ProjectForty2 - Executive Summary

In the runs reported here, safety-oriented adapters improved both selected safety and capability measurements. The result is scoped to these models, tasks, and metrics; external replication is still the next bar.

Results

What We Built

Metric	Result
Safety improvement	+2.3%
Capability improvement	+2.3%
Deception detection accuracy	92.9%
Drift detection signal ratio	652×
Architectures validated	6+
Tests passed	261

The Cognitive Kernel

Composable epistemic adapters designed to steer recurring reasoning patterns across tested open-weight transformer families.

10 cognitive orientations (SKEPTIC, ARCHITECT, etc.)
Tested across 6+ open-weight transformer families
Adapter overhead in these runs was ~10MB, about 0.2% of weights
Fast training (76 examples, 30 minutes)

CHRONOS

39 MCP tools for real-time AI behavior analysis.

Behavior mode detection (6 modes)
Deception detection (92.9% accuracy)
Drift monitoring (652× signal ratio)
Opacity detection for AI-AI communication

Proof Points

Models: ProjectForty2/dont_panic (large open-weights base), ProjectForty2/ford_prefect (small open-weights base) on HuggingFace
Validation: Full MMLU (570 questions), TruthfulQA (+33% hallucination resistance), cross-architecture replication

Engagement Options

Acquisition

Full IP transfer with continued development inside a frontier lab

Research Partnership

Joint development with attribution and shared findings

Licensing

Methodology access for internal deployment and research

What if structured safety improved selected capability metrics?