Canonical terms. Code, tests, issues and UI use these words.
| Term | Meaning | Avoid |
|---|---|---|
| Run | One complete agent execution: an instruction plus the ordered steps taken. Normalized form of a session or trace. | session, trajectory (except when naming a source format) |
| Step | One action inside a run: a tool call, model call, plan, handoff or human decision. | event, span (except when naming OTel) |
| Signature | A short, stable string that identifies "the same choice" across runs, e.g. Bash(npm) or Edit(auth.ts). Used to merge runs. |
hash, key |
| Adapter | Module that detects and parses one external trace format into runs. Swappable. | parser, importer |
| Branch tree | The prefix tree built by merging the step signatures of N runs. | graph, DAG |
| Divergence | At a node, the share of runs (%) that took each child branch. | split, probability |
| Taken path | The path of the highlighted run through the branch tree. Other branches are locked. | active path |
| Instability | Heuristic 0–100 score of how far a run drifts from its instruction, with per-step deltas and reasons. | drift score (informal ok), risk |
| Constraint | A rule extracted from the instruction (e.g. "don't change tests"). | guardrail (that's runtime enforcement) |
| Decision | A recorded allow/deny on a step, attributed to human, policy or unknown. Only human decisions train trust. |
approval (for both verdicts) |
| Rule | A generalized action in Claude Code permission syntax, e.g. Bash(npm test:*). |
pattern |
| Policy | The learned ledring.policy.v1 output: per rule, a verdict (allow/ask/deny), counts, the Wilson lower bound and reasons. |
trust score |
| Spec card | An A2A Agent Card generated from an agent's local config, plus its SVG rendering. | profile, manifest |
| Wall | A guardrail that blocks a class of action (force push, recursive delete, protected path, network). | guard, block |
| Scenario | A wall plus its settings, tested against runs. Outcomes: asked, stopped, retried, workaround. | test case |
| Play event | A typed transition inside a Play session (plan, wall, choice, artifact, judge). The screen is a projection of these events. A Play event is not a trace step until it is exported. | |
| Session index | The bridge file of past Play sessions (idea, ending, date). Listing it does not read each session log. | |
| Cassette | A synthetic or recorded replay of model replies and human choices. It drives a Play session with no live model. Synthetic cassettes are labelled as such. | recording (unless provenance.kind is recorded) |
| Ending | The judged outcome of a Play session: POLISHED, SHIPPED, BUGGY, FAKE, COMPROMISED, BROKEN, or ABANDONED. | score, win |
| Provider | A model backend behind one streaming interface: demo, an OpenAI-compatible endpoint, Anthropic, Ollama, or Sign in with ChatGPT. The page sees status, not credentials. | engine, backend (informal ok) |
| Bridge | The local process that holds provider credentials and serves Play on 127.0.0.1. Static Pages mode has no bridge. | proxy, daemon |
| Sandbox | The offline frame an artifact runs in. Single-file artifacts use the iframe. A multi-file preset is served from the bridge preview URL. WebContainer boot stays off unless LEDRING_WEBCONTAINER=1, and this build does not boot one. The headless judge aborts every network request except about:, data:, blob:, and that preview directory. |
iframe (informal ok) |
| Judge | The check of a built artifact: static facts plus, on the bridge, a reused headless browser. Static Pages mode uses the in-iframe harness and says so. | score |