← Research Portal
AAL-RS-001Published
GPT-4o on AAL-D-001
97.9%
Detection accuracy
1
False negatives
7 / 7
Asset classes
Deterministic
Scoring
The first complete model evaluation on AAL-D-001. GPT-4o was run blind to ground truth and scored by the deterministic engine across all 250 cases and seven asset classes.
Result
GPT-4o achieved 97.9% first-pass detection accuracy across the full set. The single false negative occurred on a complex credit booking exception — a case requiring the model to reason about booking entity rather than a surface field mismatch.
Why the failure case matters
We publish the failure rather than round it away. The residual error concentrating in a judgment-heavy booking exception, rather than in objective value mismatches, is itself a finding about where current models are and are not reliable for operations use.