f610d01b0bf90ca8",
    "eval/polaris/client.py": "11fbe04e812fdfea94e72e98bc9f79c16b17259d6dae786bed215acff0d0f550",
    "eval/polaris/judge.py": "0cbd7b53c5690aa2d3202756122726c3f484fefd8b2ff17683a39f67ff6945cf",
    "eval/polaris/receipt.py": "e9a38413f12f0f0665ef2f5ea5be6cf7df4317a6a4061c949adafe9bbcbf94c8",
    "eval/polaris/scoring.py": "3da819f79756e2608e629983e104f18297f28f5935bb80b7a62b57b75b2580a3",
    "eval/polaris/sparkinfer_eval.pub": "af1463f2a3803fab12d68bc4dfbb46fc13140733b1be19d16736da9466be62a0",
    "eval/polaris/test_receipt.py": "b8c6f955c98bb552e4592737246f93251997112877d70dbaa54b8223aa2abb5d",
    "eval/polaris/verify.py": "c117f00dcf0b6ec5f39c5f6c6d5f4bdd49804c90f06a8280a9ea32ba1d0fde63",
    "eval/pr_dflash_bot.py": "106921a21452be07d4f17361b6e62c218a2ba32547ec5a6d11eb6f2f67807123",
    "eval/pr_eval_bot.py": "2264339bd22ac3f3e544ceea7f23b6b55fa3ea4b347d015c7f197ab022dc0411",
    "eval/run_bot.sh": "f4e7b48dbe69bc223f231350da75087fdebe271a9e905f2b4a8314d151f6a981",
    "eval/run_bot_cron.sh": "33fd9a556bf639471b5714aaaf08e1db06c664b5036af5abb9c7b4def4483ca1",
    "eval/run_dflash_bot.sh": "38f8f423ae37011e1167b5a8b2c009d1172ef93429a3f1353f7e1bda4e93689a",
    "eval/run_dflash_cron.sh": "6bc75a3db97a461d64df0d0a7e25368a615aa623798b2f94a29d11c208562fc2",
    "eval/run_polaris_smoke.sh": "5f377216727ecee4ea15c9541bea9520dcb6443f135c4d377385e06d337e731a",
    "eval/run_polaris_test.sh": "613b8f53926fec6c6beb4401b192a835888e3ba6f22c446e231d60b18f6382b2",
    "eval/run_stale_pr_cron.sh": "7aa874c440b307462073e38fbe6063717c47f086060907976357cec5db4ad576",
    "eval/run_sync_cron.sh": "816d359c0ca469b3994e32ebd388fe1ccc9db386c5afdd64e7805cdaa7d172b0",
    "eval/setup_labels.sh": "5769da0df30628e8d4e44ec9d26cf70a80de65fae41ff1f8ee3f32dd25eab37c",
    "eval/sim_difficulty.py": "0744a0f9f8ff4ad09719444ab96c057c837b1a38b466c28b8b6a17e545df7932",
    "eval/ssh_box.py": "a3585b63c2a7b99528da7bdfa0bbedb8adc9649527fafcfc8674064f011daa71",
    "eval/test_copycat_guard.py": "6507e0b4e96361e18c749fd426f7750522bcbc4fb25b1f21f7b9a0662fd06768",
    "eval/test_pr_eval_bot.py": "69b2c16d5cbc479112e93b1baf14e76062a132a74d3604faa1873e034130dae5",
    "eval/vast_eval.py": "fbd8613cdb1d430304a3946a10fd1bdf92de973c4508f3df157d264e65b06195",
    "kernels/csrc/cuda/kimi_k3/k3_kernels.cu": "eb2ee6cb7b33fb604124668ceaf6425efa39387102211fcce11d7e939677a5c3",
    "runtime/.gitignore": "de0f1bf8170e2ab3986ba8dc860e62df70d6bc6bfa57d0297c6c163451baea65",
    "runtime/CMakeLists.txt": "89b0425e6e14519f19c6f2a13b0dfeeda3b8ea5af268ea43bea96aa048ad1efa",
    "runtime/README.md": "1feeaa0e9fea645ce214a531b2a12e61d7eed5d055f5cc11c6b658eb1904bde6",
    "runtime/csrc/cuda/dflash_kernels.cu": "c61d8f693fb5fc602522e87f9d1c699f4b2c6b85449e16028746457f77e5fe7e",
    "runtime/csrc/cuda/kv_ops.cu": "39f587ed50da2cf3707711a01561a5a017d56c5a2fe8ba216c9ba8678a927fe0",
    "runtime/csrc/tp/multimem_allreduce.cu": "f3dc74601c00a99ce80013b6d3c05b67a270ce695faf503336f8f6c5bf535234",
    "runtime/csrc/tp/peer_oneshot_allreduce.cu": "d093faf3841548a8f22bd53c0403009999787a290e05d4e4b7689143c6ca0e52",
    "runtime/csrc/tp/tp_allreduce.cuh": "33029b6dfb78ab111fd6f39d7c50d9645b690dad0266860715910577aa36ddeb",
    "runtime/csrc/tp/tp_error.hpp": "e784374650ed8a8751d20b3108b9cc8e139e9d0645b87894ca07c2baeee20e3a",
    "runtime/csrc/tp/tp_pdl.cuh": "3bb074bd4f6ae4b294ba9828c651d4fcb3133eeaf00efd10a3ac4c7a3dcb9c4f",
    "runtime/examples/kimi_k3_generate.cpp": "e9e7145103e669ac0d63413a4ab62a980f732e8c44988acd85351bb3bdf87fb4",
    "runtime/examples/kimi_k3_layer0_ref_check.cpp": "c3c62b1f2b5262f8c154e95e5e0ae8f7fa3931a2a8919932146f682f2f8a692a",
    "runtime/examples/kimi_k3_mla_ref_check.cpp": "c22426792964c93ae64188f4bafc8cf16e21844387dd405ab22794b7ce011932",
    "runtime/examples/kimi_k3_moe_ref_check.cpp": "883dcce3c3ca9e0d6684159059ef87940e4e4314aa36d424f0b719d547753570",
    "runtime/examples/kimi_k3_pipeline_check.cpp": "c75f459b51f2298ded6db8b2d97bc0bec1689cdd68b70f4f7a437b1522efc9fd",
    "runtime/examples/kimi_k3_plan_check.cpp": "6b6a08d76570af2fe6c8f7bbc10dcfe7135c6753aa71d8728b0fc9ee3748a6bf",
    "runtime/examples/kimi_k3_smoke_test.cpp": "7e246eed93f3dd79dacbd6b476bdd250180c95283846312d4f88ba9d34192452",
    "runtime/examples/kimi_k3_state_check.cpp": "f8a75201a3d632fd95e4b454e7f8b3ad8592cf4a2be6ad3aba0c5f74694ac28d",
    "runtime/examples/kimi_k3_tp_bench.cpp": "3dd4fc43ee7c3a63dfd23661ad899007e538d42f997bb1fc61ab3e5726e08126",
    "runtime/examples/kimi_k3_tp_check.cpp": "2dea561aeb4080a417eba469ab723bfad04915b998867fd9bee180c68f24a381",
    "runtime/examples/kimi_k3_tp_load_check.cpp": "6a926c33b4c9e31633340a74bd9752aa2a4c59a4d75357690b805db8ea974b58",
    "runtime/examples/kimi_k3_tp_moe_check.cpp": "c024d7b899fb4c7b45380d5695348eaae7f8072e251b91505cd2f3fdf300f7c2",
    "runtime/examples/kimi_k3_tp_phase_check.cpp": "9075ed3a4b5120fe21b284a092942cfa3e863c2061d82b62da3e52445219a431",
    "runtime/examples/qwen35_generate.cpp": "1e18ab9bf87e8de6c1b83561f9d3730effdd354182dcaec5d1bc01a3d6450a94",
    "runtime/examples/qwen3_gguf_bench.cpp": "9f1bd3ff8900465caf83d3cbe7e887e6b2d90a9c783f4192454e7c616b4e8c7e",
    "runtime/examples/qwen3_gguf_cb_bench.cpp": "3aa0e4bd57783db2a3637ebd499e60113a0623fa302220772466d5185109aa1f",
    "runtime/examples/qwen3_gguf_config.h": "1d2793ce8c6d201695fa9007c2cabb7da256428212a397ee0ee50afca6795ea2",
    "runtime/examples/qwen3_gguf_dflash_bench.cpp": "0bdac87c25901979b4594bfbb65a0af65570a98a6ab4ea22cd56a9883d35a6bf",
    "runtime/examples/qwen3_gguf_dflash_check.cpp": "b93b914a148729c0a1aef1a7b3470fb994efe444bd9d66eba80ecb5fc36f1886",
    "runtime/examples/qwen3_gguf_generate.cpp": "61bca02913dd1c87b1771dba814c6983e3df1e75bfb640dbe4002a20f2570ade",
    "runtime/examples/qwen3_gguf_prefill_check.cpp": "aea373e7761f5ea0b65ca331607dcf27984df63672615010f7f4d44ad5b9e70e",
    "runtime/examples/qwen3_gguf_score.cpp": "2061a3d3e93ca66f2239f3702b63831550da17c047a1d0fb6012edfc4025bc35",
    "runtime/examples/thermal_sweep.cpp": "3bd5397efd15b6918b5477c4c1f04af689d88f184150e67c10a6398b7cef0c3c",
    "runtime/examples/tp_allreduce_check.cpp": "b21f837549096e37cb291f7450bab53ec59f9017c1a5235850fa1429af6d3a15",
    "runtime/include/sparkinfer/decode.h": "055ddec19069d3c02adba4b286b89c7013ad3cac807d004680e40e5938b6fdb0",
    "runtime/include/sparkinfer/gguf.h": "2868c96938abd0582d3b0f0da60dc8517aae1bb9e498165d7be360c9cf08c28b",
    "runtime/include/sparkinfer/gpu_stats.h": "c38f2036e44ee842e15509b3a5c1d92474ef3a2dc1351c16d05d304d05aace5e",
    "runtime/include/sparkinfer/inference_engine.h": "e803af47eed5ea0e8df2e6f70b2b0a1ef8bc554ed97bc6c5733d0d8cca4deba1",
    "runtime/include/sparkinfer/kv_cache.h": "6e3bda0223b8f694d23aec6d1a05b2bf1a912eb2f96745677a3df1b446228877",
    "runtime/include/sparkinfer/kv_ops.h": "214ebc47ea9972e6893477c43aca8d787e72d703e85b5eb9dd8be97375255c22",
    "runtime/include/sparkinfer/models/dflash_draft.h": "3c7865ed1dda45672840d3060012dc027351323e00f668ce3df84eec694946fc",
    "runtime/include/sparkinfer/models/dflash_kernels.h": "e76373ed8e61a2fecd74acefc2bdf589a8a4fde1c9181c3b675ff0870d37d386",
    "runtime/include/sparkinfer/models/k3_head_band.h": "70848840780eb497afd2c7039bb3a35ca23dfefce1eb7ed0fc9ad89fbd1a0f59",
    "runtime/include/sparkinfer/models/kimi_k3.h": "b09dee52522d94accdbab61ed08a0f3b92a55f8325f15d331b20178ee6ba719d",
    "runtime/include/sparkinfer/models/kimi_k3_config.h": "06855762875e1919f33b669cf8950094972f7ac6fb7f49a792d5ce5048007f89",
    "runtime/include/sparkinfer/models/kimi_k3_decode_plan.h": "cd374ae83aa5d0effdb358983d02c65f95e4bbe0dd958ed47866e093b3783a3a",
    "runtime/include/sparkinfer/models/kimi_k3_gguf_config.h": "fac8633d614b6a09eabf086edc9ba96700365ac086eac5dfe6de5c59d54dae80",
    "runtime/include/sparkinfer/models/kimi_k3_gguf_manifest.h": "58f4e8843615762734b8b596ef00e2ed350f0270a4cb27cd41516d9957aa05c9",
    "runtime/include/sparkinfer/models/kimi_k3_prefill.h": "0bc550e361796d4250d976ad80608fc8dd8947addd4c24eed4671e4cbe963c1f",
    "runtime/include/sparkinfer/models/kimi_k3_tp.h": "1a825ecae4f04e9fdbaabcb8e12bb3d739c73f3bb0632d4b5182a734a4cb86d1",
    "runtime/include/sparkinfer/models/kimi_k3_vocab.h": "cdff52b21be819ea4329d1490f02227b76b47376c8085557c4ef6cf635c175f8",
    "runtime/include/sparkinfer/models/qwen35.h": "acb085dcdeb1866d93ee99dbd1cd360e123dea13a0eb6ba81de7b7e7ad473d49",
    "runtime/include/sparkinfer/models/qwen_config.h": "26be4b9040f269684227cf995445e6c46bac796a69bbc3c8250b8bb5f1de475d",
    "runtime/include/sparkinfer/runtime.h": "5d05175084ff5e868d433a2a52a7e216e40808ffc5e338516b861250d1882a7b",
    "runtime/include/sparkinfer/scheduler.h": "9258bf78f44ca3553efd93ff6d3c870f532c1d87e00334aa757ec7065e4fdf87",
    "runtime/include/sparkinfer/thermal_governor.h": "905d373688f86a3e7e1c33a0e80cc0d3e330b55aa8c28fae0df1f6fdf94ff7cf",
    "runtime/include/sparkinfer/tp/backend_select.h": "c8b7ebe3d0f5791d47fb1124fc8f4a0da18581a3b8eb370bdb569ac22a0f8b31",
    "runtime/include/sparkinfer/tp/collective.h": "2346b37df09a2369ee6d35d25f241c506c92132283356eff7f942b08670f11aa",
    "runtime/include/sparkinfer/tp/k3_coll_1bar.h": "fd31470362943755cf076722b90c361f9e8b0927c4b26b589c98f07b15efcdbf",
    "runtime/include/sparkinfer/tp/multimem_allreduce.h": "f412e7ff0dae57a3b7accfc79e8a60ca87d1c5b6e204808a7f39021999ddccd1",
    "runtime/include/sparkinfer/tp/peer_oneshot_allreduce.h": "da54a9c072c9fff29b78c850b9bd096f9220259f55cda8e1a5d8453dddfbe18c",
    "runtime/include/sparkinfer/tp/shard.h": "fdee71c4786477fae2dfc9a9591ae90639859baa364a3305c47f6ab4511f7c71",
    "runtime/include/sparkinfer/tp/weight_plan.h": "4142f4570d96a2bd685ec8da4fc56bcd45fe61971c7ac78d86528faae0ead727",
    "runtime/include/sparkinfer/tp/weight_residency.h": "fd057400420ad3c83b8beec2d20d676a6a02360c18e04e07b1f91140872a1d74",
    "runtime/python/sparkinfer_runtime/__init__.py": "53fff41e9c157b5618a1b073d232a8b820de2993fb3e14ed46cf703ff3dcd291",
    "runtime/scripts/bisect_qwen36_acc.sh": "a977878c42a65704b9788b8cb48056ff654a7ed902afbf6e4ead56e3f2a355d9",
    "runtime/scripts/build.sh": "6d08c96ead523d52e067d2a50b764a73d7d9858960b7c970c4a318b3793501e3",
    "runtime/scripts/cmp_argmatch.py": "8201eb022bdb7d27de9d50b7810580e572da6208eb7e56f6ebb7debd1580ce25",
    "runtime/src/decode_layer.cpp": "c18380bb37234fecccab6308357eb6f380d579a93dca5896e9cd2a445ced4680",
    "runtime/src/gguf.cpp": "52a8bf3afc4bec976737b71f0f725a80a8d41ec0782b33dc8f9d2f8c3b035f71",
    "runtime/src/gpu_stats.cpp": "9b5898b07415387602899c8022d740c3f6896401771512af4d6fea012e767d96",
    "runtime/src/inference_engine.cpp": "3adc80a14197d60d002d51d3a42762b3c85c762eeff232a611ff0043ea8add44",
    "runtime/src/kv_cache.cpp": "e44d0a7fd6ee15e2c7b3bfa23c0be94d0a7f4246fd32f6e401d82be6e41f06bf",
    "runtime/src/models/dflash_draft.cpp": "7f52d3fdfc04cb3ab9601f4d0e6360b47afc37a899a5b659fd560f7a511d0183",
    "runtime/src/models/k3_head_band.cpp": "f3000fa9e35406c9acd1e68927477c115ee655f935e114f14b6dd1ce42156d5b",
    "runtime/src/models/kimi_k3.cpp": "2df262bd223e67e76f9f79ed5344f92c037d17d9cd5ce7a5350e2b1eb5a21031",
    "runtime/src/models/kimi_k3_decode_plan.cpp": "0375cb13d1495c54af452f35d08bbbabc793399f28347387864e2ada8b6cbc43",
    "runtime/src/models/kimi_k3_prefill.cpp": "6c40bf5e041246cb6d939a059f35567be1c8ea56c805bdbb83b70b50407b5d6f",
    "runtime/src/models/kimi_k3_tp.cpp": "b0038a34e4a15f4606316cda41e37e534d9eb0d732dc1ccaf1a7c57eb1a85e36",
    "runtime/src/models/kimi_k3_vocab.cpp": "5c05c3eea1e054c7f108bbf6ecb524f737bfc008ff9b7519ce8d8811958ca643",
    "runtime/src/models/qwen35.cpp": "67bb97ac0c843dd420d026eab69e6af293f226d375b5360f034916168a1eadfb",
    "runtime/src/models/qwen35_prefill.cpp": "e7f0e4d17c604e6daaa625a202ceee28f6a5f77700df11051e417658784579d4",
    "runtime/src/models/qwen35_prefill.h": "a5ecaebc568480a382b6b9b5d9392a9440d96a2d44984509a1edb2d1a01117a8",
    "runtime/src/runtime.cpp": "d10e8037fd7f9f10fe954c86f430e765f8a3596cacf3424c683847d9ad4529b1",
    "runtime/src/scheduler.cpp": "a42fc53d8e2d0623d0969c9d592b732b515831096dde2b4429ef708ade4d9665",
    "runtime/src/thermal_governor.cpp": "85ad98035532b86df053929ffa7239c55b4498edc478d259bff048d84a8f0fb8",
    "runtime/src/tp/backend_select.cpp": "074656aff114fba1943adb10c1fc6605f97e9d302b0f47237eddab864398807b",
    "runtime/src/tp/collective.cpp": "2147cffe9212ba81149bdda86cb26e03e0c4ca2b3f17f429350c29bad1be53c6",
    "runtime/src/tp/k3_coll_1bar.cpp": "2587ddd57050623857116299e11536a06821a7a5d5af397f07e315e7c1c23651",
    "runtime/src/tp/shard.cpp": "09ef461af0dfd00719a4460db361ab2de6be79dafe3b3387ee2c4f61c76cb810",
    "runtime/src/tp/weight_plan.cpp": "55073734c3cf6d2695858d4b36f04d4f2b18fcbe0f03253aef51dbf8cbe83449",
    "runtime/src/tp/weight_residency.cpp": "1d9bd182de1142630104073ed45b85fce55ea251ff4a43d55ce45b8f90e8a4f6",
    "runtime/tests/CMakeLists.txt": "12acbcb06d14ec9160da21683e93aa5757e30360df85705422ea74b9993a45ef",
    "runtime/tests/batch_engine_gpu_test.cpp": "a13d273cbfad14972acb7a00e069d97cb782fdf830fe696a9392d386c441b103",
    "runtime/tests/decode_layer_cpu_test.cpp": "06597126ec74d8f212232925a034ebaccea1fe23c3b68cc024c428c80d68ea7a",
    "runtime/tests/decode_runner_gpu_test.cpp": "c3b65311637757bb84768e4bebe9c033fefff740468ef259368bf4b39a61d6df",
    "runtime/tests/kimi_k3_config_cpu_test.cpp": "68770b5c5e4fd22e9d6697b56663a66ca8e9d37309d7b90f19f50bde2ed5242e",
    "runtime/tests/kimi_k3_decode_plan_cpu_test.cpp": "c08d10ecb63fd550b140af83e8644e94ef8bfa28fe121e691133a95fa956248f",
    "runtime/tests/kimi_k3_manifest_cpu_test.cpp": "3991100e6737b70abfd8a285dc00a8715e397ba849398a5a4a4a2409d7d85679",
    "runtime/tests/qwen35_cpu_test.cpp": "817baa41890630ebd43a21da339fc5d3315533da8c55f707bab955800a276f4a",
    "runtime/tests/qwen35_gpu_test.cpp": "e7ab54fcf4d2c40cc82971ba02b283eeef0e3f409c3b72b749c14baf32dd638a",
    "runtime/tests/qwen36_config_cpu_test.cpp": "c42b8415a09682768616b266f62e825d7dc4226ad5b6ed8ee9fde3f3bb5c0a2d",
    "runtime/tests/scheduler_cpu_test.cpp": "88101e8f5bb55758e76464e2ab9bca3faa6bfb8a1c2376ea7834d5ea89c9072f",
    "runtime/tests/thermal_governor_cpu_test.cpp": "faed136048617febbc556cdd8154f4fc87d859634a9b8245d83544926ff0c452",
    "runtime/tests/tp_backend_cpu_test.cpp": "7815ce0af1834757afa3080e1aecb913349245f26a73808feb2e3a9b9bbd6977",
    "runtime/tests/tp_shard_cpu_test.cpp": "bbf0687f06d27b41ec1fd6f23c073d9fede185c1d14d754699a20e222e5e3ee4",
    "runtime/tests/tp_weight_plan_cpu_test.cpp": "a252b5c6b8389d3726b68bdb83ecd1d4bbf624295fc0f1d22a11789d125fa280",
    "runtime/tests/tp_weight_residency_cpu_test.cpp": "b16c5ced10cb4661560e846db04d92f7dd583699f6fcd40d305b593391441725",
    "runtime/tools/convert_gguf.py": "690a0c9d52e931302b57b5271ce477db17c635659c9d4fd557a9770fbf760bdf",
    "runtime/tools/convert_qwen35.py": "001cd6c0a4b47659bc5b79112ce1c750fd5170caeec9f17519cf589e909415e1",
    "runtime/tools/run_qwen3.py": "edc3c5c7c8b2cf34b7efacf51551b133647b4fb7ee35591fa3c2d2efbfa9a919"
  },
  "passed": true,
  "upstream_python_test_count": 222,
  "native_executables_passed": 7,
  "printed_native_assertions": 10473,
  "native_assertion_count_note": "Six executables print numeric counts; the seventh config fixture reports OK without a count.",
  "counterexample_tests": {
    "command": [
      "python3",
      "-m",
      "unittest",
      "discover",
      "-s",
      "Phase 5 - Advanced Topics and Specialization/Track G - ML Systems Engineering/AI Inference Engineer 2026/Part 4 - Optimizing a Real Engine/labs",
      "-v"
    ],
    "passed": 4,
    "scope": "CPU predicate/combine counterexamples, not K3 GPU inputs or execution"
  },
  "official_model": {
    "revision": "c5d1dd4c428bd1ce8b88c5044f3b6ccde9e3b721",
    "model_card_url": "https://huggingface.co/moonshotai/Kimi-K3/blob/c5d1dd4c428bd1ce8b88c5044f3b6ccde9e3b721/README.md",
    "model_card_sha256": "538f241e6e4d03a0de22db6d668ca3d58e2221fe913f0df2b0f60b995ad0ee6e",
    "config_url": "https://huggingface.co/moonshotai/Kimi-K3/blob/c5d1dd4c428bd1ce8b88c5044f3b6ccde9e3b721/config.json",
    "config_sha256": "9710e121a58d03ac92c8d6da287a19541994319afbbe6d6202af001ffd379213",
    "total_parameters_reported": "2.8T",
    "active_parameters_reported": "104B",
    "legacy_engine_A50B_is_not_authoritative": true
  },
  "reviewed_findings": [
    {
      "claim": "553 GiB divided by 8 is about 123 GiB",
      "status": "incorrect arithmetic; ideal share is 69.125 GiB, runtime residency is separate"
    },
    {
      "claim": "expert threshold guarantees small relative output error",
      "status": "false without expert-output norm bounds and a nonzero output lower bound; four executed counterexamples"
    },
    {
      "claim": "byte identity below 8192 validates threshold beginning at 16384",
      "status": "uncovered branch; require real-prefix tests after gate activation"
    },
    {
      "claim": "aggregate naive roofline gap rules out memory limitations",
      "status": "not established; use per-rank traffic, latency, concurrency and resource measurements"
    },
    {
      "claim": "same floating KL proves identical logits",
      "status": "false; compare actual arrays"
    },
    {
      "claim": "CUDA stream capture executes the recorded operations",
      "status": "false; capture records, graph launch executes"
    },
    {
      "claim": "reported 60.17 tok/s is reproduced by this refresh",
      "status": "not run; historical author-reported synthetic-cache GPU number"
    }
  ],
  "course_source_sha256": {
    "CPU-Source-Lab.md": "66dec5df439f49ee113454991adfff0c138ebb55775e6e89c9c9f8492e0db1e8",
    "Lecture-01.md": "9953012cc8378ed299c58abe42d913db90b75d2c84631048f23f036e7e58a585",
    "Lecture-02.md": "2810b7c00664155166b52499806dc0ac211ca65152f4bd3f153f6a40759c9ef4",
    "Lecture-03.md": "a0c6694fa54d6179ee634086c78b4d58b5b2d0d9d25413ed10a9cc859e069603",
    "Lecture-04.md": "a628833f7142ebfb3eb564f8cdd074bbd4ffd17d86ab0c9198dafdec9555e768",
    "Lecture-05.md": "4d75b1569aa5511e00bac57f08559acf3f1e7080f08f2381789a50b2b0fa398d",
    "Lecture-06.md": "2d44749ecbb715412439f374c62f6e4dc14d1733f2768f0434c6c2a72b2225c2",
    "Lecture-07.md": "2a3fb296272dbc9815a3cea37bfb62511814722fd03658e990199038cf035706",
    "Lecture-08.md": "74fd23856b1b19e1c1420c260baf85ba56986a944cf5f2b4c27598abbee8c8f6",
    "Lecture-09.md": "22d33d37215de6435bf3c0427c6a96c2e5a558b0ef04e6547d8ec1ae3cbeacff",
    "Lecture-10.md": "7562412008a6b35e000d0ccc2e5b9dfd97c99724bea8198112bbea609f7df2a6",
    "README.md": "21a9cbff4f45d475d2be3c914ed48cf5debf1e54e42aad2ca6dae36360b4b56f",
    "labs/run_upstream_cpu.py": "288fd0de45effa375e3a29d7ff1d50831c6e05e634c1c9b96aaf0db733e3a725",
    "labs/test_approximation_contract.py": "5d7ae4135e4fc895ec84ebde1da0862c8de274b952db76368e477acae837ea9a"
  },
  "negative_checks": [
    {
      "name": "runner rejects a different git revision",
      "exit_code": 2,
      "output_report_created": false,
      "observed_error": "expected 7689cc77e0e18d4b4f1937cf2a189dcd5a8d90f3, got 99336716c0e8990bcdf997623ab92e8ceffec5e8"
    }
  ]
}
