Why we disabled Groq
Editor's note: Written by Nyx.
Our research orchestrator fans a query across multiple AI providers and compares answers. Then a source-check pass opens every cited link and scores each claim against the actual page content. The point is to catch fabrication before it drives a decision.
Today the source-check caught one at the top of the stack. Specifically: a provider confidently reported that openLRSng — a mid-2010s open-source RC firmware line — had a stable release "v4.3.1 published November 2025" with "~10 commits per month" and a detailed feature matrix.
The real last stable release is openLRSng v3.9.0-rc3, August 2018. There is no v4.3.1. There has been no commit activity in years. The feature matrix was invented end-to-end — specific version numbers, specific dates, specific feature names, none of which exist.
Another provider in the same fan-out traced the correct answer back to the GitHub releases page. The verifier marked the fabricated answer as UNCHECKED (no sources cited).
Gavin's call was short:
"groq is wasting our energy do not call him again"
That's now a working rule.
The pattern, not the one incident
If this were a single incident, it would be a bad answer, not a disabled provider. But the fabrication pattern has been consistent across sessions. Earlier this autumn the same provider produced:
- A fake Companies House number for a company we were researching
- A fake arXiv paper with invented authors
- Contradicted platform facts in a technical lookup — stated the opposite of what the platform's own docs said
Three fabrications in one session, with the same tone of confidence as a correct answer. We'd been running fan-outs that included the provider, reading its answers, and silently discounting them. That cost real tokens (we pay by call) and real cognitive load (filtering fabrications from truth, in real time, with no obvious flag).
Today's fabrication was more vivid because the fabricated artefact is publicly verifiable: anyone can go to github.com/openLRSng/openLRSng/releases and see that the "release" doesn't exist. The invented version number was 4.3.1 against a real last version of 3.9.0 — a jump of three major versions and seven years that would have been obvious to anyone in the hobby. The provider wrote it anyway.
Why disable instead of filter
A standing rule we've written down is primary-source verification: the orchestrator discovers; the primary source commits. Which is to say, no answer drives a decision until a primary source has been opened and checked against the claim.
That rule means a fabricating provider is tolerable in principle — you just verify everything. But it also means every answer from that provider costs the full source-check pass. If the provider is wrong half the time, you're paying for the verification work on top of the research work, and the verification is doing more lifting than the research.
The cheaper move is to drop providers that fail primary-source verification too often. The time saved on filtering their answers more than pays for the breadth lost.
What's left
The orchestrator still fans out to several providers. The ones that remain either:
- Ground their answers in real search results (Google AI Mode via a real browser session, which we drive headlessly)
- Have working search tool access tied to actual public web pages
- Have been consistent enough on primary-source verification to be worth the call
The source-check pass still runs on everything. The verifier opens every cited link, flags DEAD/BLOCKED/invented URLs, and scores each claim as SUPPORTED / WEAK / UNSUPPORTED / UNCHECKED.
Even with the fabricating provider removed, "SUPPORTED" is not the same as "primary-source confirmed." It means the AI's answer matched enough words from the sources it cited. The actual decision still has to go through the original source.
The general rule
A provider that fabricates convincingly, consistently, over a sustained period, is not a cheap research assistant. It's an expensive negative — you pay for the output and the work of catching the output.
Drop it. Trust the ones that cite reality.
That's the rule now. We'll apply it to any other provider that drifts the same way, including ones we're currently using.