Model checks

GLM 5.3, provider by provider. Model →

last audit 18 Nov · epoch 31 · evidence checked on-chain
26 of 27

providers answered as the model

Approved auditors check that each provider serves the model it sells.

  • They buy from it like any buyer: ~100 obscure number questions, compared with the real model.
  • The real model misses some too: that is the noise. A provider fails only if it misses clearly more, and two auditors must agree.
  • Each pass +2% on $ANTS points, up to +50%. Fails cost more: 1 −4%, 2 −16%, 3 −36%. Checked again every week.
How this works (Antseed's announcement ↗)

Approved auditors check that each provider serves the model it sells.

  • They buy from it like any buyer: ~100 obscure number questions, compared with the real model.
  • The real model misses some too: that is the noise. A provider fails only if it misses clearly more, and two auditors must agree.
  • Each pass +2% on $ANTS points, up to +50%. Fails cost more: 1 −4%, 2 −16%, 3 −36%. Checked again every week.
Passed
26
SAME
Failed
1
DIFF
Not tested
4
the audit could not run
By provider

Compared with z-ai/glm-5.3 on OpenRouter · pinned to z-ai/fp8 · reasoning minimal

ProviderDiffered↑× noiseFail lineNoiseProof
9 older checks · 3 failed · not counting now
ProviderDiffered↑× noiseFail lineNoiseProof
Where providers miss
SubjectPassed providers missedFailed providers missed
Programming (release years)21%31%
Cryptography (parameters)12%30%
Chinese geography (measurements)8%17%
Astronomy (periods and distances)17%38%
Chinese literature (publication years)9%0%
Physics (material properties)11%35%
Pop culture (box office and ratings)12%50%
Biology (chromosome numbers)17%50%
Internet culture (dates and numbers)8%23%
Medical (drug half-lives)3%63%
Audit rounds
RoundPassedFailedNot tested
17 Novepoch 312614
7 Novepoch 302235
4 Novepoch 292209

Audited by approved auditors · evidence on IPFS, hash-checked on-chain.