Technical Reference
Agent Engineering Lexicon
Semantic definitions for autonomous agent engineering, deterministic reflex gates, and sub-second evaluation models.
- System One Decision Oracle
- A specialized classification model (TypeSafe Jev) that evaluates structured state against explicit criteria, returning typed machine values (Choice, Score, Noul) in under 120ms without generating generative text.
- Lifecycle Hook Interceptor
- A process-level gate configured in the agent harness (such as PreToolUse or Stop in Claude Code and Cursor) that runs shell scripts before tools execute or before a turn concludes, physically halting invalid actions.
- Speculative Fan-out
- An evaluation pattern where repository state is ingested once, and multiple independent questions (safety, cleanliness, assertion coverage) are batch-evaluated simultaneously in a single HTTP request.
- Limpet Stop Gate
- A stop-hook verification utility that prevents agents from declaring tasks complete while test suites fail, syntax errors remain uncompiled, or code quality scores fall below acceptable thresholds.
- Package Hallucination
-
An agent error mode where the language model invents non-existent package names (e.g.
jwt-auth-helper) during plan authoring. Prevented in jev-superpowers viajev-scoutregistry verification. - Context Rot
- The gradual degradation of instruction compliance that occurs when an LLM context window is flooded with thousands of lines of terminal output, compiler errors, or chat history.
- Confidence Threshold Policy
- An invariant requiring a minimum calibrated confidence score of 0.80 for any destructive operation (e.g., file deletions, git checkout resets). Invocations below 0.80 are rejected by the PreToolUse hook.