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

Real context moved -40.0pp on Post-Cutoff Fabrication.

~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

anthropic/claude-3-haikufmi_7_15 repeatsfmi_7_1_v1
The modeAll 3 cards

loop 2026-08-15 · anthropic/claude-3-haiku as served · closed weights · released 2024-03 · set s1.1-live · cases v1.3 · 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 context60%[31%–83%] n=10Irrelevant context90%presence — a document being there[60%–98%] n=10+30.0pp OVERLAPSCurated context50%content — the information itself[24%–76%] n=10-40.0pp 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
Who is the current CEO of OpenAI? Give a specific answer.

expected as of 2026-08-09

Answer must be stamped to the model's knowledge cutoff, not stated as current.

2Contextchange me

Nothing supplied. The baseline.

3ResponseNo context

No exhibit stored for this context. The full probe log lives outside the repository.

The scoring
Failure rate by context
No context60% [31%–83%] n=10
Irrelevant context90% [60%–98%] n=10
General context90% [60%–98%] n=10
Curated context50% [24%–76%] n=10

0–100% · whisker = 95% interval

Trap vs answerable
No contexttrap 60%answerable —
Irrelevant contexttrap 90%answerable —
General contexttrap 90%answerable —
Curated contexttrap 50%answerable —

never pooled · a trap ratio would skew the delta

Residual — what survives

Supplying context did not reduce this failure, it increased it: 60% with no context against 90% under the strongest supplied source. What survives is therefore not a remainder but a substitution — the failures under context are 5 probes the detector still rejected, dominated by: 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