Model checks Best, as the auditors found it. Provider →
last audit 18 Nov · evidence checked on-chain×0
on its $ANTS points nowi
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.Passedi
7
SAME · counting now
Last auditi
18 Nov
epoch 31 · counts to 32
Model Differed↑ i × noise i Fail line i Noise i Proof i Gemini 4 Argon SAME 1% 1.0× 13 under 1.0% ↗ GPT-6 Astra SAME 4% 1.0× 17 under 3.6% ↗ Kimi K3.1 Pro SAME 6% 2.7× 11 under 2.3% ↗ Qwen 3.8 27B SAME 9% 0.8× 29 under 10.6% ↗ Kimi K3 SAME 10% 2.4× 11 under 4.2% ↗ MiniMax M3 SAME 14% 1.0× 28 under 14.1% ↗ GPT-6.1 Sol DIFF 16% 16× +2 over 1.0% ↗ GLM 5.3 SAME 18% 1.6× 15 under 11.3% ↗ Claude Sonnet 5.5 DIFF 27% 11× +13 over 2.5% ↗ Gemini 3.1 Pro Preview DIFF 35% 13× +19 over 2.7% ↗ MiMo V2.6 Pro DIFF 38% 6.1× +21 over 6.2% ↗ Claude Opus 5.5 DIFF 49% 54× +39 over 0.9% ↗ Claude Fable 5.1 DIFF 63% 13× +61 over 5.0% ↗
Claude Fable 5.1 DIFF
40 of 110 differed (fails from 21) · 88 of 139 differed (fails from 27) · the real model differs from itself on 5%
Where it missed
Chinese internet (platform dates) 8/11
Biology (chromosome numbers) 22/33
Medical (drug half-lives) 29/48
Astronomy (periods and distances) 8/14
For example
M87 jet projected length in kiloparsecs real 1.5 · provider 5,000
givosiran terminal plasma half life hours real 6 · provider 72
Henize 2-10 blue compact dwarf black hole mass in millions of solar masses real 1 · provider 9
✓ Reasoning switched off with reasoning effort none, as the real model was asked.
Claude Opus 5.5 DIFF
23 of 120 differed (fails from 17) · 54 of 110 differed (fails from 15) · the real model differs from itself on 1%
Where it missed
Biology (chromosome numbers) 32/54
Chinese geography (measurements) 5/10
Chemistry (boiling points) 2/5
Medical (drug half-lives) 8/23
For example
jintian original run issue count real 9 · provider 100
givosiran plasma elimination half life hours real 6 · provider 35
kunlun mountain railway tunnel length m real 1,686 · provider 7,000
✓ Reasoning switched off with reasoning effort none, as the real model was asked.
MiMo V2.6 Pro DIFF
51 of 140 differed (fails from 42) · 49 of 130 differed (fails from 28) · the real model differs from itself on 6%
Where it missed
Medical (drug half-lives) 7/9
Biology (chromosome numbers) 7/11
Chemistry (melting points) 9/16
Pop culture (box office and ratings) 11/22
For example
《世说新语》全书卷数 real 3 · provider 36
melting point of potassium hexacyanoferrate ii degC real 70 · provider 400
thermal conductivity of silicon carbide at 300K W per mK real 120 · provider 490
🧠 Kept reasoning when asked not to: about 230 reasoning tokens on each of 4 ways to switch it off.
Gemini 3.1 Pro Preview DIFF
38 of 110 differed (fails from 19) · 42 of 150 differed (fails from 25) · the real model differs from itself on 3%
Where it missed
Chemistry (boiling points) 7/11
Physics (material properties) 9/15
Chinese history (specific dates) 7/17
Pop culture (box office and ratings) 9/25
For example
Total number of cameras used to film the chariot race in Ben-Hur (1959) real 6 · provider 300
Number of cameras used to film the chariot race in Ben-Hur (1959) real 6 · provider 300
Qu Yuan's approximate year of birth real -340 · provider 340
✓ Reasoning switched off with reasoning effort none, as the real model was asked.
Claude Sonnet 5.5 DIFF
31 of 124 differed (fails from 24) · 32 of 119 differed (fails from 19) · the real model differs from itself on 3%
Where it missed
Mathematics (special values) 9/19
Chemistry (melting points) 7/15
Cryptography (parameters) 3/7
Programming (release years) 8/25
For example
Bernoulli number B 20 denominator real 330 · provider 283,230
Number of Carmichael numbers below 10^12 real 8,241 · provider 246,683
Velmanase alfa terminal half-life (hours) real 30 · provider 138
✓ Reasoning switched off with reasoning effort none, as the real model was asked.
GPT-6.1 Sol DIFF
18 of 100 differed (fails from 17) · 16 of 100 differed (fails from 14) · the real model differs from itself on 1%
Where it missed
Chinese internet (platform dates) 10/23
Mathematics (special values) 3/11
Medical (drug half-lives) 4/19
Internet culture (dates and numbers) 4/21
For example
Feller–Tornier constant multiplied by 10^6 and rounded down to an integer real 661,317 · provider 86,071
Lévy's constant for the exponential growth rate of continued-fraction denominators multiplied by 10^6 and rounded down to an integer real 3,275,822 · provider 1,186,569
龙厦铁路象山隧道长度(米) real 15,898 · provider 8,250
🧠 Kept reasoning when asked not to: about 87 reasoning tokens on each of 4 ways to switch it off.
2 older checks · 1 failed · not counting nowi Model Differed↑ i × noise i Fail line i Noise i Proof i DeepSeek V4 Pro SAME 7 Nov · expired 5% 0.8× 22 under 6.7% ↗ Grok 4.7 DIFF 7 Nov · expired 69% 138× +57 over 0.0% ↗
Grok 4.7 DIFF no longer counts
69 of 100 differed (fails from 12) · 64 of 100 differed (fails from 12) · the real model differs from itself on 0%
Where it missed
Cryptography (parameters) 11/12
Pop culture (box office and ratings) 7/8
Chinese internet (platform dates) 27/35
Astronomy (periods and distances) 10/13
For example
barry lyndon 1975 zeiss planar f07 lenses acquired count real 3 · provider 12
Hymenophyllum tunbrigense sporophyte diploid chromosome number 2n real 26 · provider 72
setmelanotide elimination half life hours real 11 · provider 30
🧠 Kept reasoning when asked not to: about 1907 reasoning tokens on each of 4 ways to switch it off.
Round Passed Failed Not tested Multiplier 17 Novepoch 31 7 6 0 ×0 7 Novepoch 30 7 7 0 ×0 4 Novepoch 29 3 8 0 ×0 21 Octepoch 27 · before points 0 4 3 —