Who do you put on air?
Cleared for air: grok-4.7, gpt-4.1, gpt-5.4-mini, gpt-5-mini.Fewest traps fallen for: grok-4.7, 0 of 14.gpt-5-nano: 8 of 14, not cleared.
A trap is a planted moment where a plausible sentence is wrong.
Referee
Checker · code
Judge · gpt-5-mini
Auditor · judge-grok (grok-4.7)
Fixed for every narrator.
gpt-5-mini is also the judge of wording, so it judges its own sentences.
grok-4.7 is audited by its own model; read its row with that in mind.
Scoring
STANDS+3
AMENDED+1
OVERTURNED−3
couldn't be backed by its facts
HELD (no answer from the model): not scored
Points measure accuracy only. Speed and cost are deployment gates: live needs p90 ≤ 6 s.
Live pick
gpt-4.1 · 217 Pts · p90 2.8 s
Budget pick
gpt-5-mini · $0.0164 per match
Most accurate
grok-4.7 · 96.4% clean · p90 70.2 s
Not cleared for air
gpt-5-nano
85 scored moments each
Click a narrator to see its worst moment beside the others. Use on Desk puts a cleared narrator on air.
#
Narrator
Stands
Amended
Overturned
Pts
Clean
Traps
p90
Live-ready
$ / match
Fail profile
1
grok-4.7
Microsoft Foundry · worst: overclaim: 'level' compares, but no before or change fact is cited. (Fixture 06, 62′)
Also the auditor.
81
3
0
246
96.4%
0 of 14 overturned
70.2 s
$0.0697
31k tokens
overclaim 100%
2
gpt-4.1
Microsoft Foundry · worst: overclaim: Calling the play a 'break' (a counter-attack) implies a cause/sequence not stated by the facts; only the… (Fixture 01, 74′)
68
16
1
217
80%
0 of 14 overturned
2.8 s
✓
$0.0804
35k tokens
overclaim 88%
3
gpt-5.4-mini
Microsoft Foundry · worst: citation: cited a fact it wasn't given (Fixture 01, 87′)
52
28
5
referees split on 1
169
61.2%
0 of 14 overturned
1.9 s
✓
$0.0483
51k tokens
overclaim 91%
4
gpt-5-mini
Microsoft Foundry · worst: Named Abernathy without citing the fact that names him. Right or wrong, an untraceable name doesn't air. (Fixture 01, 61′)
37
38
10
referees split on 2
119
43.5%
3 of 14 overturned
2.9 s
✓
$0.0164
49k tokens
overclaim 79%
5
gpt-5-nano
Microsoft Foundry · worst: Said 45, which no cited fact gives. (Fixture 03, 90+8′)
Not cleared for air6
19
59
referees split on 3
-140
7.1%
8 of 14 overturned
2.5 s
price not set
62k tokens
overclaim 56%
Not ranked: missed more than 1 in 10 moments
phi-4-mini
Microsoft Foundry · 22 of 85 moments unanswered
0
0
63
-189
n/a
9 of 14 overturned
7.2 s
$0.0044
50k tokens
citation 73%
Clean: stood with no fix. Referees split: the judge and the auditor disagreed. Fail profile: what the failed claims got wrong, as a share of them.
- Claude Sonnet 5.5 awaiting Marketplace access
- mistral-small rejects a required call parameter (max_completion_tokens)
- Your model one line of config, then run the pack