What's new at AgnosLogic.

Major

• Trained MLP factual probe on 10,416 diverse pairs from 7 sources (HaluEval, TruthfulQA, HellaSwag, Winogrande, MNLI, ARC, hand-crafted)
• 98.9% F1 on HaluEval QA (10K samples) — up from 94.3% in v51g
• 100% comparison accuracy on FreshQA with 0% false positive rate
• Resolved bf16 deployment mismatch via forward() normalization path

Improvements

• Sub-100ms inference latency on RunPod A100 80GB serverless
• Fixed reasoning text showing logic contradictions on verified facts
• Added auto-polarity correction at handler startup

Major

• Introduced MLP factual probe to replace linear probe — non-linear decision boundary
• 94.3% F1 on HaluEval QA, up from 76.5% with linear probe

• US Provisional Patent filed covering the CAEF architecture (FAH + CWMI + ESR + MLP probe)

Major

• Dual probe architecture: separate logic and factual probes
• ESR-only retraining on frozen backbone
• 62% strict accuracy on hand-crafted adversarial test set

Major

• First production-stable model: 76.5% Logic Detection, 70% adversarial catch rate
• 0.601 epsilon spread between true and false pairs
• Deployed on RunPod serverless infrastructure