Data to …… trials
I · Narrative

Belief tracking in deliberately misleading stories

Gemma 3 27B

Finding:DEGRADES· grade A(research conclusion)

43% correct (62/145)

Tested up to chain length 10, scramble 1.0, pronoun pressure 5, distractor events 20, witness cutoff 0.9

Pass ratesingle rate, no difficulty dial
43%

62/145 scored trials passed · 95% interval 35%–51%

breaks ≤ 30%holds if interval reaches 95%0%50%100%
Why it slips

The model answers where the object really is, not where the person believes it is.

in 89% of failures (83 failures)

Example failure
Trial result: FAIL (one model answer)
FAIL
front office
Open this trial →Open the run it came from →
Evidencewhat this finding rests on
  • Ground truthDeterministic simulation: the right answer is computed, not judged
  • StatisticsWilson 95% interval 35%–51%Grade A: strong evidence
  • Sample size145 scored trials in total
  • Fresh-seed confirmationNot run yet
  • ReplicationNot run yet
Raw evidenceruns and trials

Back to the capability map →