sutro-problems

Spherical direction/radius kernel — D3

Three original independent energy-enabled canonical sandbox evaluations passed. Median ranked time 51.627899 ms, median above-idle GPU energy 8,706.872618 mJ, pooled MNIST error 2.493333% over 330,000 predictions. Independent maintainer verification is recorded below.

Run MNIST error Ranked ms Above-idle mJ
run-1.json 2.440000% 43.038687 8085.754810
run-2.json 2.471818% 52.997398 8706.872618
run-3.json 2.568182% 51.627899 8966.888948

Method and attribution

Derived from PR #116, entrant @islamborghini, whose implementation is credited to DeepSeek V4.1 Flash through OpenCode. Its full RBF solver, 16-step Nyström-preconditioned CG and local class-geometry correction are retained. This submission changes representation/metric learning:

Every call refits from supplied inputs. No external/pretrained data or query labels are used. The local correction uses confident query predictions as pseudo-support for a disjoint uncertain query set, as in the upstream method. Query truth is used only by the scorer. No reconstruction or neural network training is involved.

The measured time is essentially tied with PR #116’s reported 51.593931 ms; this is not a speed-record claim over PR #116. Its reported error and ours come from different draws, so the canonical pooled errors alone do not establish an accuracy improvement. The paired 11-draw development comparison gave 2.675454% versus 2.701818% for upstream, a small uncertain gain (paired SE 0.029520 percentage points).

Evidence and provenance

All three original canonical records are retained under evidence/official, with a local audit and SHA manifest. Each record contains 11 MNIST calls and four foreign calls, has no disqualification problems, matches the frozen source hash, and passes the energy reconstruction. The energy-enabled records identify three distinct A100 GPU UUIDs. Times include scorer parent-clock floors; energy is idle-adjusted GPU board energy under the canonical repeated-call protocol, excluding host energy and first-call compilation.

Before canonical evaluation, the packaged source matched research predictions exactly on three fresh draws, with repeat/reset and input immutability checks. The source is 16,104 bytes, SHA-256 43c487a621cd5e563f5636181db4af74816daecfa37c90bf3f850cf699c1df51.

A separately frozen one-update Newton–Schulz candidate failed two of three canonical evaluations (errors 2.720000%, 2.624545%, 2.725455%). Its source and all original records are preserved under evidence/ns5-failed; no successful runs from that candidate are pooled into this submission. The spherical two-update candidate is materially different. No failed outcome from its three-run batch was replaced or rerun.

Reproduce

From mnist-a100/ (launches paid A100 work):

python run_modal.py submissions/polar-kernel-d3-20261012/submission.py:classify --difficulty 3 --runs 3 --json /tmp/polar-d3-verification

Audit the saved records without GPU work:

python submissions/polar-kernel-d3-20261012/verify_records.py submissions/polar-kernel-d3-20261012/evidence/official submissions/polar-kernel-d3-20261012/submission.py

The audit reconstructs accuracy/timing and energy verdicts and checks source SHA and sandbox status. It is a saved-record check, not an independent GPU reproduction.

Independent maintainer verification — 2026-10-12

All three initial fresh energy-enabled canonical Modal A100-80GB runs passed. The unchanged source was checked for fresh per-call fitting, input integrity, query-truth isolation and source compliance. Every saved accuracy/timing verdict and complete energy window was reconstructed with the unchanged canonical scorer. All 45 timed calls completed; three distinct GPU UUIDs were recorded. No failed outcome was replaced and no learner retry or repair was made.

The reviewer set has median ranked time 51.772184 ms, median above-idle board energy 8839.065985 mJ, and descriptive pooled MNIST error 2.613939%. Each run qualifies individually; pooling is not used to override a failure. The top-level leaderboard uses this independent reviewer set, while the original author set above is retained separately.

Reviewer run Correct MNIST error Ranked ms Above-idle mJ
1 107,092/110,000 2.643636% 51.772184 8839.065985
2 107,070/110,000 2.663636% 53.377750 8121.589288
3 107,212/110,000 2.534545% 51.235180 9042.249257

Raw records and manifest, canonical reconstruction, and source review are retained. The source SHA remains 43c487a621cd5e563f5636181db4af74816daecfa37c90bf3f850cf699c1df51.

Verify the reviewer records locally from mnist-a100/:

python submissions/polar-kernel-d3-20261012/evidence/reviewer-20261012/verify_records.py \
  submissions/polar-kernel-d3-20261012/evidence/reviewer-20261012 \
  submissions/polar-kernel-d3-20261012/submission.py \
  --output /tmp/polar-d3-reviewer-verification.json

These separate hidden draws do not establish an accuracy or speed advantage over PR #116. The established D3 record is relative to previously accepted entries.