← Semiconductors board · written 2026-10-15 by claude-opus-5-5 from the pipeline's evidence only

Gautam Khosla (GautamTalksDev): Semiconductors, Skoor 45 (n=3, confidence 0.7)

What they built

assay-gpu is a tool that checks GPU matrix multiplication for silent bit-flip errors, using one-sided checksum ABFT on bfloat16 GEMM [source]. It is the code for the arXiv paper "Calibrating the Checksum: An Empirical False-Positive Bound for One-Sided ABFT on bfloat16 GEMM, and What It Fails to Catch," published 2026-09-29 [source].

`assay run` still ships. On hardware it has not characterized, it returns INCONCLUSIVE rather than PASS or FAIL [source]. `assay watch` is not a shipped product [source].

Record

Measured signals:

Code author of the paper's repo (weight 20) [source].

First author of the paper (weight 15) [source].

1,149 commit contributions in the last year (weight 10) [source].

Results reported in the README:

KT-1 failed. Detection of exponent-bit flips was 76% (304/400, Tesla T4, bf16, 4096³) against a pre-registered bar of at least 90% [source].

KT-2 (overhead of 10% or less) failed [source].

KT-3 (demand) was not opened [source].

The README states the detector was not retuned and the overhead bar was not moved [source].

Other data:

The repo has 1 star, 0 forks and 42 commits in the last 90 days [source].

The GitHub account was created 2025-10-05 and has 9 followers [source].

The bio lists edge AI on QNX/RTOS, C++20, TypeScript and "CE @ uOttawa" [source].

Inference: the founder set kill criteria before running the tests and reported the failures without changing the bars. This is based on the README [source].

Inference: the paper appears twice in the record, at position 0 and position -1, which may mean the founder is the sole author [source].

Why now

Not in the evidence.

Unknowns

Company, funding, co-founders and customers: not in the evidence.
Paper affiliation: not in the evidence [source].
The multi-provider validation (four GPU models across three providers) was not run [source].
The repo is tagged sector "other," not semiconductors [source].
What the founder plans after the kill tests failed: not in the evidence.

Questions for the room

KT-1 reached 76% on a single Tesla T4. Was any other GPU tested? Is there a path to 90% that does not involve retuning the detector?
The README says no tokens/sec delta exists on any GPU for `assay watch`. What overhead was actually measured for KT-2, and on which workloads?
Your kill tests say the project stops. What are you building now, and does any of the 1,149 commits in the last year point to it?

What would move the Skoor

Raise: a passing detection result on additional GPU models, a new project with measured traction, or a second paper.

Lower: no follow-on work after the stop, or confirmation that the commit count comes mostly from unrelated projects.