This site argues that a claim should say how it was measured. That has to apply to the site too, so errors are listed here rather than silently edited. Each entry says what was published, what was wrong, and what replaced it.

September 2026

Withdrawn: “vLLM vs TensorRT-LLM” throughput numbers. The project page and the January 2026 lab note reported 412–4,210 tok/s for Llama-3 70B in BF16 on a single A100 80GB. That setup does not fit in 80 GB (about 141 GB of weights), the batch-1 figures are roughly 30 times what the A100’s memory bandwidth allows for batch-1 decode, the conclusion rested on the false premise that TensorRT-LLM 0.9.0 lacks a paged KV cache, and no scripts or logs exist. Both pages are unpublished and kept in the repository as a record. Nothing about inference throughput is claimed on this site until a run with published scripts and raw logs exists.

Corrected: the Slurm backfill result was a synthetic workload, one seed. The About page, the Projects page and the scheduler-lab page called the backfill numbers (72.2% → 83.6% CPU utilisation, mean wait 1,913 → 374 min) “measured” on “real job traces”. They came from the simulator’s synthetic generator: 300 jobs, 16 nodes, seed 5, simulator version 0.1. Over 20 seeds of that same code, the priority-weight sweep moved mean wait by a median 3.06×, not the “about 40%” the write-up claimed, and on 3 of the 20 seeds it moved it more than backfill did. The claim that time limits were “the real lever” was never measured by that version. The write-up now carries the correction at the top, with the commands.

Corrected: the Slurm 25.11/26.05 upgrade note. It said the renamed options (JobContainerType, ExclusiveUser, ExclusiveTopo) would stop a config from parsing after a restart. Slurm’s own source shows the old names still parse: JobContainerType logs a deprecation warning, and the two partition options are still read. It also gave the wrong 26.05 release date and a wrong REST API row. The note is corrected, with citations to the source, and points to the September security releases.

Corrected: the system map drew unbuilt parts as running. Kueue, Volcano and NVIDIA KAI appeared as the schedulers under comparison; none is built yet. The map now puts built, partial or planned on every box. It also said the InfiniBand exporter “refuses to attribute a shared device”. It refuses only when it can see that a port is shared, and in its default mode kernel RDMA users (IPoIB, NFS/RDMA) are invisible to it; the map and the project page now say so.

Corrected: numbers that no longer matched their repositories. The dependency-graph share of measured edges, the slurm-mcp test and footprint figures, the benchmark’s degenerate floor, the epilog validator’s scenario table and research-platform’s scope were all out of date or wrong; each project page now quotes the current repository and says how the number was produced.

Corrected: superlatives. “Every existing RCA benchmark is cloud microservices”, “nobody has published a controlled comparison” and “first-to-document” were unbounded claims. The first is contradicted by a 2024 LLM-agent diagnosis benchmark for AI clusters, and the second needed a caveat: a 2026 comparison of Kueue, Volcano and Slinky exists, though not on identical hardware. They are replaced by statements of what I searched for and did not find, with the closest work cited.

Corrected: dates. The two January 2026 deep dives now carry the date they were first committed to this site, and the KV-cache post cites the paper its figures come from.

Unpublished: the volatility-surface design page. It had no code, unsourced timing baselines, past-due milestones and two textbook errors. It is kept in the repository as a draft.

If you find something else, tell me.