4. Where judgement belongs in a machine system
A craft question he treats as philosophical. Standard agent orchestration puts a language model in charge; he thinks that's wrong because judgement is expensive, unreliable and hard to inspect. Instead: a deterministic program that summons a model only where a plan has broken and a decision is needed. His design question: "which nodes need to be agentic at all." Built this way: mgr.
Same instinct elsewhere: astrology-product computes the chart mechanically, maps configurations to a fixed authored corpus, and lets the model write only from that ("constrained improvisation"); this research wiki carries provenance, dispositions and written rules for changing them (research-values).
Principle in his words: an agent system offers "predicate satisfaction" (output can be verified against a test), never reproducibility. So annotate every boundary by contract strength and keep the unverifiable parts small. See harness-idiom.
Open: a general account of where the boundary goes. See thread-under-the-threads.
Live: one of his open threads. Re-check with James before relying on it.
<!-- exitmusic-2026-10-26 -->
From his notes
Two of his published pieces bear on this thread: ai-risk-as-runaway-it (accountability delegated to systems that can't be questioned; failures that can't be diagnosed) and dark-discourse (a design grammar with no words for harms to non-users).
Source: James's notes at exitmusic.world (read 2026-10-26); his own position, disposition stated.
Provenance: james-open-threads.md, uploaded by James 2026-10-11, assembled from his own descriptions, autumn 2026 (unverified). Disposition: live.
Related: harness-idiom, mgr, astrology-product, research-values, thread-under-the-threads