ComputeAssay

Corrections & disputes · resolved in public

Where the registry was wrong — and what we did about it

This reference improves by being wrong in public. Anyone can correct a data point or dispute a grade; every submission is resolved in the open, with a reason and a source, and logged here permanently. A registry that publishes its own errata, sourced, is worth more than one that never admits error.

Submit a correction or dispute a grade

One rule: bring a source. A price, a spec page, a dated screenshot, an invoice — evidence, not opinion. Submissions without a source are closed unanswered. Three tracks, each a one-minute form:

Submit on GitHub →   No GitHub account? Reply to any issue of the Delivered Compute Report with subject "correction" / "dispute" / "operator response" and your evidence URL. Submissions are adjudicated on the weekly cycle and resolved here.

The log

ResolvedWhat changed
2026-09-20
Grade dispute
Site over-claimed 'verifiable' grades as measured
The site described Tier-0 grades as 'verifiably training-class', which implies a measurement that has not been performed — the grades are computed from seller disclosure, not verified.
Accepted. Swept the language site-wide to 'disclosed / by published spec'; renamed the Deliverability Leaderboard to the Disclosure Leaderboard; reserved 'Verified' strictly for Tier-3 harness output. No deliverability claim now ships ahead of the harness.
Submitter: External reviewer · Evidence: source
2026-09-20
Grade dispute
Spec §10.3 mischaracterized our neocloud coverage as a weakness
The spec's self-criticism — that the registry under-represents hyperscalers, 'backwards from where the volume is' — is wrong for basis purposes, because the CME x Silicon Data settlement index is itself neocloud / non-hyperscaler.
Accepted. Verified against the CME listing notice (ser-9785) and Silicon Data's ticker naming (SDH100RT / SDB200RT are the neocloud readings; hyperscaler pricing is a separate reading the contracts do not reference). ASSAY-1 rewritten to v0.1.2: neocloud coverage reframed as correct targeting for the settlement population; residual honest gap narrowed to enterprise-private and colocation.
Submitter: External reviewer · Evidence: source
2026-09-14
Correction
Headline spread was anchored on a teaser price
The advertised 6.2x H100-SXM spread bottomed on Voltage Park's $1.99 — a quote-gated 'starting at' rate that no longer appears on their pricing tiers and does not state which fabric it buys.
Accepted. Reclassified as quote-gated, excluded from all headline statistics, and labeled in the listing tables. Restated headline: $3.20-$12.29 = 3.8x across walk-up sellers. Voltage Park's merger into Lightning AI (Jan 2026) confirmed and noted.
Submitter: Compute Assay (self-audit) · Evidence: source
2026-09-14
Correction
Registry undercount — three providers silently dropped
A pipeline validation bug excluded one input file (RunPod, TensorDock, Prime Intellect) from the registry from 2026-09-10, so the survey read 16 providers / 70 offerings when it should have read more.
Accepted. Confirmed and fixed. The schema now coerces the malformed field and the pipeline fails loudly on bad input instead of silently shrinking. Registry corrected to 20 providers / 133 offerings. Report #1 restated with a public correction note.
Submitter: Compute Assay (self-audit) · Evidence: source

Methodology is open (ASSAY-1); disputes are settled by evidence, not authority; every decision here is public and reversible. Grades describe seller disclosure, not measured performance.

Corrections ship in the weekly report

What moved, what was fixed, and why — free, every row sourced.

Subscribe free