Skip to content

Latest commit

 

History

History
15 lines (13 loc) · 2.33 KB

File metadata and controls

15 lines (13 loc) · 2.33 KB

Glossary

Terms the documentation glosses on first mention, collected in one place for quick lookup.

  • System 1 decision model: A fast classifier that reads environment state and returns probabilities over caller-enumerated options in one forward pass without generating freeform text. docs/decision-models.md
  • Decision model: The abstract interface (DecisionModel) standardizing how fronts query backends (jev, laya, cua, baselines) for typed decisions. docs/decision-models.md
  • Model slot: The constructor argument (model) of an openJiuwen agent where a decision model or chat model is plugged in to take actions on decision turns. docs/architecture.md
  • Front (tool, browser, rail): An agent interaction layer (ToolDecisionModel, BrowserDecisionModel, DecisionModelRail) that extracts state, presents candidate options to a decision model, and executes actions. docs/architecture.md
  • Rail: A policy or safety check that evaluates state at lifecycle hooks to enforce constraints, block loops, or fail closed. docs/architecture.md
  • Head (choice, noul, score): A specialized prediction head of a decision model; choice selects an offered key, noul estimates proposition probability, and score ranks state against a rubric. docs/decision-models.md
  • Arm: An experimental configuration or driver pipeline tested on identical benchmark tasks to measure comparative latency, cost, and accuracy. docs/benchmarks.md
  • Tick: A single decision turn in an agent episode recording the chosen key, probabilities, confidence, latency, tokens, and active state flags. evals/README.md
  • Replay: A visual or step-by-step reconstruction of an evaluation run, rendering states, chosen actions, and model probabilities side by side to HTML or GIF. evals/README.md
  • Episode: A single complete run of an agent through a task, environment, or game from start to terminal state or timeout. evals/README.md
  • Labelled set: A dataset of pre-recorded environment states and ground-truth options in JSONL format used to benchmark precision, recall, and accuracy without live execution. evals/README.md