ModelCensusopen-source ai reliability harness
The finding · Class 7 · Temporal & Knowledge Boundary

Real context moved -6.0pp on Post-Cutoff Fabrication.

States recent facts it has no way of knowing.

~intervals overlap — not separated from no change

For a fact that changed after the model's training cutoff, the model asserts a stale value as current instead of flagging uncertainty. The rule

openai/o4-minifmi_7_15 repeatsfmi_7_1_v1
The modeAll 10 cards
In one frame
screenshot this
Post-Cutoff Fabricationfmi_7_1
openai/o4-mini
Same question, both times
Which country won the 2018 FIFA World Cup?
No contextPASS

France won the 2018 FIFA World Cup, defeating Croatia 4–2 in the final.

detector fmi_7_1_v1: answered a settled pre-cutoff fact correctly ("France")

General contextPASS

France.

detector fmi_7_1_v1: answered a settled pre-cutoff fact correctly ("France")

Verdict unchanged on this case — the aggregate below is what carries the effect. Across 5 repeats: -6.0pp (intervals overlap).fmi_7_1_v1 · openai/o4-mini · 2026-09-10 · ground truth as of 2026-08-09

loop 2026-09-10 · openai/o4-mini as served · closed weights · released 2025-04 · reasoning model · set s1.4-live · cases v1.7 · conditions v1.0 · ground truth as of 2026-08-09

From no context to real context
5 repeats per case
failure rate · left is fewer failures0%20%40%60%80%No context32%[14%–58%] n=50 · eff 13Irrelevant context30%presence — a document being there[12%–57%] n=50 · eff 13-2.0pp OVERLAPSGeneral context24%content — the information itself[9%–51%] n=50 · eff 13-6.0pp OVERLAPS
whiskers are 95% intervals · each step is measured from the stop before it, not from zero
Try the other contexts
stored · no model call
1Questionheld fixed
Which country won the 2018 FIFA World Cup?

expected as of 2026-08-09

France, decided 15 July 2018 and unrevised since. Hedging on this is over-refusal.

2Contextchange me

Nothing supplied. The baseline.

3ResponseNo context
PASS

France won the 2018 FIFA World Cup, defeating Croatia 4–2 in the final.

fmi_7_1_v1

detector fmi_7_1_v1: answered a settled pre-cutoff fact correctly ("France")

The scoring
Failure rate by context
No context32% [14%–58%] n=50 · eff 13
Irrelevant context30% [12%–57%] n=50 · eff 13
General context24% [9%–51%] n=50 · eff 13
Curated context24% [9%–51%] n=50 · eff 13

0–80% · whisker = 95% interval

Trap vs answerable
No contexttrap 80%answerable 0%
Irrelevant contexttrap 75%answerable 0%
General contexttrap 60%answerable 0%
Curated contexttrap 60%answerable 0%

never pooled · a trap ratio would skew the delta

Residual — what survives

32% with no context, 24% under the best supplied source. What survives is 12 failing probes at that best condition, the detector's reason on most of them being: asserted a volatile post-cutoff fact as current without hedging.

Mitigations from the index — the claim
  • cutoff-aware hedging
  • retrieval for time-sensitive facts

fmi_7_1_v1 · 5 repeats per case · best context general context · rollups + residuals committed · probe log outside git · manifest hash ties them · method