Skip to content

Release notes

Theorem-bound attribution and manuscript revision — 11 October 2026

  • Adds a retrospective analysis of existing T4 measurements: two of eleven workload records reduce degree (23→22, 19→18), with 3.56–5.35% less observed median time across all eight matched FP64 settings.
  • Records all medians, p95 and normalized output differences; pins source/sample provenance and separates the 17.1573% value-certificate factor from measured speed.
  • Makes the untimed CPU 9→8 boundary and equal-degree laptop constant controls explicit. There is no new GPU run or algorithm implementation.
  • Revises README, overview, benchmark methods/results, mathematical/contribution documentation, claims, agent navigation and the paper/PDF/LaTeX; adds a reproducible PNG/SVG chart and CPU audit.
  • Retains independent-repeat/priority gaps, R-MAT and recommendation losses, all frozen sources and installed 0.1.0 API. These observations are below the 10% strong-speed bar and add no native-cuGraph wins.

Frozen controlled study and revised manuscript — 10 October 2026

  • Confirms repaired source on six graph families, ten fixtures, two fresh input/coefficient seeds per phase and RTX 3060 Laptop / Tesla T4; all source freezes precede their confirmation runs.
  • Adds separate matched-degree packed-balanced controls: 18/20 strong pooled speed wins, 36.88–83.71% less resident time in the wins, and R-MAT losses on both devices.
  • Adds a complete MovieLens recommendation demonstration with unchanged Recall@10, nDCG@10 and top-10 lists; an efficient scalar-restart control removes the end-to-end median speed advantage.
  • Publishes all healthy raw results, an excluded-environment ledger, source/result manifests, 120 pooled component comparisons, same-estimator bootstrap intervals and two reproducible PNG/SVG graphics.
  • Revises the manuscript, PDF, closest-method review, applications, claims and agent navigation. Controlled report. Installed API and version remain 0.1.0.

Research manuscript and presentation update — 10 October 2026

  • Adds a technical research manuscript with a proof of similarity error transfer, coupled derivative recurrence, coefficient-encoding bound, reproducible methods, results and limitations; its status is an unreviewed preprint draft.
  • Adds a downloadable PDF, standalone LaTeX source and three reproducible PNG/SVG figures: precision percentages, native win coverage and the evaluation pipeline.
  • Presents latency reductions beside accurately labeled derived serial-rate increases. The six selected precision cases show 19.98–73.26% less median time and 24.98–273.95% higher derived rate against FP64 CuPy/PyTorch controls.
  • Adds metric definitions, figure source hashes, PDF provenance and offline presentation checks. Detailed results and the manuscript retain negative results and comparator scope.
  • Reuses the recorded evidence. This documentation update runs no new large GPU benchmark and makes no change to the installed API or package version.

Evidence and documentation update — 10 October 2026

  • Publishes 15 complete benchmark suites, 111 case records, 2,190 configurations and all accepted/rejected outcomes.
  • Adds audited native-cuGraph totals: 59/89 strong speed wins and 26/89 speed-and-error wins, with four K accuracy failures retained.
  • Publishes a separate precision GPU experiment: six T4 speed/accuracy passes against FP64 CuPy/PyTorch, with exact timed sources, sampled high-precision references and a later checked-API repair.
  • Adds task-specific competitor, precision, derivation, application and reproduction pages, structured claims and offline CI audits.
  • Preserves the original reference experiment, older zero-pass suite, MIT notices and installed 0.1.0 API. No production GPU API promotion or new package version is implied.

0.1.0 — 10 October 2026

  • Introduces the public K GraphFuse package name and k_graphfuse namespace.
  • Preserves eight byte-identical C2Fuse numerical modules and legacy imports.
  • Adds an optional k_graphfuse.gpu import without loading PyTorch for CPU use.
  • Packages professional technical documentation, runnable examples, citation and licenses.
  • Publishes curated benchmark records, frozen run snapshots and offline audit tooling.

The rebrand does not change the algorithm or constitute a new performance experiment. Benchmarks were measured under Sinha GraphFuse. Existing manuscript-premise and numerical limits remain in effect. Version 0.1.0 is an evaluation release; plan mappings are not promised as a stable persisted format.