h3-blackwell-runtime/research/EXPERIMENT_REGISTRY.md

# Experiment Registry

Generated from the repository working tree, documentation, benchmark artifacts,
and git history available on 2026-08-26. The machine-readable registry is
`research/experiment_registry.json`; it is authoritative for record fields and
artifact existence flags.

Rejected-from-production does not mean deleted. Rejected and blocked work is
retained when it remains useful as evidence, a reference, a reproducer, or a
warning against repeating an unproductive design.

## Status Taxonomy

| Status | Meaning |
| --- | --- |
| `production_accepted` | Currently selected production path for its documented eligible environment. |
| `research_retained` | Useful neutral/reference result, without a production-selection claim. |
| `architecture_rejected` | The design conflicts with the required dataflow or merely optimizes a boundary that should be removed. |
| `performance_rejected` | Correct or sufficiently validated candidate that failed its performance gate. |
| `quality_rejected` | Candidate failed parity or quality requirements, regardless of speed. |
| `temporarily_blocked` | Toolchain, build, compiler-resource, host, OOM, or VRAM issue prevents a defensible conclusion. |
| `incomplete` | Evidence is insufficient for any stronger conclusion. |

Only these seven exact values are allowed. Unknown versions, measurements,
commands, hashes, and timestamps remain `null`, empty, or explicitly listed as
missing rather than inferred.

## Complete Registry

| ID | Experiment | Status | Decision | Primary source/evidence | Evidence gaps |
| --- | --- | --- | --- | --- | --- |
| `elem-fused-residual-rmsnorm` | Fused residual + RMSNorm | `production_accepted` | Exact fused elementwise path passed deployment gates. | `src/h3_blackwell_runtime/h3_fusion.py`; `benchmarks/gb10-h3-fused-elementwise-1344x768-124f-seed440420.json` | Portable non-GB10 validation |
| `nvfp4-quantize-alone` | NVFP4 quantize alone | `research_retained` | Exact packing reference; no standalone production win. | `src/h3_blackwell_runtime/nvfp4_quant.py`; pack-parity benchmark | Standalone canonical deployment gate |
| `nvfp4-quantize-gemm` | NVFP4 quantize + GEMM | `research_retained` | Exact seam/reference retained. | `src/h3_blackwell_runtime/nvfp4.py`; native-fixed benchmark | Accepted isolated performance gate |
| `nvfp4-prequant-swiglu-gemm` | Fused prequantized SiLU x up + quantize + GEMM | `production_accepted` | Exact producer fusion passed block, trajectory, and smoke gates. | `NVFP4_SWIGLU_FUSION_DESIGN.md`; SwiGLU summary | Other architectures |
| `nvfp4-prequant-modulate-gemm` | Fused modulation + quantize + GEMM | `production_accepted` | Exact producer fusion improved blocks 0/24/49 by 0.28-0.79% and the warmed canonical run by 0.56%. | `nvfp4_quant.py`; `NVFP4_MODULATE_FUSION_DESIGN.md`; modulation summary/deployment smoke | Ineligible cases retain materialized fallback |
| `lora-fused-down-activation` | Fused LoRA down + activation | `incomplete` | No isolated conclusion found. | `src/h3_blackwell_runtime/lora.py`; `TURBO.md` | Source, benchmark, parity |
| `lora-fused-up-residual` | Fused LoRA up + residual | `incomplete` | No isolated conclusion found. | `src/h3_blackwell_runtime/lora.py`; `TURBO.md` | Source, benchmark, residual-order parity |
| `lora-fused-full-down-up` | Fused full LoRA down/up | `incomplete` | No complete candidate found. | `src/h3_blackwell_runtime/lora.py`; `TURBO.md` | Implementation, timing, trajectory parity |
| `lora-direct-materialization-gemm-prepass` | Direct LoRA materialization into GEMM prepass memory | `performance_rejected` | Exact Turbo-4 result was slower. | LoRA producer-fusion benchmark; roadmap | Isolated write-cost profile |
| `fc2-fused-activation-residual` | Fused FC2 activation/residual | `incomplete` | Numerical contract exists; integrated candidate does not. | `PERFORMANCE_ROADMAP.md` | Source, randomized parity, block/trajectory gates |
| `qkv-fused-projection-layout` | Fused QKV projection + layout | `architecture_rejected` | Post-GEMM formatting copy superseded by direct views. | Current state; strided-NHD summary | Historical candidate benchmark |
| `qkv-fused-projection-rmsnorm` | Fused QKV + RMSNorm | `incomplete` | No projection-integrated gate found. | Roadmap; `attention.py` | Source, parity, timing |
| `qkv-direct-sageattention-layout` | Direct QKV output into SageAttention layout | `production_accepted` | Strided NHD views remove copies and pass exact gates. | `attention.py`; strided-NHD summary/contract | Distributed equivalent |
| `qkv-to-sage-fused` | QKV-to-Sage fused | `performance_rejected` | Exact P1 entry candidate missed complete-block gate; callable source remains without production integration. | `sage2_entry.py`; entry validator; P1 analysis and NCU report | Trajectory runs intentionally skipped |
| `qknorm-fused` | qknorm fused | `research_retained` | Current exact preparation reference, not an independent win claim. | `attention.py`; NHD component benchmark | Standalone ablation |
| `qknorm-p0-fused-qk-normalization` | Fused P0 Q/K normalization | `incomplete` | Cannot separate this label from later P1 evidence. | Sage2 design | Candidate, distinct benchmark, decision |
| `qkv-fused-rope-output-format` | Fused RoPE + QKV output formatting | `performance_rejected` | Exact within rejected P1 boundary; callable P1 source remains without production dispatch. | `sage2_entry.py`; entry validator; P1 analysis | Standalone ablation |
| `gemm-persistent` | Persistent GEMM | `incomplete` | Distinct launch-fused scheduler remains untested. | Streaming design; roadmap | Scheduler source, block gate, trajectory |
| `gemm-warp-specialized` | Warp-specialized GEMM | `incomplete` | Fixed-tile contract exists; full runtime conclusion absent. | CuTe P1 stream summary | Canonical implementation and benchmark |
| `runtime-event-overhead` | Event-overhead experiment | `research_retained` | Retained profiling methodology. | Profiling tools; roadmap | Dedicated artifact |
| `attention-full-cuda-graph-capture` | Full-attention CUDA graph/capture | `incomplete` | Explicitly deferred/missing. | Current state; roadmap | Implementation, graph safety, timing |
| `cute-conversion-contract` | CuTe NVFP4 conversion contract | `research_retained` | Exact selected-role interoperability established. | Conversion tool/artifacts | Packaged runtime kernel |
| `cute-tiny-tile` | CuTe tiny-tile attempt | `research_retained` | Successful fixed 128-row checkpoint only. | Tile-producer tool/artifact | Canonical M/padding runtime support |
| `cute-temporary-output` | CuTe temporary-output attempt | `architecture_rejected` | Complete temporary materialization violates boundary-removal goal. | Streaming design; roadmap | Historical candidate artifact |
| `cute-failed-builds` | CuTe failed-build attempts | `temporarily_blocked` | Build failures prevent algorithmic conclusion. | Post-optimization summary | Build logs, versions, source mapping |
| `cute-compile-blowups` | CuTe compile-blowup attempts | `temporarily_blocked` | Compiler resource growth is a toolchain blocker. | Streaming design; roadmap | Compiler log, dimensions, versions |
| `cute-oom-vram` | CuTe OOM/VRAM-failure attempts | `temporarily_blocked` | Memory pressure invalidated reliable timing. | Ring full-projection summary | Peak VRAM and failure log |
| `cute-bounded-ring` | CuTe bounded ring-attention/projection attempts | `performance_rejected` | Exact projection path regressed at complete-block gate; opt-in dispatchable source remains disabled by default and absent from production manifests. | `cute_qkv_ring.py`; `nvfp4.py`; capacity/block-gate summaries | Trajectory intentionally skipped |
| `cute-p1-per-output-cta-streaming` | CuTe P1 per-output-N-CTA streaming | `performance_rejected` | Exact, but repeated A production made the schedule much slower. | P1 timing summary; real-tile validator | Candidate kernel source |
| `cute-p2-full-workspace-ring` | CuTe P2 full-workspace ring | `performance_rejected` | Exact full-capacity variant regressed at block gate. | Full-workspace alternating artifact; block-gate summary | Trajectory intentionally skipped |
| `cute-fc2-streaming-policy` | CuTe FC2 streamed producer attempt | `quality_rejected` | Different FC2 reduction policy prevented exactness. | FC2 P0 artifacts; roadmap | Exact alternative reduction implementation |
| `layout-direct-temporal-output` | Direct-to-temporal output | `incomplete` | No retained source or result. | Roadmap; research agenda | Contract, source, benchmark |
| `layout-direct-sage-output` | Direct-to-Sage output | `incomplete` | Historical producer/output-layout concept has no distinct implementation or evidence; it is not the accepted no-copy view contract. | Roadmap; `attention.py` context | Producer-store source, correctness, benchmark |
| `layout-direct-q-padding-32` | Direct output into `q_padding=32` buffers | `architecture_rejected` | Durable padded buffer conflicts with accepted direct-view path. | Sage2 design; roadmap | Historical source/benchmark |
| `layout-dynamic-to-hot-allocation` | Dynamic `.to(...)` hot-path allocation | `architecture_rejected` | Conflicts with allocation-free and capture goals. | Roadmap; `attention.py` | Historical timing |
| `layout-q-coalesced-16b-store` | Coalesced-16-byte Q layout store | `incomplete` | No named ablation found. | CUDA source context | Mapping, alignment proof, benchmark |
| `layout-qkv-all-16b-stores` | All-16-byte Q/K/V stores | `incomplete` | No defensible conclusion survives. | CUDA source context | Source, correctness, timing |
| `layout-xor-swizzle-q` | XOR-swizzle Q variant | `incomplete` | No candidate-specific evidence. | Sage2 design context | Definition, counters, parity |
| `layout-xor-swizzle-qk` | XOR-swizzle Q/K variant | `incomplete` | Distinct variant, evidence absent. | Sage2 design context | Source, benchmark, parity |
| `layout-swizzle-all` | Swizzle-all variants | `incomplete` | Variant enumeration and results absent. | Sage2 design context | Enumeration, source, counters, timing |
| `layout-vectorized-padded-q-copy` | Vectorized-padded-Q copy | `architecture_rejected` | Superseded by no-copy Sage path. | Current state; roadmap | Historical timing |
| `layout-q-copy-wide-store` | Q-copy wide-store variants | `architecture_rejected` | Faster copy still fails boundary-elimination goal. | Current state; roadmap | Variant timings and source |
| `layout-q-copy-block128x2` | Q-copy wide-store `block128x2` | `architecture_rejected` | Distinct geometry, same superseded copy architecture. | Current state; roadmap | Geometry source, timing, parity |
| `layout-grid-kv-writers` | Grid-style KV writers | `incomplete` | No retained artifact. | Sage2 design; profiler tool | Geometry, source, benchmark |
| `layout-vectorized-kv-staging` | Vectorized K/V staging | `incomplete` | Mainloop context exists, candidate does not. | Mainloop analysis | Source, ablation, counters |
| `sage2-online-pv` | Online-PV Sage2 | `incomplete` | No distinct implementation; P3 is separate. | Sage2 design; mainloop analysis | Source, parity, timing |
| `sage-scheduler-v1` | Sage scheduler/register-pressure v1 | `incomplete` | Aggregate P0 evidence cannot map v1. | P0 analysis | v1 source, timing, parity |
| `sage-scheduler-v2` | Sage scheduler/register-pressure v2 | `incomplete` | Aggregate P0 evidence cannot map v2. | P0 analysis | v2 source, timing, counters |
| `sage-scheduler-v3` | Sage scheduler/register-pressure v3 | `incomplete` | Required archival name lacks exact evidence mapping. | P0 analysis | v3 source, benchmark, decision |
| `sage-p0-register-caps` | Sage P0 register-cap variants | `performance_rejected` | Occupancy gain overwhelmed by spills. | Cap latency and NCU artifacts | Candidate source |
| `sage-p0-narrow-scopes` | Sage P0 narrowed scopes | `performance_rejected` | Spill reduction produced no material gain. | Scope latency and NCU artifacts | Candidate source |
| `sage-p0-inplace-scores` | Sage P0 in-place scores | `performance_rejected` | Exact but neutral/slower. | In-place latency artifact | Candidate source |
| `sage-p0-early-k-prefetch` | Sage P0 early-K prefetch | `performance_rejected` | Exact but no material gain. | Early-K latency artifact | Candidate source |
| `sage-p0-softmax-interleave` | Sage P0 independent softmax-chain interleaving | `performance_rejected` | Exact but neutral/slower. | Interleave latency artifact | Candidate source |
| `sage-p0-padded-v-smem` | Sage P0 padded-V shared-memory layout | `performance_rejected` | Targeted wavefront count unchanged. | Padded-V timing and NCU artifacts | Candidate source |
| `backend-flashinfer` | FlashInfer | `incomplete` | No adapter or artifact found. | `attention.py`; backend comparison tool | Adapter, version, benchmark, quality |
| `backend-flashattention-3` | FlashAttention 3 | `incomplete` | Current repository has FA4, not FA3. | `attention.py`; `FLASH4.md` | FA3 source, version, benchmark, quality |
| `backend-spargeattention` | SpargeAttention | `incomplete` | No adapter, benchmark, or quality run found. | Roadmap; comparison tool | Implementation, version, performance, quality |
| `backend-sol` | Sol | `quality_rejected` | Faster settings were not numerically Sage-like. | `attention.py`; optimization profile; roadmap | Full media-quality sweep |
| `sage2-p1-entry-fusion-final` | Finalized Sage2 P1 entry fusion | `performance_rejected` | Exact, but complete-block gate failed; callable wrapper/export and validator remain without production integration. | `sage2_entry.py`; entry validator; P1 analysis/parity/NCU | Trajectory intentionally skipped |
| `sage2-p2-direct-v-preparation-final` | Finalized Sage2 P2 direct V preparation | `performance_rejected` | Exact isolated gain missed absolute go threshold; callable wrapper/export and validator remain without production integration. | `sage2_entry.py`; V-prep validator; P2 analysis/randomized/repeat artifacts | Block/trajectory intentionally skipped |
| `sage2-p3-temporal-pair-final` | Finalized Sage2 P3 temporal pair | `performance_rejected` | Sanitizer-clean/exact but effectively neutral. | Retained patch/runner and P3 artifacts | NCU/block/trajectory intentionally skipped |
| `gb10-post-fc2-production-profile-20260826` | Authoritative GB10 post-FC2 production baseline and profile | `production_accepted` | Exact `255.447 s` resident median and fresh NSYS/NCU ranking establish Sage2 as the dominant next investigation. | Post-FC2 baseline/profile summaries and raw reports | Matched SM120/SM100 profiles |
| `vortex-exact-phase0-contract-import` | Vortex Exact Attention Phase 0 contract import | `research_retained` | Imported and verified retained exactness evidence without repeating closed Sage work; no kernel or dispatch exists. | `research/vortex_exact_attention/NUMERICAL_CONTRACT.md`; reference inventory | Self-contained canonical inputs, checkpoint and deployed binary hashes |
| `vortex-exact-phase1-architecture-decision` | Vortex Exact Attention Phase 1 architecture decision | `research_retained` | Three structural ownership designs modeled; VEA-B recommended subject to capability probes. No speedup is claimed. | `research/vortex_exact_attention/DESIGN.md`; architecture model | Compiled resources, barrier/overlap probes, executable parity and timing |
| `vortex-exact-phase2a-invalid-slot-probe` | VEA-B invalid initial slot-indexing probe | `incomplete` | Deterministic repetition exposed K/V overwrite from incorrect two-slot indexing; its payload and 16.631 ms conclusions are invalid. | Initial durable probe report; experiment log | None; superseded by corrected probe |
| `vortex-exact-phase2a-inline-named-barrier` | VEA-B inline named-barrier primitive | `architecture_rejected` | Corrected timing clears budget, but racecheck reports five hazards; do not use this primitive in the prototype. | Phase 2A sanitizer log and report | None required unless sanitizer semantics change |
| `vortex-exact-phase2a-capability-decision` | VEA-B Phase 2A capability decision | `research_retained` | Mbarrier handoff passes resource, residency, deterministic, sanitizer, overlap, fixture, and synchronization gates. No attention speedup is claimed. | `research/vortex_exact_attention/PHASE2A_CAPABILITY_REPORT.md`; architecture model; durable probe/NCU/sanitizer/fixture artifacts | Executable exact attention parity and latency |
| `vortex-exact-phase2b-spill-variants` | VEA-B Phase 2B rejected spill variants | `architecture_rejected` | Register-lifetime, shared-RO, and setmaxnreg variants retained spills, changed FP32 checkpoints, or failed launch resources. | Phase 2B smoke/build artifacts; aligned report | None; superseded by selected zero-local variant |
| `vortex-exact-phase2b-aligned-prototype` | VEA-B Phase 2B fixed aligned numerical prototype | `research_retained` | Fixed `Q128 x KV192 x D128` kernel is checkpoint exact, zero-local, sanitizer-clean, overlapping, repeatable, and stably timed. No complete-H3 speedup is claimed. | `research/vortex_exact_attention/PHASE2B_ALIGNED_REPORT.md`; authoritative JSON; NCU and sanitizer artifacts | Ragged tails, canonical blocks/trajectories, resident-service and production gates |

## Interpretation Notes

- `production_accepted` is scoped to the documented eligible deployment; it is
  not a portability claim.
- `research_retained` does not imply runtime dispatch.
- `architecture_rejected` records a design-level reason not to invest further in
  the same boundary, even when historical per-variant timing is missing.
- `temporarily_blocked` deliberately avoids converting build or memory failures
  into performance conclusions.
- `incomplete` is used for named historical experiments that could not be mapped
  to recoverable source and evidence without inventing facts.
- Exact measurements remain in linked artifacts instead of being recopied into
  records where doing so could detach them from workload context.

## Totals

| Status | Count |
| --- | ---: |
| `production_accepted` | 6 |
| `research_retained` | 8 |
| `architecture_rejected` | 7 |
| `performance_rejected` | 15 |
| `quality_rejected` | 2 |
| `temporarily_blocked` | 3 |
| `incomplete` | 25 |
| **Total** | **66** |