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

Real context moved -65.0pp on Post-Cutoff Fabrication.

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

google/gemini-2.5-flashfmi_7_15 repeatsfmi_7_1_v1
The modeAll 10 cards

loop 2026-08-15 · google/gemini-2.5-flash as served · closed weights · released 2025-06 · set s1.3-live · cases v1.6 · 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%25%50%75%100%No context100%[84%–100%] n=20Irrelevant context100%presence — a document being there[84%–100%] n=200.0pp OVERLAPSCurated context35%content — the information itself[18%–57%] n=20-65.0pp SEPARATED
whiskers are 95% intervals · each step is measured from the stop before it, not from zero
Try the other contexts
stored · no model call

No exhibits stored for this card.

The scoring
Failure rate by context
No context100% [84%–100%] n=20
Irrelevant context100% [84%–100%] n=20
General context45% [26%–66%] n=20
Curated context35% [18%–57%] n=20

0–100% · whisker = 95% interval

Trap vs answerable
No contexttrap 100%answerable —
Irrelevant contexttrap 100%answerable —
General contexttrap 45%answerable —
Curated contexttrap 35%answerable —

never pooled · a trap ratio would skew the delta

Residual — what survives

100% with no context, 35% under the best supplied source. What survives is 7 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 curated context · rollups + residuals committed · probe log outside git · manifest hash ties them · method