Two official milestones
The raw source hash identifies file bytes. A submission payload content hash identifies the packaged submission and is a different value.
| Source | Official result | Source SHA-256 |
|---|---|---|
| 036 · 1.301203× | d296d0c1b1774e5e | 19a67530e519a9ca65db953c69c0ce5423bd0782d6629918cd5e6e40ed8d77e3 |
| 039 · 1.323430× | f7219946bf314f37 | 2e570ef2c4f3df70306a191e6eea79c75858e0379db056af582490de66857f77 |
Read the experiment records
The filterable evaluation ledger labels checks, scored runs, failures, and queued rows. The original 001–038 ledger is frozen; 039 is a separate addition.
- Historical ledger · CSV
- Combined evaluation records · JSON
- Thirty official 036 shapes · CSV
- 036 aggregate, machine, and evaluator identity
- 039 scored result and historical board snapshot
- Saved board read
Preserved source files
Presence in this list means a source was preserved. It does not by itself establish a successful correctness check or a performance win.
- 001-first-candidate.py
- 002-split-router.py
- 003-selective-gemv.py
- 004-router-dispatch.py
- 005-attention-fused-prep.py
- 006-moe-local-tiles.py
- 007-dense-splitk.py
- 008-layer-integration.py
- 009-packed-router.py
- 010-moe-fused-postnorm.py
- 011-layer-validated-components.py
- 012-attention-local-splits.py
- 013-moe-middle-tiles.py
- 014-hopper-gemv.py
- 015-transposed-dense.py
- 016-wide-k-dense.py
- 017-quack-large-dense.py
- 018-quack-skinny-swapped.py
- 019-quack-large-batched.py
- 020-quack-skinny-batched.py
- 021-measured-dense-bundle-ca23de4a9240.py
- 022-cublaslt-plan-search.py
- 023-vector-plan-search.py
- 024-router-plan-search.py
- 025-router-single-launch.py
- 026-expanded-bundle-c0f93b5a0027.py
- 027-paired-router-search.py
- 028-cublaslt-expanded-search.py
- 029-c1-bestof.py
- 030-c1-atomic-router.py
- 031-c1-tinygemm-unpooler.py
- 032-c1-cluster-qkv.py
- 033-c1-deep-cluster-qkv.py
- 034-c2-packed-cublaslt.py
- 035-c2-pooler-lt-layouts.py
- 036-c2-streaming-vector.py
- 037-c2-lean-preparation.py
- 038-c2-vector32.py
- 039-c20-pdl-router-lt15-skinny.py
Saved problem contracts
- muse_attention_decode/TASK.md
- muse_attention_decode/baseline.py
- muse_attention_decode/checks.py
- muse_attention_decode/reference.py
- muse_bf16_proj/TASK.md
- muse_bf16_proj/baseline.py
- muse_bf16_proj/checks.py
- muse_bf16_proj/reference.py
- muse_decode_layer/TASK.md
- muse_decode_layer/baseline.py
- muse_decode_layer/checks.py
- muse_decode_layer/reference.py
- muse_moe_w4a16/TASK.md
- muse_moe_w4a16/baseline.py
- muse_moe_w4a16/checks.py
- muse_moe_w4a16/reference.py
- muse_prefill_layer/TASK.md
- muse_prefill_layer/baseline.py
- muse_prefill_layer/checks.py
- muse_prefill_layer/reference.py
Scope and limitations
Official aggregate outcomes establish the evaluated bundles. They do not isolate every component’s causal percentage. Local measurements stay labeled local. Hardware counters were unavailable on the Runpod host. Historical rank 66 and the archived competition schedule are dated observations, not live account state.
Copied result exports omit transport attachments and verbose logs while retaining available score, correctness, and timing fields. The hash-pinned winner files are copied byte for byte. Original SVG explanations are schematic; only figures explicitly labeled measured use recorded measurements.