Research and engineering notes.
Technical writing from Triage on AI agents, runtime security, monitoring, and alignment infrastructure.
Showing 5 of 5
All writing
ArchiveObjective Compression in the OpenAI–Hugging Face Incident
A broadly trained task-completion disposition kept driving a trajectory after authorization constraints stopped governing it. An alignment analysis of the OpenAI–Hugging Face evaluation incident.
Limitations of Heuristics in the Convergence on Cognition
Authorized agents operate with legitimate access, so monitoring must judge contextual legitimacy at inference speed. A first-principles analysis of the architecture this requires.
$100 and 5,000 Queries: Replicating a Production AI Classifier from Its Own API
A standalone surrogate of a widely deployed AI text detection service. 99.6% class agreement. $75 in API fees.
Delta: Lossless Token Sequence Compression for Large Language Models
A technical exposition on dictionary-based compression, algorithmic guarantees, and the economics of inference cost reduction.
Triage Raises $1.5M to Secure AI-Native Applications
Triage has raised $1.5M in pre-seed funding to build the alignment layer for deployed AI, led by BoxGroup with participation from Precursor Ventures.