Terminology - Pome

Every coined term in these docs, defined once. Linked pages go deeper.

Term	Definition
Digital twin	A deterministic, stateful emulation of one real SaaS API (GitHub, Slack, Stripe, Gmail, Linear). Same request/response shapes and error semantics as the live service, no production side effects. See How Pome works.
Task	The exam: a markdown file with three required sections — `## Config`, `## Prompt`, `## Success Criteria` — and an optional `## Seed State`. See Tasks.
Seed state	JSON the digital twin loads before the agent starts — the initial world (repos, channels, charges).
Examinee	The clone of the agent under test. It works the task against the twins and never sees the grading criteria.
Coach	The agent driving the test — your own coding agent, working over the Pome MCP. It installs, launches runs, finalizes, and reads reports. The coach can see criteria; the examinee cannot.
Grader	The deterministic engine behind a `[code]` criterion — SQL over the digital twin’s final state and its tape. It is the only thing that scores, and every verdict it returns ships with the evidence pointer behind it.
Narrator	The LLM. It writes each run’s summary and handoff, and an advisory reading of every `[model]` criterion. It holds no score authority: nothing it writes moves a number.
Judge	The LLM scoring a `[model]` criterion rather than reading it. Beta, off by default, chosen per agent (`evaluator_id`, which defaults to `narrator`). A `[code]` criterion has a grader and never a judge.
Run	One execution of a task by one agent: boot the twin with the seed, record every call, score the result. Artifacts land in `runs/<task>/<run-id>/` and on the dashboard.
Sandbox	What you start: one id, one TTL, one price, and the 1–3 digital twins a run talks to. Defined in full on Sandboxes. It expires; finalize the run while the sandbox is still live, or it cannot be scored. The wire spells it `/v1/sessions`, `session_id`, and the `ses_` id prefix.
Finalize	The call (`finalize_run`) that closes a run and triggers scoring.
Criteria	A task’s acceptance checks, and what owns the verdict. `[code]` criteria are deterministic checks on twin state or events, and they are what the score is out of; `[model]` criteria are plain-English judgment calls the narrator reads and answers in prose.
Score	0–100 satisfaction across the `[code]` criteria that were evaluated — `[model]` readings are advisory and never enter the denominator. A run passes when the score meets the task’s `passThreshold` (default 100). A task with no `[code]` criteria reports narrative only — 0 scored criteria instead of a percentage.
Trace	The recorded stream of everything the examinee did (`events.jsonl`): twin HTTP calls, LLM calls, run lifecycle events.
Surface	A place you drive the loop from: the CLI, the coach skills, the MCP, or the dashboard. Same loop and the same run row, wherever you start it.
`pome.json`	The manifest that connects a repo to Pome: the agent’s slug, its start command, and the twins it uses. Written by `pome init`, read by every run. See Your own agent.