Glossary

Plain-language definitions for terms used in the manuscript, kept consistent with real (2026) engineering practice. Use this to check that dialogue and narration use terms correctly — characters may explain these informally, but the underlying concept should never be fudged for drama.

• Agent / agent loop — a system in which a model plans, calls tools, observes results, and re-plans iteratively until it judges a task complete or hits a limit. Called “the Loop” internally at Meridian.
• Autonomous multi-agent system — an architecture where multiple agent loops coordinate (as sub-agents, peers, or a supervisor/worker hierarchy) on a shared task. Increases capability and increases the surface area for compounding failure.
• Circuit breaker — a deterministic, non-model-mediated limit (step count, cost ceiling, tool-call rate, wall-clock time) that halts a runaway process and escalates rather than trusting the system’s own judgment of completion.
• Context window / context saturation — the fixed span of tokens a model can attend to at once. Saturation is the degradation that occurs as that window fills with low-signal, stale, or contradictory content, distinct from simply “running out of space.”
• Context drift — behavioral degradation that emerges gradually as accumulated context pulls the model’s effective behavior away from its original instructions, without any single edit being at fault.
• Deterministic scaffolding — the umbrella term (and thesis of the book) for engineering structure — schemas, evals, circuit breakers, scoped permissions, human review gates — wrapped around a probabilistic model so the system’s behavior is bounded even though the model’s own output is not.
• Eval / synthetic continuous evaluation — an automated, versioned test suite of task scenarios (ordinary and adversarial) run against a model or prompt version to produce a quantitative pass rate, used the way a regression suite is used for conventional software.
• Hallucination — fluent, plausible-sounding output that is factually or contextually false, distinct from a model expressing uncertainty; dangerous specifically when acted upon by a downstream tool call.
• Human-in-the-loop exception cache — a queue that routes low-confidence or out-of-distribution cases to a human reviewer instead of letting the system guess, treated in the book as an instrumentation problem, not a staffing problem.
• Least-privilege tool access — scoping a tool grant to the minimum required for a specific task type, rather than granting broad standing access “in case it’s needed.”
• Prompt injection — adversarial or incidental instructions embedded in data the model processes (an email body, a wiki page, a ticket) that the model interprets as legitimate operator instructions rather than untrusted content.
• Prompt patch — an ad hoc edit to a system prompt intended to fix a specific observed failure, usually made without a corresponding regression test, and therefore carrying an unmeasured risk of breaking other behavior.
• Queueing theory — the branch of applied mathematics studying systems where tasks (or requests) arrive, wait, and are served; Victor’s preferred lens for reasoning about agent workloads, retries, and cascading backlogs.
• Regression benchmark — a fixed, repeatable test suite run against every change to catch newly introduced failures, standard in conventional software and, per the book’s thesis, systematically missing from most production LLM systems circa the book’s setting.
• Schema-validated tool interface — a tool-calling contract in which every request from the model is checked against a strict, machine-enforced schema before execution, so malformed or out-of-policy calls are deterministically rejected rather than executed on trust.
• Single point of failure (SPOF) — in this book, specifically the organizational pattern of one engineer (Devon) holding undocumented, irreproducible knowledge required to operate a critical system.
• Stochastic variance / error tail — the fact that a probabilistic system’s outputs form a distribution with a nonzero-probability “tail” of bad outcomes, which cannot be eliminated but can be bounded, measured, and defended against — the crux of Claire’s Act I category error (treating the system as having deterministic uptime rather than a measurable failure rate).
• Tool use / tool call / function calling — the mechanism by which a model requests execution of an external action (query a database, send an email, run code) rather than only producing text.