Answers are grounded in this paper and cite their sources.
Autoregressive frontier models optimize conditional sequence likelihood over unstructured token histories, creating a fundamental architectural impedance mismatch when deployed inside production software systems governed by strict state invariants and cross-service contracts. Under partial observability or under-specified requirements, unconstrained coding agents exhibit three characteristic failure modes: epistemic interpolation (silently fabricating unstated business rules rather than halting), context-entropy drift (progressive dilution of specification constraints by verbose tool traces), and cascading regression loops (local symptom patching that violates upstream or downstream invariants). This working paper formalizes a Zero-Trust Agentic Synthesis Protocol designed to constrain stochastic language models within deterministic engineering boundaries: Asymmetric Context Ingestion: Pairing 100% full-context ingestion of non-compositional design specifications with coordinate-indexed surgical scope isolation for source code to maximize constraint signal density. Epistemic Halting Automaton: Modeled on Dijkstra's guarded commands, treating specification ambiguity as a hard non-mutating halting state rather than an invitation to approximate. Forced Three-Stage Artifact Pipeline & Orchestrator–Verifier Topologies: Materializing grounded symbol inventories, set-theoretic gap analyses, and placeholder-free execution schedules executed across disjoint repository worktrees with read-only adversarial verifiers and compute-optimal model allocation.
No takes yet. Share an insight, caveat, or question.
Van Le Nhat (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: