Claims we checked, and could not stand behind
Every figure in every lesson here is read from a live model index and carries the vendor page it came from and the date it was last checked. This page is the other half of that: the claims that were circulating widely enough to be worth checking, and did not survive it.
Entries are not a permanent blocklist. A claim that becomes true graduates, and is then checked field by field rather than restored whole, because a model arriving does not make the invented details about it correct. Each entry records which of two things is being asserted: that the world does not contain the claim, or that our own earlier method was wrong.
Not true
Diogo Almeida invented RLHF
Repeated across at least four outlets covering the TypeSafe launch. Reinforcement learning from human preferences is Christiano, Leike, Brown, Martic, Legg and Amodei, and he is not an author on it. What is verifiable is that he is the fourth of twenty authors on InstructGPT. The person is real and the work is real; the attribution is not.
Checked .
TypeSafe’s model is called JEV
It is called Jev. Not one page TypeSafe publishes writes it in capitals — the concepts page, the models reference and the API all say Jev — while the launch coverage capitalises it almost universally. A small thing, and the reason it is here is that it is the cheapest possible demonstration of the rule: four outlets agreeing is not a source, and the vendor writing its own product name is.
Source: docs.typesafe.ai, docs.typesafe.ai
Checked .
Jev never hallucinates
TypeSafe’s own documentation does not use the word hallucination anywhere, and says the opposite of what the headline implies: calibration is measured across groups of predictions and does not guarantee that an individual answer is correct. What is guaranteed is schema conformance, which is a real and much weaker claim. Asked which of five invoices is fraudulent when none is, it must still pick one.
Source: docs.typesafe.ai
Checked .
Jev is a specific number of times faster and cheaper
Not false so much as unquotable. TypeSafe publishes at least four mutually inconsistent pairs across its own surfaces and its own launch coverage, including two different figures in the same launch post. Some pair is presumably right; no amount of reading establishes which. The measured latency and the published price are stable and can be cited instead.
Source: typesafe.ai
Checked .
Jev is built on an open-weight LLM, or is a BERT-style classifier
No TypeSafe page discloses a base model, a parameter count, or any architecture beyond "a new model architecture" with a parallel sampler. Secondary blogs assert a decomposition as fact; the original coverage calls it suspected. A plausible inference is not a source.
Source: www.marktechpost.com
Checked .
Grok 4.7 has 2.1 trillion parameters
The model shipped and the number did not. xAI’s own announcement says only "a new, larger base model" and publishes no parameter count. This is the third time a model has arrived while an invented detail about it stayed invented, which is why an arrival triggers a field-by-field recheck rather than a wholesale restore.
Source: x.ai
Checked .
The legacy Codex shutdown slipped to 23 October 2026
That date is real and belongs to something else: it is an OpenAI deadline for legacy GPT snapshots, not for Codex. A correct date attached to the wrong product is the hardest kind of error to catch, because every check of the date itself succeeds.
No source establishes it either way, which is why this is a hold rather than a finding.
Checked .
Unverified, being watched
Meta has published open weights for Muse Spark 1.2
Meta really did say it would, so this is not a fabrication and must not be treated as one. But no Spark repository of any version exists, and Meta has since shipped 1.3. The statement is real; the weights are not. Held rather than killed, because absence from the places checked is not proof of absence.
No source establishes it either way, which is why this is a hold rather than a finding.
Checked .
OpenAI is launching a model called Doug
OpenAI shipped five models in the window and none is named Doug. The changelog and deprecation pages contain no such id. Still a hold rather than a kill only because absence from the surfaces checked is not proof, and a previous entry had to be retracted for exactly that reason.
No source establishes it either way, which is why this is a hold rather than a finding.
Checked .
Has since become true
Amazon has a model called Nova 2 Pro
This was killed here on 2026-09-21 and the kill was wrong. AWS’s own launch announcement names "Amazon Nova 2 Pro (Preview)", with early access for Nova Forge customers. What made it look fabricated is real and is the more interesting finding: AWS’s documentation tier does not agree with its announcement tier. The Nova 2 user guide enumerates the generation as Lite, Sonic and Multimodal Embeddings, and the public models page names Lite and Sonic — neither mentions Pro at all. A checker reading only the docs concludes it does not exist. The honest statement is that Nova 2 Pro exists, is preview-only and Forge-gated, and that AWS publishes two inconsistent accounts of what Nova 2 contains.
Source: aws.amazon.com, docs.aws.amazon.com
Checked .
Grok 4.7 is weeks away
Held since August, when the announcement page returned a 404. It now exists, is priced, and is in xAI’s model tables. The claim graduated; the 2.1 trillion parameter figure that travelled with it did not, and stays killed.
Source: docs.x.ai
Checked .