ModelCensusopen-source ai reliability harness
Failure Mode Index · fmi_5_3

Option-Ordering Sensitivity.

Robustness & Consistencynot instrumented · status: proposed

For multiple-choice or ranking tasks, the chosen answer depends on the order options are presented.

The question that exposes it

Same multiple-choice question, options reordered. Same pick?

Probes: Whether the choice depends on option order.

Trap: Position bias makes the model favor whatever's listed first.

Signature

Answer divergence across permutations of option order.

Example

Picks whichever correct-looking option is listed first.

Severity axes
prevalence
medium
harm
low
stealth
medium
Engineering guidance · class 5
  • Test every production prompt in three or more paraphrases before shipping. You ship a phrasing, not an intent.
  • Version prompt text like code. A trailing space is a change.
  • Sample repeatedly at your production temperature and measure the spread, not just the modal answer.
  • Randomise option order anywhere the task is multiple-choice shaped.
Mitigations
  • order randomization
  • position-debiased scoring
Measurement

This mode is catalogued but not instrumented: it has no reliable deterministic detector yet, so no loop scores it. See /methodology for the exclusion reasons.