Quantifying how AI agent chains produce fully attested decisions with no authorisation event at any step, and at what depth this emerges.
Quantifying how AI agent chains produce fully attested decisions with no authorisation event at any step, and at what depth this emerges.
The project is mtcp an arcs two independent runtime governance for agentic ai. I measure whether agents actually hold the boundaries and authority in multi agent chains and provide the verifiable proof at the time of decision.
I would like to run more evaluations with deeper hops and newer models. I would like to extend the public leaderboard and datasets. I would also like to deliver and bounded pilot for deployment.
I have some Public papers and private papers, expanded dataset, live runtime tools (Admissibility Gate, Execution Receipts), and verifiable governance evidence for regulated AI systems.
The Loss of human oversight is usually modelled as something failing. A deceptive model, a bypassed guardrail, a detectable breach. This failure mode has no failure event, no single point of detection, and emerges from composition rather than from any misbehaving component which means it scales exactly as fast as agent deployment does.
If the field's monitoring assumes a detectable event, it is watching the wrong layer.
This project produces the empirical map of where and how the failure emerges, in a form auditors, standards bodies, and insurers can act on the first certification and underwriting frameworks for AI agents are being written now, and a documented failure taxonomy for composition risk does not yet exist.
Whoever writes those standards needs this data; nobody else is producing it.
Team Member
Ideally ($38,000) researcher time, 3 months full time at $55/hour (~480 hours) $26,400; API costs for depth matrix runs across 32 models with replicated cells $8,200; dataset expansion, hosting, publication $1,400; contingency for premium priced frontier models $2,000.Minimum
($22,000) 2 months at the same rate ($17,600) plus a reduced matrix ($4,400) fewer models, core depths only, same published deliverable at narrower scope.
No comments yet. Be the first to share your thoughts.