ANTIBODY: An Immune-System Architecture for Long-Horizon AI Agent Oversight (Doctoral Research Proposal)

Correction (10 September 2026). The first version of this record's abstract stated that the candidate held twelve months of production data across more than fifty routed models. That was not accurate: the multi-model routing system referenced had been running for days, not months, and the fifty-model figure conflated a one-time model-comparison exercise with ongoing production use. This version replaces that claim with figures verified directly against the candidate's own cost-ledger files: 10,827 individually priced calls across 15 distinct models from 5 providers, in logs currently spanning roughly 18 days. Long-horizon autonomous agents degrade in ways current oversight cannot catch, and the tools used to catch them are themselves degrading. This proposal investigates an external, model-agnostic oversight layer built on five principles drawn from immunology and reliability engineering: a dormant failure memory that activates only on derailment and grows more informative with runtime; predictive, hazard-based intervention in place of fixed-schedule context management; a sequential monitor whose statistical guarantees strengthen the longer it runs; oversight instruments that are cryptographically keyed so they cannot be learned away by future model generations; and a controlled cross-generation test of whether keyed instruments survive where fixed ones do not. The candidate operates a multi-agent content-generation pipeline with deterministic quality-gating and cost-tracking infrastructure whose current logs capture over ten thousand individually priced model calls across fifteen distinct models from five providers; Year 1 extends this instrumentation with per-step outcome logging to build the derailment-hazard corpus the first study analyses.