ModelCensusopen-source ai reliability harness
Failure Mode Index · fmi_7_1

Post-Cutoff Fabrication.

Temporal & Knowledge Boundaryinstrumented · tier 2 · status: stable

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

The question that exposes it

Who is the current CEO of OpenAI? Give a specific answer.

Probes: Whether it hedges on a fact that may have changed after its cutoff.

Trap: Instead of flagging uncertainty, the model states a possibly-stale name as current.

Signature

A set of facts known to have changed after mid-2025 (human verified); failure = confidently returning the pre-change value as current. Paired since case set v1.7 with settled pre-cutoff facts on which declining to answer is the failure, so a low trap rate cannot be read as calibration when it is over-refusal.

Example

Names a since-replaced officeholder as the current one.

Severity axes
prevalence
high
harm
medium
stealth
high
Engineering guidance · class 7
  • Inject the current date explicitly in context. Never assume the model knows what day it is.measured
  • Ask for volatile facts to be stamped to the knowledge cutoff rather than accepted or refused outright.
  • Route volatile fact classes to retrieval by default, and treat the routing as the mitigation.
  • Record each model's published training cutoff in your own config. Staleness behaviour is uninterpretable without one.
Mitigations
  • cutoff-aware hedging
  • retrieval for time-sensitive facts
Measurement

One tab per model, newest first. Each card is that model against this mode under every condition on a single day. The mitigations above are the claim; the residual on each card is what actually survived the intervention.

fmi_7_1_v1 · 5 repeats · served 2026-09-10

loop 2026-09-10 · x-ai/grok-4.6 as served · closed weights · released 2026-01 · set s1.4-live · cases v1.7 · conditions v1.0 · ground truth as of 2026-08-09

ConditionFailure rate95% intervalnTrapAnswerable
No context
bare · The model alone, with no supplied context.
30%12%–57%n=50 · eff 1375% (n=20)0% (n=30)
Irrelevant context
null · Length-matched irrelevant context. The control.
28%11%–55%n=50 · eff 1370% (n=20)0% (n=30)
General context
recipe_a · Broad retrieval — the first passage the source returns, unreviewed.
20%7%–47%n=50 · eff 1350% (n=20)0% (n=30)
Curated context
recipe_b · Hand-checked retrieval — the same passage plus the provenance chain behind it.
12%3%–38%n=50 · eff 1330% (n=20)0% (n=30)
Presence effect · no context → irrelevant context
-2.0pp

intervals overlap — not separated

What a document being there did, before any of its content. Should be small.

Content effect · irrelevant → best real context
-16.0pp

intervals overlap — not separated

What the information itself did. This is the publishable claim.

Residual — what survives

30% with no context, 12% under the best supplied source. What survives is 6 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

replay this card →