h3-blackwell-runtime/research/experiment_registry.json

{
  "schema": "h3-blackwell-runtime-experiment-registry",
  "version": "1.0.0",
  "generated_at": "2026-08-25",
  "allowed_statuses": [
    "production_accepted",
    "research_retained",
    "architecture_rejected",
    "performance_rejected",
    "quality_rejected",
    "temporarily_blocked",
    "incomplete"
  ],
  "experiments": [
    {
      "id": "elem-fused-residual-rmsnorm",
      "name": "Fused residual + RMSNorm",
      "family": "elementwise-fusion",
      "status": "production_accepted",
      "hypothesis": "Combining exact H3 normalization/modulation and residual-gate work reduces tensor traffic without changing model arithmetic.",
      "implementation_strategy": "Use the deployed fused H3 elementwise kernels while preserving explicit BF16 boundaries.",
      "source_locations": ["src/h3_blackwell_runtime/block.py", "src/h3_blackwell_runtime/h3_fusion.py", "PERFORMANCE_ROADMAP.md"],
      "active_source_location": "src/h3_blackwell_runtime/h3_fusion.py",
      "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-h3-fused-elementwise-1344x768-124f-seed440420.json", "exists": true}],
      "profiler_artifacts": [{"path": "benchmarks/gb10-fused-block-residual-profile-component2-1344x768-124f-seed440420.json", "exists": true}],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "dirty working tree at registry generation"},
      "metrics": {"result": "accepted; exact measurements remain in linked artifacts"},
      "correctness_evidence": ["Repository current-state documentation records bit-exact blocks and checksum-identical trajectories."],
      "decision_rationale": "The exact path passed block, trajectory, deployment, and memory gates and is selected for Spark production.",
      "reproducer_commands": [],
      "timestamp": null,
      "evidence_missing": ["Portable validation outside GB10/SM121"],
      "production_behavior": "Enabled for the selected Spark runtime through the fused elementwise path.",
      "source_recovery": "Active implementation and validator remain in the working tree."
    },
    {
      "id": "nvfp4-quantize-alone",
      "name": "NVFP4 quantize alone",
      "family": "nvfp4-projection",
      "status": "research_retained",
      "hypothesis": "An owned standalone quantizer can match the reference packing contract and expose scale/packing cost.",
      "implementation_strategy": "Discover scales and emit native QDATA/SFA independently of GEMM.",
      "source_locations": ["src/h3_blackwell_runtime/nvfp4_quant.py", "tools/profile_nvfp4_linear.py", "PERFORMANCE_ROADMAP.md"],
      "active_source_location": "src/h3_blackwell_runtime/nvfp4_quant.py",
      "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-nvfp4-pack-parity-component2-fixed-1344x768-124f-seed440420.json", "exists": true}],
      "profiler_artifacts": [],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "prototype infrastructure in dirty working tree"},
      "metrics": {"conclusion": "packing contract established; standalone speed not consistently favorable"},
      "correctness_evidence": ["Current roadmap records packed-bit exactness at H3 widths."],
      "decision_rationale": "Useful as reference infrastructure, but not selected as an independent production optimization.",
      "reproducer_commands": [], "timestamp": null,
      "evidence_missing": ["Standalone canonical deployment gate"],
      "production_behavior": "Not dispatched as a standalone replacement.",
      "source_recovery": "Implementation is present in nvfp4_quant.py and historical commits named in git log."
    },
    {
      "id": "nvfp4-quantize-gemm",
      "name": "NVFP4 quantize + GEMM",
      "family": "nvfp4-projection",
      "status": "research_retained",
      "hypothesis": "Pairing the owned packing contract with the reference GEMM establishes an exact projection baseline.",
      "implementation_strategy": "Wrap native packed activations for Comfy Kitchen/CUBLAS NVFP4 GEMM and compare outputs.",
      "source_locations": ["src/h3_blackwell_runtime/nvfp4.py", "src/h3_blackwell_runtime/nvfp4_quant.py", "benchmarks/gb10-nvfp4-native-fixed-component2-1344x768-124f-seed440420.json"],
      "active_source_location": "src/h3_blackwell_runtime/nvfp4.py", "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-nvfp4-native-fixed-component2-1344x768-124f-seed440420.json", "exists": true}],
      "profiler_artifacts": [{"path": "benchmarks/gb10-nvfp4-component2-profile-1344x768-124f-seed440420.json", "exists": true}],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "reference/prototype path"},
      "metrics": {"conclusion": "exact baseline retained; no standalone production win documented"},
      "correctness_evidence": ["Roadmap records exact outputs at all four H3 projection widths after swizzle correction."],
      "decision_rationale": "Retained as the exact baseline and seam for producer fusion rather than promoted as a separate backend.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Isolated accepted performance gate"],
      "production_behavior": "Production continues to use the established NVFP4 linear dispatch.",
      "source_recovery": "Active wrapper and quantizer are present."
    },
    {
      "id": "nvfp4-prequant-swiglu-gemm",
      "name": "Fused prequantized SiLU x up + quantize + GEMM",
      "family": "nvfp4-producer-fusion",
      "status": "production_accepted",
      "hypothesis": "Recomputing exact SwiGLU BF16 boundaries inside NVFP4 production removes the complete FC2 activation intermediate.",
      "implementation_strategy": "Fuse SiLU/multiply into scale and packing while retaining the exact Comfy FC2 GEMM.",
      "source_locations": ["src/h3_blackwell_runtime/block.py", "src/h3_blackwell_runtime/nvfp4_quant.py", "NVFP4_SWIGLU_FUSION_DESIGN.md"],
      "active_source_location": "src/h3_blackwell_runtime/nvfp4_quant.py", "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-nvfp4-swiglu-fusion-summary.json", "exists": true}],
      "profiler_artifacts": [{"path": "benchmarks/gb10-fully-fused-fresh-nsight-summary.json", "exists": true}],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": "sha256:1d340e14cb6fc45ccfdbe63dde8db2a2b3aea94b493702c8a08e1f8d5b4f7b83", "commit_context": "dirty working tree at registry generation"},
      "metrics": {"result": "accepted; measurements are in the linked summary"},
      "correctness_evidence": ["Producer bytes and warmed trajectory outputs are documented as bit-identical."],
      "decision_rationale": "Passed exactness, block, canonical trajectory, and deployment smoke gates.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Non-GB10 architecture validation"],
      "production_behavior": "Selected Spark deployments enable H3_NVFP4_SWIGLU_FUSION=1 for eligible base requests.",
      "source_recovery": "Active source and design document are present."
    },
    {
      "id": "nvfp4-prequant-modulate-gemm",
      "name": "Fused modulation + quantize + GEMM",
      "family": "nvfp4-producer-fusion",
      "status": "production_accepted",
      "hypothesis": "Recomputing the exact BF16 AdaLN modulation boundary inside NVFP4 scale discovery and packing removes the materialized modulated QKV and FC1 inputs.",
      "implementation_strategy": "Fuse segment-specific modulation into native QDATA/SFA production while retaining the unchanged Comfy NVFP4 GEMMs.",
      "source_locations": ["src/h3_blackwell_runtime/block.py", "src/h3_blackwell_runtime/nvfp4.py", "src/h3_blackwell_runtime/nvfp4_quant.py", "NVFP4_MODULATE_FUSION_DESIGN.md"],
      "active_source_location": "src/h3_blackwell_runtime/nvfp4_quant.py",
      "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-nvfp4-modulate-fusion-summary.json", "exists": true}, {"path": "benchmarks/gb10-nvfp4-modulate-fusion-deployment-smoke.json", "exists": true}],
      "profiler_artifacts": [],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": "sha256:5f879c43374bcedb89745971d7d95d82afc8fcf9c41d30f11b257c2863b9fe28", "commit_context": "active source and production manifests"},
      "metrics": {"packed_tokens": 37810, "block_24_producer_p50_ms": 6.018611951731145, "alternating_block_p50_improvement_percent": {"block_0": 0.2794895723218782, "block_24": 0.6017027236613792, "block_49": 0.7939711616055223}, "warmed_two_step_seconds": {"baseline": 47.66834012803156, "candidate": 47.41923755296739, "improvement_percent": 0.5225744684944189}, "warmed_twelve_step_seconds": {"baseline": 285.99590051098494, "candidate": 284.4078275830252, "improvement_percent": 0.5552782138213641}},
      "correctness_evidence": ["The linked summary records equal tensor scales, zero QDATA and SFA differences for blocks 0/24/49, and bit-identical warmed two-step and twelve-step video/audio tensors."],
      "decision_rationale": "The exact producer fusion improved all three alternating block medians and warmed two-step and canonical twelve-step trajectories, then passed the deployment smoke test.",
      "reproducer_commands": [],
      "timestamp": null,
      "evidence_missing": ["Distributed, chunked-MLP, full-precision, pre-scale, autograd, and active-LoRA cases retain the materialized fallback"],
      "production_behavior": "Spark single-GPU manifests enable H3_NVFP4_MODULATE_FUSION=1; ineligible cases retain the materialized fallback.",
      "source_recovery": "Active dispatch, wrapper, quantizer, validators, design document, and benchmark artifacts remain in the working tree."
    },
    {
      "id": "lora-fused-down-activation",
      "name": "Fused LoRA down + activation",
      "family": "lora-fusion", "status": "incomplete",
      "hypothesis": "Fusing LoRA down projection with activation production may avoid a materialized BF16 input.",
      "implementation_strategy": "Produce LoRA-down operands at the activation boundary without changing LoRA arithmetic.",
      "source_locations": ["src/h3_blackwell_runtime/lora.py", "TURBO.md", "PERFORMANCE_ROADMAP.md"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [],
      "environment": {"gpu": null, "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [],
      "decision_rationale": "The repository documents materialized LoRA fallback behavior but no isolated conclusion for this fusion boundary.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Recoverable candidate source", "Benchmark artifact", "Parity evidence"],
      "production_behavior": "Not selected; active LoRA uses the materialized fallback.",
      "source_recovery": "Search history and external experiment storage; no candidate source was found in this checkout."
    },
    {
      "id": "lora-fused-up-residual",
      "name": "Fused LoRA up + residual",
      "family": "lora-fusion", "status": "incomplete",
      "hypothesis": "Applying LoRA-up output directly at the residual boundary may eliminate an intermediate.",
      "implementation_strategy": "Fuse LoRA up projection accumulation with the exact residual update order.",
      "source_locations": ["src/h3_blackwell_runtime/lora.py", "TURBO.md", "PERFORMANCE_ROADMAP.md"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [],
      "environment": {"gpu": null, "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [],
      "decision_rationale": "No defensible performance or exactness conclusion is present for this materially distinct boundary.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Candidate source", "Benchmark", "Residual-order parity"],
      "production_behavior": "Not selected; production applies LoRA through the existing fallback.",
      "source_recovery": "No candidate source found in current tree or searched git history."
    },
    {
      "id": "lora-fused-full-down-up",
      "name": "Fused full LoRA down/up",
      "family": "lora-fusion", "status": "incomplete",
      "hypothesis": "A single fused down/up path can reduce launch and materialization overhead for active adapters.",
      "implementation_strategy": "Execute both low-rank projections in one owned path while preserving rank, alpha, and BF16 boundaries.",
      "source_locations": ["src/h3_blackwell_runtime/lora.py", "TURBO.md"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [],
      "environment": {"gpu": null, "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [], "decision_rationale": "No complete candidate or conclusive artifact was found.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Implementation", "Timing", "Turbo trajectory parity"],
      "production_behavior": "Not selected.", "source_recovery": "Recover from external research checkout if it existed."
    },
    {
      "id": "lora-direct-materialization-gemm-prepass",
      "name": "Direct LoRA materialization into GEMM prepass memory",
      "family": "lora-fusion", "status": "performance_rejected",
      "hypothesis": "Materializing exact BF16 LoRA input inside fused producer prepass memory can retain producer fusion for Turbo requests.",
      "implementation_strategy": "Add optional BF16 materialization while modulation/SwiGLU producers pack NVFP4.",
      "source_locations": ["src/h3_blackwell_runtime/nvfp4_quant.py", "PERFORMANCE_ROADMAP.md", "CURRENT_STATE.md"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-nvfp4-lora-producer-fusion-turbo4-isolated.json", "exists": true}], "profiler_artifacts": [],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "prototype reverted from deployment"},
      "metrics": {"result": "bit-exact but slower; exact measurements remain in artifact"},
      "correctness_evidence": ["Canonical Turbo-4 trajectory documented as bit-exact."],
      "decision_rationale": "Extra BF16 writes cost more than the removed standalone producer.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Profiler report isolating write cost"],
      "production_behavior": "Rejected; active LoRA retains exact materialized fallback and disables producer fusion.",
      "source_recovery": "Historical behavior is documented; prototype is not an active dispatch."
    },
    {
      "id": "fc2-fused-activation-residual",
      "name": "Fused FC2 activation/residual",
      "family": "projection-epilogue", "status": "incomplete",
      "hypothesis": "An exact FC2 epilogue can combine activation projection, gate, and residual while reducing traffic.",
      "implementation_strategy": "Round GEMM output to BF16, apply gate with the reference boundary, then add BF16 residual in reference order.",
      "source_locations": ["PERFORMANCE_ROADMAP.md", "NVFP4_SWIGLU_FUSION_DESIGN.md", "src/h3_blackwell_runtime/block.py"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [],
      "environment": {"gpu": null, "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [],
      "decision_rationale": "The numerical contract is documented, but no integrated epilogue candidate passed a gate.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Owned FC2 epilogue source", "Randomized parity", "Block and trajectory benchmarks"],
      "production_behavior": "FC2 GEMM and residual gate remain separate production operations.",
      "source_recovery": "Design requirements are in PERFORMANCE_ROADMAP.md."
    },
    {
      "id": "qkv-fused-projection-layout",
      "name": "Fused QKV projection + layout",
      "family": "qkv-layout", "status": "architecture_rejected",
      "hypothesis": "A post-projection fused layout kernel can remove three QKV materializations.",
      "implementation_strategy": "Write or copy interleaved projection output into backend-specific Q/K/V layouts.",
      "source_locations": ["CURRENT_STATE.md", "PERFORMANCE_ROADMAP.md", "src/h3_blackwell_runtime/attention.py"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-sage-strided-nhd-summary-1344x768-124f-seed440420.json", "exists": true}], "profiler_artifacts": [],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "superseded by views"},
      "metrics": {"conclusion": "post-GEMM copy does not remove the projection boundary"},
      "correctness_evidence": ["The accepted strided-NHD path demonstrates copies are unnecessary for Sage2."],
      "decision_rationale": "A separate layout writer is architecturally inferior to direct strided views or a true GEMM epilogue.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Historical candidate-specific benchmark"],
      "production_behavior": "Production uses projection-strided NHD views instead.",
      "source_recovery": "Superseded concept; use current attention.py for the accepted architecture."
    },
    {
      "id": "qkv-fused-projection-rmsnorm",
      "name": "Fused QKV + RMSNorm",
      "family": "qkv-preparation", "status": "incomplete",
      "hypothesis": "Fusing QKV output production with Q/K RMSNorm can avoid reading and writing full Q/K tensors.",
      "implementation_strategy": "Apply per-head RMSNorm as part of projection output handling before attention layout formatting.",
      "source_locations": ["PERFORMANCE_ROADMAP.md", "src/h3_blackwell_runtime/attention.py", "SAGE2_BLACKWELL_DESIGN.md"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [],
      "environment": {"gpu": null, "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [], "decision_rationale": "No projection-integrated implementation or gate was found.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Candidate source", "Projection-integrated parity", "Timing"],
      "production_behavior": "Production runs the established fused RMSNorm/RoPE operator after projection.",
      "source_recovery": "Only the design target is recoverable from current docs."
    },
    {
      "id": "qkv-direct-sageattention-layout",
      "name": "Direct QKV output into SageAttention layout",
      "family": "qkv-layout", "status": "production_accepted",
      "hypothesis": "Sage2 can consume projection-strided NHD Q/K/V views without QKV copies.",
      "implementation_strategy": "View interleaved projection output as NHD, normalize/RoPE in place, and call Sage2 NHD.",
      "source_locations": ["src/h3_blackwell_runtime/attention.py", "CURRENT_STATE.md", "PERFORMANCE_ROADMAP.md"], "active_source_location": "src/h3_blackwell_runtime/attention.py", "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-sage-strided-nhd-summary-1344x768-124f-seed440420.json", "exists": true}, {"path": "benchmarks/gb10-sage-strided-nhd-contract-1344x768-124f-seed440420.json", "exists": true}],
      "profiler_artifacts": [{"path": "benchmarks/gb10-component2-nsys-summary-1344x768-124f-seed440420.json", "exists": true}],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "active source in dirty working tree"},
      "metrics": {"result": "accepted; measurements in linked summary"},
      "correctness_evidence": ["Blocks 0/24/49 and two-step/canonical trajectories documented as bit-exact."],
      "decision_rationale": "Removes three large copies and passed production gates.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Distributed-path equivalent"],
      "production_behavior": "Selected for single-GPU Sage2 with H3_SAGE_QKV_LAYOUT=strided_nhd.",
      "source_recovery": "Active in H3SageAttention.forward."
    },
    {
      "id": "qkv-to-sage-fused",
      "name": "QKV-to-Sage fused",
      "family": "qkv-preparation", "status": "performance_rejected",
      "hypothesis": "Fusing direct strided Q/K preparation through Sage2 Q quantization can remove an entry launch and traffic.",
      "implementation_strategy": "Fuse NHD Q/K RMSNorm, RoPE, and Q INT8 quantization while retaining K/V preparation and mainloop.",
      "source_locations": ["src/h3_blackwell_runtime/sage2_entry.py", "tools/validate_sage2_entry_fusion.py", "SAGE2_BLACKWELL_DESIGN.md", "CURRENT_STATE.md", "benchmarks/gb10-sage2-p1-entry-fusion-analysis.json"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-sage2-p1-entry-fusion-analysis.json", "exists": true}],
      "profiler_artifacts": [{"path": "benchmarks/gb10-sage2-p1-blocks-candidate-l2.ncu-rep", "exists": true}],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "runtime branch removed"},
      "metrics": {"result": "exact but below complete-block gate"},
      "correctness_evidence": ["Randomized lengths, real tensors, quantized buffers, attention output, and blocks are documented bit-exact."],
      "decision_rationale": "Entry speedup did not reach the required complete-block improvement.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Two-step and canonical trajectory runs were intentionally skipped"],
      "production_behavior": "Callable wrapper/export and validator source remain, but production integration and dispatch are absent; production preparation is unchanged.",
      "source_recovery": "Callable prepare_qk/attention_nhd source remains in sage2_entry.py and the validator remains in tools/validate_sage2_entry_fusion.py; evidence artifacts are retained, but production integration is absent."
    },
    {
      "id": "qknorm-fused",
      "name": "qknorm fused",
      "family": "qkv-preparation", "status": "research_retained",
      "hypothesis": "Fusing Q/K RMSNorm and split-half RoPE reduces preparation launches while preserving H3 arithmetic.",
      "implementation_strategy": "Use the Comfy Kitchen in-place rms_rope_split_half operator.",
      "source_locations": ["src/h3_blackwell_runtime/attention.py", "CURRENT_STATE.md"], "active_source_location": "src/h3_blackwell_runtime/attention.py", "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-post-optimization-sage2-nhd-components.json", "exists": true}], "profiler_artifacts": [],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "active dependency operator"},
      "metrics": {"role": "current exact preparation reference"},
      "correctness_evidence": ["Used by the accepted exact Sage2 path."],
      "decision_rationale": "Retained as the current exact preparation operation; it is not independently claimed as a production experiment win.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Standalone gate specific to qknorm fusion"],
      "production_behavior": "Called by production attention preparation.", "source_recovery": "Invocation remains in attention.py; implementation belongs to Comfy Kitchen."
    },
    {
      "id": "qknorm-p0-fused-qk-normalization",
      "name": "Fused P0 Q/K normalization",
      "family": "qkv-preparation", "status": "incomplete",
      "hypothesis": "A P0-specific Q/K normalization kernel may reduce preparation overhead before broader entry fusion.",
      "implementation_strategy": "Fuse Q and K normalization at the earliest preparation checkpoint.",
      "source_locations": ["SAGE2_BLACKWELL_DESIGN.md", "tools/profile_attention_components.py"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [],
      "environment": {"gpu": null, "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [],
      "decision_rationale": "The current checkout has no candidate-specific artifact separating this label from later P1 entry fusion.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Candidate source", "Distinct benchmark", "Decision record"],
      "production_behavior": "No distinct P0 normalization dispatch.", "source_recovery": "Check external or deleted research branches."
    },
    {
      "id": "qkv-fused-rope-output-format",
      "name": "Fused RoPE + QKV output formatting",
      "family": "qkv-preparation", "status": "performance_rejected",
      "hypothesis": "Combining RoPE with attention-format stores removes a Q/K pass.",
      "implementation_strategy": "Apply split-half RoPE while writing prepared Q/K into the Sage entry representation.",
      "source_locations": ["src/h3_blackwell_runtime/sage2_entry.py", "tools/validate_sage2_entry_fusion.py", "benchmarks/gb10-sage2-p1-entry-fusion-analysis.json", "SAGE2_BLACKWELL_DESIGN.md", "src/h3_blackwell_runtime/attention.py"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-sage2-p1-entry-fusion-analysis.json", "exists": true}], "profiler_artifacts": [],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "covered by rejected P1 entry candidate"},
      "metrics": {"result": "component was exact within P1 but complete-block gate failed"},
      "correctness_evidence": ["P1 prepared Q/K and complete output parity."],
      "decision_rationale": "As implemented in the fused P1 boundary, formatting fusion did not deliver enough block-level benefit.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Standalone ablation"],
      "production_behavior": "Callable P1 wrapper/export and validator source remain without production integration; production keeps the existing RMSNorm/RoPE preparation on strided views.", "source_recovery": "The P1 prepare_qk wrapper remains in sage2_entry.py and its validator remains in tools/validate_sage2_entry_fusion.py; production dispatch is absent."
    },
    {
      "id": "gemm-persistent",
      "name": "Persistent GEMM",
      "family": "cute-nvfp4", "status": "incomplete",
      "hypothesis": "A persistent scheduler can amortize activation production and launch overhead across output tiles.",
      "implementation_strategy": "Use a persistent work queue with bounded tile reuse rather than one producer per output-N CTA.",
      "source_locations": ["NVFP4_STREAMING_DESIGN.md", "PERFORMANCE_ROADMAP.md", "src/h3_blackwell_runtime/cute_qkv_ring.py"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-cute-qkv-runtime-block-gate-summary.json", "exists": true}], "profiler_artifacts": [],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "proposed successor to rejected ring schedule"},
      "metrics": {}, "correctness_evidence": [],
      "decision_rationale": "The bounded ring was tested, but the distinct launch-fused persistent scheduler remains unimplemented or unconcluded.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Persistent scheduler source", "Block gate", "Trajectory validation"],
      "production_behavior": "No persistent GEMM production dispatch.", "source_recovery": "Design target remains in roadmap; ring prototype is not equivalent."
    },
    {
      "id": "gemm-warp-specialized",
      "name": "Warp-specialized GEMM",
      "family": "cute-nvfp4", "status": "incomplete",
      "hypothesis": "Separating producer and consumer warps can overlap packing with tensor-core GEMM.",
      "implementation_strategy": "Assign DMA/packing and MMA roles to different warps with staged shared-memory handoff.",
      "source_locations": ["NVFP4_STREAMING_DESIGN.md", "benchmarks/gb10-cute-p1-stream-a-summary.json"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-cute-p1-stream-a-summary.json", "exists": true}], "profiler_artifacts": [],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "validator prototype"},
      "metrics": {"result": "fixed-tile contract demonstrated; complete warp-specialized runtime conclusion absent"},
      "correctness_evidence": ["P1 shared-memory producer contract is documented bit-exact."],
      "decision_rationale": "The available evidence covers a fixed tile and a rejected per-CTA schedule, not a complete production candidate.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Canonical runtime implementation", "Integrated benchmark"],
      "production_behavior": "Not selected.", "source_recovery": "Validator evidence remains under benchmarks."
    },
    {
      "id": "runtime-event-overhead",
      "name": "Event-overhead experiment",
      "family": "runtime-overhead", "status": "research_retained",
      "hypothesis": "CUDA event instrumentation may distort small kernel and launch-boundary measurements.",
      "implementation_strategy": "Compare synchronized component probes with uninstrumented end-to-end timing.",
      "source_locations": ["PERFORMANCE_ROADMAP.md", "tools/profile_h3_block.py", "tools/profile_attention_components.py"], "active_source_location": "tools/profile_h3_block.py", "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-fully-fused-fresh-nsight-summary.json", "exists": true}], "profiler_artifacts": [],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "profiling methodology"},
      "metrics": {"conclusion": "component probes are attribution-only; final claims require separate uninstrumented runs"},
      "correctness_evidence": [], "decision_rationale": "Retained as measurement methodology, not a production feature.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Dedicated event-overhead artifact"],
      "production_behavior": "No production behavior change.", "source_recovery": "Profiling tools and methodology remain."
    },
    {
      "id": "attention-full-cuda-graph-capture",
      "name": "Full-attention CUDA graph/capture",
      "family": "cuda-graph", "status": "incomplete",
      "hypothesis": "Capturing full attention can reduce repeated launch overhead for stable shape buckets.",
      "implementation_strategy": "Capture the complete attention preparation/mainloop/output path with preallocated buffers.",
      "source_locations": ["CURRENT_STATE.md", "PERFORMANCE_ROADMAP.md"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [],
      "environment": {"gpu": null, "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [],
      "decision_rationale": "CUDA graph work is explicitly listed as missing and deferred until shape/kernel policies stabilize.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Capture implementation", "Graph safety validation", "Timing"],
      "production_behavior": "No full-attention graph capture.", "source_recovery": "Roadmap-only target."
    },
    {
      "id": "cute-conversion-contract",
      "name": "CuTe NVFP4 conversion contract",
      "family": "cute-nvfp4", "status": "research_retained",
      "hypothesis": "CUTLASS DSL can consume the same logical H3 block-scaled FP4 data with an exact epilogue policy.",
      "implementation_strategy": "Validate E2M1/E4M3 conversion, scale layouts, and alpha-before-BF16 epilogue on real tiles.",
      "source_locations": ["tools/validate_cute_nvfp4_conversion.py", "NVFP4_STREAMING_DESIGN.md", "PERFORMANCE_ROADMAP.md"], "active_source_location": "tools/validate_cute_nvfp4_conversion.py", "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-cute-nvfp4-conversion-contract.json", "exists": true}, {"path": "benchmarks/gb10-cute-p0-h3-summary.json", "exists": true}], "profiler_artifacts": [],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "validator-only"},
      "metrics": {"conclusion": "QKV, attention output, and FC1 exact; FC2 policy differs"},
      "correctness_evidence": ["Real-tile packed data, scales, and outputs are documented exact for selected roles."],
      "decision_rationale": "Retained as foundational interoperability evidence, not runtime dispatch.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Packaged runtime kernel"],
      "production_behavior": "No CuTe production dispatch.", "source_recovery": "Validation tools and artifacts are present."
    },
    {
      "id": "cute-tiny-tile",
      "name": "CuTe tiny-tile attempt",
      "family": "cute-nvfp4", "status": "research_retained",
      "hypothesis": "A fixed tiny real tile can prove the producer-consumer conversion contract before full shapes.",
      "implementation_strategy": "Replace A/SFA TMA input for a fixed 128-row tile and compare every packed byte and output.",
      "source_locations": ["tools/validate_cute_nvfp4_tile_producer.py", "benchmarks/gb10-cute-nvfp4-tile-producer.json", "PERFORMANCE_ROADMAP.md"], "active_source_location": "tools/validate_cute_nvfp4_tile_producer.py", "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-cute-nvfp4-tile-producer.json", "exists": true}, {"path": "benchmarks/gb10-cute-p1-stream-a-summary.json", "exists": true}], "profiler_artifacts": [],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "validator-only"},
      "metrics": {"scope": "fixed 128-row checkpoint"},
      "correctness_evidence": ["Every real 128-K tile and complete selected-role output documented bit-exact."],
      "decision_rationale": "Successful research checkpoint, but deliberately not a full runtime candidate.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Canonical M/padding runtime support"],
      "production_behavior": "Validator only.", "source_recovery": "Tool and benchmark artifacts remain."
    },
    {
      "id": "cute-temporary-output",
      "name": "CuTe temporary-output attempt",
      "family": "cute-nvfp4", "status": "architecture_rejected",
      "hypothesis": "A temporary global output can bridge an owned CuTe GEMM to existing downstream code.",
      "implementation_strategy": "Materialize complete intermediate output and copy/reshape after GEMM.",
      "source_locations": ["CURRENT_STATE.md", "PERFORMANCE_ROADMAP.md", "NVFP4_STREAMING_DESIGN.md"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [],
      "environment": {"gpu": null, "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [],
      "decision_rationale": "Complete temporary materialization fails the stated boundary-removal architecture even if numerically correct.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Historical candidate artifact"],
      "production_behavior": "Not selected.", "source_recovery": "No source found; architecture is superseded by direct/bounded producer designs."
    },
    {
      "id": "cute-failed-builds",
      "name": "CuTe failed-build attempts",
      "family": "cute-toolchain", "status": "temporarily_blocked",
      "hypothesis": "Candidate CuTe kernels can be compiled for SM121 with the available DSL/toolchain.",
      "implementation_strategy": "Build generated CUTLASS DSL kernels for the target architecture.",
      "source_locations": ["CURRENT_STATE.md", "SAGE2_BLACKWELL_DESIGN.md", "benchmarks/gb10-post-optimization-profile-summary.json"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-post-optimization-profile-summary.json", "exists": true}], "profiler_artifacts": [],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "build failures summarized, logs external or absent"},
      "metrics": {"result": "one or more candidate builds rejected"},
      "correctness_evidence": [], "decision_rationale": "Build/toolchain failure prevents a defensible architecture or performance decision.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Per-attempt build logs", "Exact compiler versions", "Candidate source mapping"],
      "production_behavior": "No failed candidate is dispatched.", "source_recovery": "Recover run-specific build directories from external Spark storage."
    },
    {
      "id": "cute-compile-blowups",
      "name": "CuTe compile-blowup attempts",
      "family": "cute-toolchain", "status": "temporarily_blocked",
      "hypothesis": "Larger static schedules can be generated without prohibitive compile-time/resource growth.",
      "implementation_strategy": "Instantiate wider tile/schedule variants in CUTLASS DSL.",
      "source_locations": ["NVFP4_STREAMING_DESIGN.md", "PERFORMANCE_ROADMAP.md"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [],
      "environment": {"gpu": null, "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [],
      "decision_rationale": "Compiler resource blowup is a temporary implementation/toolchain blocker, not proof the algorithm is wrong.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Compiler log", "Exact variant dimensions", "Toolchain versions"],
      "production_behavior": "Not selected.", "source_recovery": "No local build log found; recover external build workspace if available."
    },
    {
      "id": "cute-oom-vram",
      "name": "CuTe OOM/VRAM-failure attempts",
      "family": "cute-resource-limit", "status": "temporarily_blocked",
      "hypothesis": "Full-shape duplicate-model or workspace trials fit available GB10 memory.",
      "implementation_strategy": "Run full projection and model-level variants with large outputs/workspaces.",
      "source_locations": ["PERFORMANCE_ROADMAP.md", "benchmarks/gb10-cute-p2-ring-full-projection-summary.json"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-cute-p2-ring-full-projection-summary.json", "exists": true}], "profiler_artifacts": [],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "duplicate-model harness memory pressure"},
      "metrics": {"result": "FC1 timing unusable; memory pressure caused variance and a container stop"},
      "correctness_evidence": [], "decision_rationale": "Resource failure blocked reliable timing and does not establish a candidate quality or architecture decision.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Peak VRAM measurement", "Container failure log"],
      "production_behavior": "No affected candidate selected.", "source_recovery": "Summary artifact remains; detailed container logs are absent."
    },
    {
      "id": "cute-bounded-ring",
      "name": "CuTe bounded ring-attention/projection attempts",
      "family": "cute-ring", "status": "performance_rejected",
      "hypothesis": "A bounded global packed-tile ring can produce each activation tile once and reuse it across N consumers.",
      "implementation_strategy": "Use caller-owned QDATA/SFA ring slots with allocation-free production and chunked GEMM consumption.",
      "source_locations": ["src/h3_blackwell_runtime/cute_qkv_ring.py", "tools/validate_cute_nvfp4_ring.py", "PERFORMANCE_ROADMAP.md"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-cute-p2-ring-capacity-summary.json", "exists": true}, {"path": "benchmarks/gb10-cute-qkv-runtime-block-gate-summary.json", "exists": true}], "profiler_artifacts": [],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "opt-in disabled prototype"},
      "metrics": {"result": "projection parity passed; alternating complete-block gate regressed"},
      "correctness_evidence": ["Full projections in blocks 0/24/49, including tail chunk, documented bit-exact."],
      "decision_rationale": "Isolated projection gains did not survive chunk launch and scheduler overhead at the block gate.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Trajectory validation intentionally skipped"],
      "production_behavior": "Opt-in dispatchable code remains in cute_qkv_ring.py and nvfp4.py, but H3_CUTE_QKV_RING is disabled by default and absent from production manifests.", "source_recovery": "Dispatchable prototype source remains in src/h3_blackwell_runtime/cute_qkv_ring.py and src/h3_blackwell_runtime/nvfp4.py; validation artifacts are also present."
    },
    {
      "id": "cute-p1-per-output-cta-streaming",
      "name": "CuTe P1 per-output-N-CTA streaming",
      "family": "cute-nvfp4", "status": "performance_rejected",
      "hypothesis": "Each output-N CTA can produce its own A/SFA tile cheaply enough to eliminate global activation packing.",
      "implementation_strategy": "Use the DMA warp to convert BF16 A directly into staged E2M1/SFA shared memory independently in every output-N CTA.",
      "source_locations": ["PERFORMANCE_ROADMAP.md", "benchmarks/gb10-cute-p1-stream-a-timing-summary.json", "tools/validate_cute_nvfp4_real_tiles.py"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-cute-p1-stream-a-timing-summary.json", "exists": true}], "profiler_artifacts": [],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "validator/timing prototype"},
      "metrics": {"result": "exact selected-role outputs, but repeated producer work made the schedule much slower"},
      "correctness_evidence": ["Real QKV, attention-output, and FC1 tile/output identity is documented in the P1 artifacts."],
      "decision_rationale": "Every output-N CTA rereads and repacks A; required N reuse is not viable under the accumulator/register budget.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Candidate kernel source is external to this checkout"],
      "production_behavior": "Rejected schedule is not dispatched.", "source_recovery": "Timing and contract artifacts remain; recover external CuTe kernel checkout for source."
    },
    {
      "id": "cute-p2-full-workspace-ring",
      "name": "CuTe P2 full-workspace ring",
      "family": "cute-ring", "status": "performance_rejected",
      "hypothesis": "A full-activation-sized ring/workspace removes chunk recycling overhead and improves the complete block.",
      "implementation_strategy": "Expand ring capacity to cover the canonical activation while retaining caller-owned packed buffers.",
      "source_locations": ["src/h3_blackwell_runtime/cute_qkv_ring.py", "benchmarks/gb10-cute-qkv-block24-full-workspace-alternating.json", "benchmarks/gb10-cute-qkv-runtime-block-gate-summary.json"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-cute-qkv-block24-full-workspace-alternating.json", "exists": true}, {"path": "benchmarks/gb10-cute-qkv-runtime-block-gate-summary.json", "exists": true}], "profiler_artifacts": [],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "opt-in capacity variant"},
      "metrics": {"result": "exact capacity variant with a documented complete-block regression"},
      "correctness_evidence": ["Alternating block artifact records equality for the candidate path."],
      "decision_rationale": "Removing recycling did not rescue the architecture and substantially worsened the controlled block result.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Trajectory validation intentionally skipped"],
      "production_behavior": "Not selected; bounded-ring dispatch remains disabled.", "source_recovery": "Runtime prototype and capacity artifact remain."
    },
    {
      "id": "cute-fc2-streaming-policy",
      "name": "CuTe FC2 streamed producer attempt",
      "family": "cute-nvfp4", "status": "quality_rejected",
      "hypothesis": "The same streamed CuTe conversion/epilogue policy used for QKV, attention output, and FC1 can preserve FC2 exactly.",
      "implementation_strategy": "Run FC2 through the owned E2M1/E4M3 producer and alpha-before-BF16 epilogue contract.",
      "source_locations": ["PERFORMANCE_ROADMAP.md", "benchmarks/gb10-cute-p0-mlp-fc2-128rows-alpha.json", "benchmarks/gb10-cute-p0-h3-summary.json"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-cute-p0-mlp-fc2-128rows-alpha.json", "exists": true}, {"path": "benchmarks/gb10-cute-p0-h3-summary.json", "exists": true}], "profiler_artifacts": [],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "P0 numerical-contract experiment"},
      "metrics": {"result": "non-exact because FC2 reference reduction policy differs"},
      "correctness_evidence": ["P0 artifacts isolate the FC2 mismatch while other selected roles match."],
      "decision_rationale": "The candidate fails the exact FC2 numerical contract; FC2 streaming is explicitly excluded.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["No accepted alternative FC2 reduction implementation"],
      "production_behavior": "FC2 retains the Comfy/CUBLAS fallback.", "source_recovery": "P0 artifacts and fallback policy are retained in the roadmap."
    },
    {
      "id": "fc2-cublaslt-splitk1-schedule",
      "name": "FC2 cuBLASLt public split-K-1 schedule",
      "family": "nvfp4-library-scheduling",
      "status": "production_accepted",
      "hypothesis": "A documented cuBLASLt schedule can preserve the exact FC2 reduction result while avoiding the production heuristic's traffic and synchronization regression.",
      "implementation_strategy": "Reproduce the exact Comfy Kitchen descriptors in an isolated extension, enumerate checked cuBLASLt algorithms, and compare one selected public split-K-1 schedule against the accepted FC2 path.",
      "source_locations": ["research/fc2_nvfp4_scheduling/README.md", "research/fc2_nvfp4_scheduling/RESULTS.md", "src/h3_blackwell_runtime/csrc/fc2_nvfp4_lt.cpp", "tools/benchmark_fc2_nvfp4_algorithms.py"],
      "active_source_location": "src/h3_blackwell_runtime/csrc/fc2_nvfp4_lt.cpp",
      "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-fc2-nvfp4-library-sweep-20260825.json", "exists": true}, {"path": "benchmarks/gb10-fc2-nvfp4-production-block-gate-20260825.json", "exists": true}, {"path": "benchmarks/gb10-fc2-nvfp4-production-trajectory-2step-20260825.json", "exists": true}, {"path": "benchmarks/gb10-fc2-nvfp4-production-trajectory-12step-20260825.json", "exists": true}, {"path": "benchmarks/gb10-fc2-nvfp4-shape-gate-20260825.json", "exists": true}],
      "profiler_artifacts": [{"path": "benchmarks/gb10-fc2-nvfp4-baseline-20260825.ncu-rep", "exists": true}, {"path": "benchmarks/gb10-fc2-nvfp4-splitk1-20260825.ncu-rep", "exists": true}, {"path": "benchmarks/gb10-fc2-nvfp4-baseline-20260825.csv", "exists": true}, {"path": "benchmarks/gb10-fc2-nvfp4-splitk1-20260825.csv", "exists": true}],
      "environment": {"gpu": "NVIDIA GB10", "cuda": "13", "driver": null, "pytorch": "2.9.1+cu130", "triton": null, "container_image": "sha256:d61c8d29e8b59306d140bfd279b0de31d827ad13bd92447ea72ea0a04182494d", "commit_context": "guarded canonical production integration"},
      "metrics": {"fc2_p50_ms_baseline": 53.618, "fc2_p50_ms_candidate": 15.636, "production_block_improvement_pct": [9.36, 8.16, 7.86], "production_two_step_seconds": {"baseline": 47.30056222799976, "candidate": 43.727866895999796, "improvement_percent": 7.553177306389591}, "production_canonical_12_step_seconds": {"baseline": 286.43065266899976, "candidate": 262.97871506399997, "improvement_percent": 8.187649396624085}, "production_12_step_dispatches": 600, "production_12_step_fallbacks": 0, "l2_hit_rate_pct_baseline": 53.32, "l2_hit_rate_pct_candidate": 91.10, "l2_read_miss_reduction_bytes": 9853094784},
      "correctness_evidence": ["Production blocks 0, 24, and 49 are byte-exact against paired baseline and retained traversal.", "Production two-step and canonical 12-step video and audio latents are byte-exact.", "Nearby M=37711/37725/37742 shape probes differ in two BF16 elements and are excluded by the canonical guard."],
      "decision_rationale": "The shape-guarded library schedule passed production-method block and trajectory gates without requiring a custom kernel. Noncanonical shapes retain the accepted fallback because AlgoCheck alone does not guarantee byte exactness.",
      "reproducer_commands": ["python tools/benchmark_fc2_nvfp4_algorithms.py --mode production-block-gate --candidate research/fc2_nvfp4_scheduling/candidate_splitk1.json --rounds 20 --workspace-bytes 0", "python tools/benchmark_fc2_nvfp4_algorithms.py --mode production-trajectory --candidate research/fc2_nvfp4_scheduling/candidate_splitk1.json --steps 12 --workspace-bytes 0", "python tools/benchmark_fc2_nvfp4_algorithms.py --mode shape-gate --candidate research/fc2_nvfp4_scheduling/candidate_splitk1.json --workspace-bytes 0"],
      "timestamp": "2026-08-25",
      "evidence_missing": ["Resident-service deployment validation deferred until performance work completes", "Trajectory timings are single baseline-then-candidate pairs rather than repeated AB/BA trials", "Shape-gate artifact predates runtime-version fields in build_info", "Portable validation outside GB10/SM121"],
      "production_behavior": "H3_NVFP4_FC2_LT_SPLITK1 selects algorithm 70 only for the validated canonical FC2 descriptor; every other shape, device, library, LoRA, and failure case falls back to Comfy Kitchen.",
      "source_recovery": "The direct cuBLASLt extension, benchmark harness, selected candidate, raw NCU reports, and full gate artifacts are retained in this checkout."
    },
    {
      "id": "layout-direct-temporal-output",
      "name": "Direct-to-temporal output",
      "family": "attention-layout", "status": "incomplete",
      "hypothesis": "Writing QKV or attention results directly to a temporal consumer layout avoids a copy.",
      "implementation_strategy": "Use consumer-native temporal indexing in the producer store path.",
      "source_locations": ["PERFORMANCE_ROADMAP.md", "VORTEX_RESEARCH_AGENDA.md"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [],
      "environment": {"gpu": null, "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [], "decision_rationale": "No source or artifact supports a final decision for this named layout.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Layout contract", "Candidate source", "Benchmark"],
      "production_behavior": "Not selected.", "source_recovery": "Search external research branches/artifact storage."
    },
    {
      "id": "layout-direct-sage-output",
      "name": "Direct-to-Sage output",
      "family": "attention-layout", "status": "incomplete",
      "hypothesis": "A QKV producer can write a Sage-specific output layout directly rather than expose views over interleaved projection output.",
      "implementation_strategy": "Change the producer store contract to emit consumer-native Sage buffers without a separate formatting copy.",
      "source_locations": ["PERFORMANCE_ROADMAP.md", "src/h3_blackwell_runtime/attention.py"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [],
      "environment": {"gpu": null, "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "historical producer/output-layout concept distinct from accepted strided views"},
      "metrics": {},
      "correctness_evidence": [], "decision_rationale": "No distinct producer-store implementation or candidate-specific evidence was found. The accepted no-copy strided-NHD view contract is recorded once under qkv-direct-sageattention-layout and is not evidence that this producer/output-layout concept was accepted.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Distinct producer-store source", "Candidate-specific correctness evidence", "Candidate-specific benchmark"],
      "production_behavior": "No separate direct-to-Sage producer/output-layout acceptance; production uses the strided-NHD view contract recorded by qkv-direct-sageattention-layout.", "source_recovery": "Only the historical concept and the separate accepted view implementation are recoverable in this checkout."
    },
    {
      "id": "layout-direct-q-padding-32",
      "name": "Direct output into q_padding=32 buffers",
      "family": "attention-layout", "status": "architecture_rejected",
      "hypothesis": "Writing directly into padded Q buffers can remove a later padding copy.",
      "implementation_strategy": "Allocate/store producer output in buffers with q_padding=32.",
      "source_locations": ["SAGE2_BLACKWELL_DESIGN.md", "PERFORMANCE_ROADMAP.md"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [],
      "environment": {"gpu": null, "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [],
      "decision_rationale": "Durable padded producer buffers conflict with the accepted strided-view/direct-consumer architecture and no contrary evidence was found.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Historical benchmark", "Candidate source"],
      "production_behavior": "Production does not allocate this direct padded-Q boundary.", "source_recovery": "No candidate source found."
    },
    {
      "id": "layout-dynamic-to-hot-allocation",
      "name": "Dynamic .to(...) hot-path allocation",
      "family": "runtime-allocation", "status": "architecture_rejected",
      "hypothesis": "On-demand device/dtype conversion in the hot path is acceptable for layout preparation.",
      "implementation_strategy": "Call dynamic tensor .to(...) operations during repeated attention execution.",
      "source_locations": ["PERFORMANCE_ROADMAP.md", "src/h3_blackwell_runtime/attention.py"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [],
      "environment": {"gpu": null, "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [],
      "decision_rationale": "Repeated dynamic allocation/conversion is incompatible with allocation-free hot-path and graph-capture goals.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Historical candidate-specific timing"],
      "production_behavior": "Not an accepted optimization strategy.", "source_recovery": "Conceptual rejection; no retained candidate source located."
    },
    {
      "id": "layout-q-coalesced-16b-store",
      "name": "Coalesced-16-byte Q layout store",
      "family": "attention-layout-microvariant", "status": "incomplete",
      "hypothesis": "Coalesced 16-byte Q stores improve layout-writer throughput.",
      "implementation_strategy": "Vectorize Q output stores to aligned 16-byte transactions.",
      "source_locations": ["PERFORMANCE_ROADMAP.md", "src/h3_blackwell_runtime/csrc/nvfp4_scale.cu"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [], "environment": {"gpu": null, "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [], "decision_rationale": "No named artifact or recoverable ablation was found.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Source mapping", "Alignment proof", "Benchmark"],
      "production_behavior": "No distinct dispatch.", "source_recovery": "Inspect external CUDA experiment source if available."
    },
    {
      "id": "layout-qkv-all-16b-stores",
      "name": "All-16-byte Q/K/V stores",
      "family": "attention-layout-microvariant", "status": "incomplete",
      "hypothesis": "Using aligned 16-byte stores for Q, K, and V maximizes layout write bandwidth.",
      "implementation_strategy": "Vectorize every Q/K/V output path to 16-byte transactions.",
      "source_locations": ["PERFORMANCE_ROADMAP.md", "src/h3_blackwell_runtime/csrc/nvfp4_scale.cu"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [], "environment": {"gpu": null, "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [], "decision_rationale": "No defensible conclusion survives in repository evidence.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Candidate source", "Correctness", "Timing"],
      "production_behavior": "No distinct dispatch.", "source_recovery": "External/deleted experiment source required."
    },
    {
      "id": "layout-xor-swizzle-q",
      "name": "XOR-swizzle Q variant",
      "family": "attention-layout-microvariant", "status": "incomplete",
      "hypothesis": "XOR swizzling Q addresses reduces partition/bank conflicts.",
      "implementation_strategy": "Apply an XOR mapping to Q store coordinates.",
      "source_locations": ["SAGE2_BLACKWELL_DESIGN.md", "src/h3_blackwell_runtime/csrc/nvfp4_scale.cu"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [], "environment": {"gpu": null, "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [], "decision_rationale": "No candidate-specific evidence was found.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Swizzle definition", "Profiler counters", "Parity"],
      "production_behavior": "Not selected.", "source_recovery": "Search external experiment source."
    },
    {
      "id": "layout-xor-swizzle-qk",
      "name": "XOR-swizzle Q/K variant",
      "family": "attention-layout-microvariant", "status": "incomplete",
      "hypothesis": "Applying the XOR mapping to both Q and K improves aggregate preparation throughput.",
      "implementation_strategy": "Swizzle Q and K stores while retaining V layout.",
      "source_locations": ["SAGE2_BLACKWELL_DESIGN.md", "src/h3_blackwell_runtime/csrc/nvfp4_scale.cu"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [], "environment": {"gpu": null, "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [], "decision_rationale": "Materially distinct from Q-only, but no retained evidence supports a conclusion.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Candidate source", "Benchmark", "Parity"],
      "production_behavior": "Not selected.", "source_recovery": "Search external experiment source."
    },
    {
      "id": "layout-swizzle-all",
      "name": "Swizzle-all variants",
      "family": "attention-layout-microvariant", "status": "incomplete",
      "hypothesis": "Swizzling all output components avoids systemic store conflicts.",
      "implementation_strategy": "Apply the candidate swizzle across Q/K/V and associated padded regions.",
      "source_locations": ["SAGE2_BLACKWELL_DESIGN.md", "src/h3_blackwell_runtime/csrc/nvfp4_scale.cu"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [], "environment": {"gpu": null, "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [], "decision_rationale": "No retained source or benchmark identifies which swizzle-all variant was viable.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Variant enumeration", "Source", "Counters", "Timing"],
      "production_behavior": "Not selected.", "source_recovery": "Recover external experiment notes/builds."
    },
    {
      "id": "layout-vectorized-padded-q-copy",
      "name": "Vectorized-padded-Q copy",
      "family": "attention-layout-copy", "status": "architecture_rejected",
      "hypothesis": "A wide copy into padded Q reduces the cost of unavoidable preparation.",
      "implementation_strategy": "Vector-copy Q into a separate padded destination.",
      "source_locations": ["CURRENT_STATE.md", "PERFORMANCE_ROADMAP.md"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [], "environment": {"gpu": null, "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "superseded by direct strided views"},
      "metrics": {}, "correctness_evidence": [], "decision_rationale": "Optimizing a complete copy is architecturally superseded by the accepted no-copy Sage layout.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Historical timing"],
      "production_behavior": "No padded-Q copy on accepted direct path.", "source_recovery": "No candidate source retained."
    },
    {
      "id": "layout-q-copy-wide-store",
      "name": "Q-copy wide-store variants",
      "family": "attention-layout-copy", "status": "architecture_rejected",
      "hypothesis": "Wider Q copy stores reduce standalone layout conversion latency.",
      "implementation_strategy": "Sweep vector widths and block geometry for a separate Q copy kernel.",
      "source_locations": ["CURRENT_STATE.md", "PERFORMANCE_ROADMAP.md"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [], "environment": {"gpu": null, "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "superseded by no-copy path"},
      "metrics": {}, "correctness_evidence": [], "decision_rationale": "A faster standalone copy still fails the boundary-elimination objective.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Variant timings", "Source"],
      "production_behavior": "Not selected.", "source_recovery": "Historical microbenchmarks were not found."
    },
    {
      "id": "layout-q-copy-block128x2",
      "name": "Q-copy wide-store block128x2",
      "family": "attention-layout-copy", "status": "architecture_rejected",
      "hypothesis": "The block128x2 geometry improves wide-store Q copy occupancy/coalescing.",
      "implementation_strategy": "Use a 128-by-2 block mapping for the Q copy writer.",
      "source_locations": ["CURRENT_STATE.md", "PERFORMANCE_ROADMAP.md"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [], "environment": {"gpu": null, "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "superseded microvariant"},
      "metrics": {}, "correctness_evidence": [], "decision_rationale": "Materially distinct geometry, but still a complete-copy architecture superseded by strided views.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Exact geometry source", "Timing", "Parity"],
      "production_behavior": "Not selected.", "source_recovery": "No source found in current tree/history search."
    },
    {
      "id": "layout-grid-kv-writers",
      "name": "Grid-style KV writers",
      "family": "attention-layout-microvariant", "status": "incomplete",
      "hypothesis": "A 2D grid writer improves K/V store coalescing and load balance.",
      "implementation_strategy": "Map K/V rows and heads across a grid instead of a linear writer.",
      "source_locations": ["SAGE2_BLACKWELL_DESIGN.md", "tools/profile_attention_components.py"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [], "environment": {"gpu": null, "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [], "decision_rationale": "No retained artifact supports a decision.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Grid geometry", "Candidate source", "Benchmark"],
      "production_behavior": "No distinct dispatch.", "source_recovery": "External experiment source required."
    },
    {
      "id": "layout-vectorized-kv-staging",
      "name": "Vectorized K/V staging",
      "family": "attention-layout-microvariant", "status": "incomplete",
      "hypothesis": "Vectorized K/V staging reduces preparation memory transactions.",
      "implementation_strategy": "Use aligned vector loads/stores during K/V preparation.",
      "source_locations": ["SAGE2_BLACKWELL_DESIGN.md", "benchmarks/gb10-sage2-block24-mainloop-analysis.json"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [], "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [], "decision_rationale": "The profiler identifies V staging behavior, but not a conclusive vectorized K/V candidate.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Candidate source", "Ablation timing", "Counters"],
      "production_behavior": "Production Sage2 staging unchanged.", "source_recovery": "Use mainloop analysis as context; candidate source absent."
    },
    {
      "id": "sage2-online-pv",
      "name": "Online-PV Sage2",
      "family": "sage2-mainloop", "status": "incomplete",
      "hypothesis": "An online PV schedule can overlap or reduce score materialization while preserving Sage2 accumulation semantics.",
      "implementation_strategy": "Integrate PV consumption into online-softmax progress without changing per-warp numerical order.",
      "source_locations": ["SAGE2_BLACKWELL_DESIGN.md", "benchmarks/gb10-sage2-block24-mainloop-analysis.json"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [], "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [], "decision_rationale": "No distinct online-PV implementation or final gate was found; P3 temporal pairing is recorded separately.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Candidate source", "Parity", "Mainloop timing"],
      "production_behavior": "Production uses the public Sage2 mainloop.", "source_recovery": "Check isolated SageAttention worktrees."
    },
    {
      "id": "sage-scheduler-v1",
      "name": "Sage scheduler/register-pressure v1",
      "family": "sage2-scheduler", "status": "incomplete",
      "hypothesis": "A first scheduler rewrite can reduce dependency stalls without increasing spills.",
      "implementation_strategy": "Reorder QK/PV/softmax work and narrow live ranges.",
      "source_locations": ["SAGE2_BLACKWELL_DESIGN.md", "benchmarks/gb10-sage2-p0-register-scheduler-analysis.json"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [], "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [], "decision_rationale": "The aggregate P0 analysis survives, but it does not map v1 to an exact retained variant.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["v1 source mapping", "v1-specific timing", "v1 parity"],
      "production_behavior": "Not selected.", "source_recovery": "Recover original SageAttention variant worktree/build logs."
    },
    {
      "id": "sage-scheduler-v2",
      "name": "Sage scheduler/register-pressure v2",
      "family": "sage2-scheduler", "status": "incomplete",
      "hypothesis": "A second scheduling/register-pressure balance improves eligible-warp availability.",
      "implementation_strategy": "Revise live ranges and instruction ordering relative to v1.",
      "source_locations": ["SAGE2_BLACKWELL_DESIGN.md", "benchmarks/gb10-sage2-p0-register-scheduler-analysis.json"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [], "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [], "decision_rationale": "No variant-specific artifact permits a conclusion independent of aggregate P0 rejection.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["v2 source mapping", "v2 timing", "v2 counters"],
      "production_behavior": "Not selected.", "source_recovery": "Recover isolated variant checkout if available."
    },
    {
      "id": "sage-scheduler-v3",
      "name": "Sage scheduler/register-pressure v3",
      "family": "sage2-scheduler", "status": "incomplete",
      "hypothesis": "A third scheduler variant can cross the mainloop gate after earlier register-pressure tradeoffs.",
      "implementation_strategy": "Apply the final recorded scheduling/register allocation variant.",
      "source_locations": ["SAGE2_BLACKWELL_DESIGN.md", "benchmarks/gb10-sage2-p0-register-scheduler-analysis.json"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [], "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [], "decision_rationale": "The name is required for archival completeness, but exact evidence mapping is missing.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["v3 source", "v3 benchmark", "v3 decision"],
      "production_behavior": "Not selected.", "source_recovery": "Recover external experiment notes/builds."
    },
    {
      "id": "sage-p0-register-caps",
      "name": "Sage P0 register-cap variants",
      "family": "sage2-scheduler", "status": "performance_rejected",
      "hypothesis": "Capping registers can increase CTA residency enough to hide dependency stalls.",
      "implementation_strategy": "Sweep register caps through the occupancy cliff and measure spills/latency.",
      "source_locations": ["SAGE2_BLACKWELL_DESIGN.md", "benchmarks/gb10-sage2-p0-register-scheduler-analysis.json"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-sage2-p0-register-cap-latency-interleaved.json", "exists": true}],
      "profiler_artifacts": [{"path": "benchmarks/gb10-sage2-p0-cap168-counters.ncu-rep", "exists": true}],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "isolated SageAttention P0"},
      "metrics": {"result": "occupancy gain at the cliff was overwhelmed by spills"},
      "correctness_evidence": ["P0 variants documented byte-exact where timed."], "decision_rationale": "No cap crossed the mainloop gate; the residency-changing cap caused severe spill traffic.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Candidate source in this checkout"],
      "production_behavior": "Production compile policy unchanged.", "source_recovery": "NCU and latency artifacts retained."
    },
    {
      "id": "sage-p0-narrow-scopes",
      "name": "Sage P0 narrowed scopes",
      "family": "sage2-scheduler", "status": "performance_rejected",
      "hypothesis": "Narrowing C++ variable scopes reduces live ranges, spills, and scheduler stalls.",
      "implementation_strategy": "Refactor exact mainloop scopes without changing arithmetic.",
      "source_locations": ["SAGE2_BLACKWELL_DESIGN.md", "benchmarks/gb10-sage2-p0-register-scheduler-analysis.json"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-sage2-p0-scopes-latency-interleaved.json", "exists": true}],
      "profiler_artifacts": [{"path": "benchmarks/gb10-sage2-p0-scopes-counters.ncu-rep", "exists": true}],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "isolated P0 variant"},
      "metrics": {"result": "spill reduction without material latency gain"},
      "correctness_evidence": ["Documented byte-exact."], "decision_rationale": "Reduced spills did not improve mainloop enough and scheduler eligibility worsened.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Candidate source in current checkout"],
      "production_behavior": "Not integrated.", "source_recovery": "Artifacts remain; source lived in an isolated checkout."
    },
    {
      "id": "sage-p0-inplace-scores",
      "name": "Sage P0 in-place scores",
      "family": "sage2-scheduler", "status": "performance_rejected",
      "hypothesis": "Reusing score storage in place lowers register pressure and data movement.",
      "implementation_strategy": "Alias/reuse exact score fragments while preserving operation order.",
      "source_locations": ["SAGE2_BLACKWELL_DESIGN.md", "benchmarks/gb10-sage2-p0-inplace-scores-latency-interleaved.json"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-sage2-p0-inplace-scores-latency-interleaved.json", "exists": true}], "profiler_artifacts": [],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "isolated P0 variant"},
      "metrics": {"result": "neutral or slower"}, "correctness_evidence": ["Documented byte-exact."],
      "decision_rationale": "Did not cross the required mainloop improvement gate.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Candidate source"],
      "production_behavior": "Not integrated.", "source_recovery": "Latency artifact retained."
    },
    {
      "id": "sage-p0-early-k-prefetch",
      "name": "Sage P0 early-K prefetch",
      "family": "sage2-scheduler", "status": "performance_rejected",
      "hypothesis": "Prefetching K earlier overlaps staging with compute.",
      "implementation_strategy": "Move exact K staging ahead in the mainloop schedule.",
      "source_locations": ["SAGE2_BLACKWELL_DESIGN.md", "benchmarks/gb10-sage2-p0-early-k-latency-interleaved.json"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-sage2-p0-early-k-latency-interleaved.json", "exists": true}], "profiler_artifacts": [],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "isolated P0 variant"},
      "metrics": {"result": "slower or below gate"}, "correctness_evidence": ["Documented byte-exact."],
      "decision_rationale": "No material latency gain was measured.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Candidate source"],
      "production_behavior": "Not integrated.", "source_recovery": "Latency artifact retained."
    },
    {
      "id": "sage-p0-softmax-interleave",
      "name": "Sage P0 independent softmax-chain interleaving",
      "family": "sage2-scheduler", "status": "performance_rejected",
      "hypothesis": "Interleaving independent softmax chains hides scalar dependency latency.",
      "implementation_strategy": "Reorder independent online-softmax chains without changing each chain's arithmetic.",
      "source_locations": ["SAGE2_BLACKWELL_DESIGN.md", "benchmarks/gb10-sage2-p0-interleave-softmax-latency.json"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-sage2-p0-interleave-softmax-latency.json", "exists": true}], "profiler_artifacts": [],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "isolated P0 variant"},
      "metrics": {"result": "neutral or slower"}, "correctness_evidence": ["Documented byte-exact."],
      "decision_rationale": "Did not alleviate the measured scheduler bottleneck enough to pass.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Candidate source"],
      "production_behavior": "Not integrated.", "source_recovery": "Latency artifact retained."
    },
    {
      "id": "sage-p0-padded-v-smem",
      "name": "Sage P0 padded-V shared-memory layout",
      "family": "sage2-scheduler", "status": "performance_rejected",
      "hypothesis": "Padding V shared-memory stride removes excessive wavefronts and improves latency.",
      "implementation_strategy": "Increase V shared stride to a 128-byte padded layout.",
      "source_locations": ["SAGE2_BLACKWELL_DESIGN.md", "benchmarks/gb10-sage2-p0-pad-v-smem-latency.json"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-sage2-p0-pad-v-smem-latency.json", "exists": true}],
      "profiler_artifacts": [{"path": "benchmarks/gb10-sage2-p0-pad-v-smem-wavefronts.ncu-rep", "exists": true}],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "isolated P0 variant"},
      "metrics": {"result": "excess wavefront count unchanged; no material latency gain"},
      "correctness_evidence": ["Exact mainloop contract retained."], "decision_rationale": "The targeted counter was unaffected and latency gate failed.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Candidate source"],
      "production_behavior": "Original Sage2 shared layout retained.", "source_recovery": "Latency and NCU artifacts retained."
    },
    {
      "id": "backend-flashinfer",
      "name": "FlashInfer",
      "family": "attention-backend", "status": "incomplete",
      "hypothesis": "FlashInfer may provide a faster compatible dense attention backend for H3 shapes.",
      "implementation_strategy": "Adapt H3 Q/K/V layout and non-causal head-dimension-128 contract to FlashInfer.",
      "source_locations": ["src/h3_blackwell_runtime/attention.py", "tools/compare_attention_backends.py"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [], "environment": {"gpu": null, "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [], "decision_rationale": "FlashInfer is not an available backend in current source and no artifact was found.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Adapter", "Version", "Benchmark", "Quality evidence"],
      "production_behavior": "Unavailable.", "source_recovery": "No source/history match found in this checkout."
    },
    {
      "id": "backend-flashattention-3",
      "name": "FlashAttention 3",
      "family": "attention-backend", "status": "incomplete",
      "hypothesis": "FlashAttention 3 may outperform generic SDPA while retaining dense attention quality.",
      "implementation_strategy": "Integrate the FA3 API for H3 BF16 Q/K/V and compare against Sage2.",
      "source_locations": ["src/h3_blackwell_runtime/attention.py", "FLASH4.md", "tools/compare_attention_backends.py"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [], "environment": {"gpu": null, "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [], "decision_rationale": "Current source integrates FlashAttention 4, not FlashAttention 3; no FA3 evidence was found.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["FA3 source", "Package version", "Benchmark", "Quality evidence"],
      "production_behavior": "Unavailable; FlashAttention 4 is a separate optional backend.", "source_recovery": "No FA3 source/history match found."
    },
    {
      "id": "backend-spargeattention",
      "name": "SpargeAttention",
      "family": "attention-backend", "status": "incomplete",
      "hypothesis": "SpargeAttention sparsity may reduce long-sequence H3 attention cost with acceptable quality.",
      "implementation_strategy": "Integrate backend routing and perform full latent/audio/visual quality gates.",
      "source_locations": ["PERFORMANCE_ROADMAP.md", "tools/compare_attention_backends.py"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [], "profiler_artifacts": [], "environment": {"gpu": null, "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": null},
      "metrics": {}, "correctness_evidence": [], "decision_rationale": "No adapter, benchmark, or quality run was found.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Implementation", "Version", "Performance", "Media quality gate"],
      "production_behavior": "Unavailable.", "source_recovery": "No source/history match found in this checkout."
    },
    {
      "id": "backend-sol",
      "name": "Sol",
      "family": "attention-backend", "status": "quality_rejected",
      "hypothesis": "Sparse Sol routing can materially accelerate long H3 attention while preserving acceptable output quality.",
      "implementation_strategy": "Route sparse non-causal attention with configurable threshold and optional residual INT8 Q/K.",
      "source_locations": ["src/h3_blackwell_runtime/attention.py", "PERFORMANCE_ROADMAP.md", "VORTEX_RESEARCH_AGENDA.md"], "active_source_location": "src/h3_blackwell_runtime/attention.py", "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-37810-token-optimization-profile-seed440420.json", "exists": true}], "profiler_artifacts": [],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "experimental backend available"},
      "metrics": {"conclusion": "no measured setting was both faster and numerically Sage-like"},
      "correctness_evidence": ["Relative-output comparisons against Sage2 are summarized in PERFORMANCE_ROADMAP.md."],
      "decision_rationale": "Performance potential does not satisfy parity/quality requirements; complete media gates are absent.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Full latent/audio/lip-sync/visual quality sweep"],
      "production_behavior": "Experimental opt-in only with fallback; Sage2 remains default.", "source_recovery": "Adapter remains active as an experimental backend."
    },
    {
      "id": "sage2-p1-entry-fusion-final",
      "name": "Finalized Sage2 P1 entry fusion",
      "family": "sage2-finalized", "status": "performance_rejected",
      "hypothesis": "One exact entry kernel can fuse strided Q/K RMSNorm, RoPE, and Q quantization for a block-level gain.",
      "implementation_strategy": "Leave K/V preparation and SM89 mainloop unchanged while replacing the entry sequence.",
      "source_locations": ["src/h3_blackwell_runtime/sage2_entry.py", "tools/validate_sage2_entry_fusion.py", "SAGE2_BLACKWELL_DESIGN.md", "CURRENT_STATE.md", "benchmarks/gb10-sage2-p1-entry-fusion-analysis.json"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-sage2-p1-entry-fusion-analysis.json", "exists": true}, {"path": "benchmarks/gb10-sage2-p1-real-parity.json", "exists": true}],
      "profiler_artifacts": [{"path": "benchmarks/gb10-sage2-p1-blocks-candidate-l2.ncu-rep", "exists": true}],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "candidate runtime branch removed"},
      "metrics": {"result": "entry improved, but no complete block reached required gate"},
      "correctness_evidence": ["Randomized and real prepared tensors, quantization, Sage output, and complete blocks documented bit-exact."],
      "decision_rationale": "Failed the required complete-block median improvement despite exactness.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Trajectory validation intentionally skipped"],
      "production_behavior": "Callable wrapper/export and validator source remain, but production integration and dispatch are absent; the public Sage2 entry remains.", "source_recovery": "Callable prepare_qk/attention_nhd source remains in sage2_entry.py and the validator remains in tools/validate_sage2_entry_fusion.py; analysis and parity artifacts are retained, but production integration is absent."
    },
    {
      "id": "sage2-p2-direct-v-preparation-final",
      "name": "Finalized Sage2 P2 direct V preparation",
      "family": "sage2-finalized", "status": "performance_rejected",
      "hypothesis": "Direct NHD BF16-to-padded/permuted FP8 V preparation can remove the full BF16 transpose and save enough block time.",
      "implementation_strategy": "Use a three-stage maxima, scale-finalization, and direct E4M3 writer.",
      "source_locations": ["src/h3_blackwell_runtime/sage2_entry.py", "tools/validate_sage2_vprep.py", "SAGE2_BLACKWELL_DESIGN.md", "CURRENT_STATE.md", "benchmarks/gb10-sage2-p2-vprep-analysis.json"], "active_source_location": null, "commit_hash": null,
      "benchmark_artifacts": [{"path": "benchmarks/gb10-sage2-p2-vprep-analysis.json", "exists": true}, {"path": "benchmarks/gb10-sage2-vprep-randomized.json", "exists": true}, {"path": "benchmarks/gb10-sage2-vprep-canonical-repeat.json", "exists": true}], "profiler_artifacts": [],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "validator-only candidate"},
      "metrics": {"result": "exact and faster in isolation, but missed absolute go threshold"},
      "correctness_evidence": ["FP8 bytes and FP32 scales documented exact across boundary lengths."],
      "decision_rationale": "Absolute saving could not reliably clear the complete-block gate, so integration was skipped.",
      "reproducer_commands": [], "timestamp": null, "evidence_missing": ["Complete-block and trajectory validation intentionally skipped"],
      "production_behavior": "Callable wrapper/export and validator source remain, but production integration and dispatch are absent; production V preparation is unchanged.", "source_recovery": "Callable prepare_v source remains in sage2_entry.py and the validator remains in tools/validate_sage2_vprep.py; evidence artifacts are retained, but production integration is absent."
    },
    {
      "id": "sage2-p3-temporal-pair-final",
      "name": "Finalized Sage2 P3 temporal pair",
      "family": "sage2-finalized", "status": "performance_rejected",
      "hypothesis": "Two warp pairs alternating QK/online-softmax and prior-tile PV can overlap INT and FP tensor work while preserving exact per-warp state.",
      "implementation_strategy": "Keep private warp scores/state/accumulators and alternate pair phases with exact barriers.",
      "source_locations": ["research/sage2_temporal_pair/patches/0001-sm89-temporal-two-pair-qk-pv.patch", "research/sage2_temporal_pair/README.md", "benchmarks/gb10-sage2-p3-temporal-pair-analysis.json"],
      "active_source_location": null,
      "commit_hash": "d1a57a546c3d395b1ffcbeecc66d81db76f3b4b5",
      "benchmark_artifacts": [{"path": "benchmarks/gb10-sage2-p3-temporal-pair-analysis.json", "exists": true}, {"path": "benchmarks/gb10-sage2-temporal-pair-short-20260825-p0.json", "exists": true}, {"path": "benchmarks/gb10-sage2-temporal-pair-timing-20260825-p0.json", "exists": true}],
      "profiler_artifacts": [],
      "environment": {"gpu": "NVIDIA GB10", "cuda": null, "driver": null, "pytorch": null, "triton": null, "container_image": null, "commit_context": "isolated SageAttention source commit recorded by experiment"},
      "metrics": {"result": "sanitizer-clean and exact, but effectively neutral and above gate"},
      "correctness_evidence": ["Thirteen short shapes and real block-24 output SHA documented exact; sanitizer errors zero."],
      "decision_rationale": "Temporal separation did not produce material overlap or meet the absolute mainloop gate.",
      "reproducer_commands": ["bash research/sage2_temporal_pair/run_spark_experiment.sh"],
      "timestamp": null,
      "evidence_missing": ["NCU, block integration, and trajectory validation intentionally skipped"],
      "production_behavior": "No production integration; public Sage2 mainloop remains.",
      "source_recovery": "Patch, runner, validator, README, and result artifacts are retained under research/ and benchmarks/."
    }
  ]
}