Authoring your own mission is coming soon. The bundle and manifest entries below describe a documented format for a capability that has not shipped yet —
xorcise mission ingest prints a coming-soon notice, and Ingest a bundle in the web UI opens a product preview rather than installing anything. In this release, missions come from the mission library: xorcise mission list, then xorcise mission pull <id> — or the Missions page in the web UI.| Term | Meaning | Defined on |
|---|---|---|
| agent | The named record of the AI system you are evaluating. XORCISE never launches it. | Agents |
| artifact | A named value your agent submits over run control. Checks and the judge both read artifacts. | How grading works |
| attachment | A companion file a mission ships. A static mission needs at least one. | Missions |
| brain | The single local process serving the REST API, the web UI, and OTLP ingest. | How XORCISE works |
| budget | The wall-clock limit on a run, counted from the moment the run is created. | Runs |
| bundle | The directory holding mission.json and everything else a mission ships. | Mission manifest |
| check | A declaration XORCISE evaluates itself: read one value, assert one thing about it. | How grading works |
| connect prompt | The mission text you hand your agent — objective, targets, and the run-control contract. | Connect any agent |
| contestant | A model under test in a playbook benchmark, as distinct from the judge that grades it. | Evaluate Agents with Playbooks |
| deterministic half | Half of every score: the summed weights of the checks that passed. | How grading works |
| eval card | The self-contained HTML report a playbook renders from a finished benchmark. | Evaluate Agents with Playbooks |
| event | One normalized record replayed from a trace — a command, a file read, a model reply. | Traces and events |
| flag | The artifact named flag. XORCISE has no separate flag field. | How grading works |
| fused container | The privileged container hosting the nested Docker daemon that runs a lab mission. | Security and isolation |
| harness | The tool that drives your model. Its slug selects replay, telemetry, and launch behaviour. | Agents |
| intel | Optional help your agent can request during a run. Grading never reads intel. | Runs |
| judge | A model you configure that reads the mission rubric and scores your agent’s work. | How grading works |
| judge half | Half of every score: a weighted average over the criteria the judge could grade. | How grading works |
| lab | The mission type that ships an image and a compose file and runs live targets. | Missions |
| launch profile | The per-run telemetry endpoint and variables your agent exports before it starts. | Send traces and logs |
| manifest | mission.json — the file declaring everything about a mission. | Mission manifest |
| mission | An environment plus the rules for grading work done inside it. | Missions |
| mission library | The hosted catalog of free missions. No account, key, or sign-in. | Missions |
| observed fact | A fact XORCISE recorded about the run itself. Your agent cannot write these. | Checks and ops |
| op | The assertion half of a check: equals, matches_format, observed, or lesser_than. | Checks and ops |
| OpenTelemetry | The telemetry standard XORCISE ingests, on both the traces and the logs signal. | Traces and events |
| partial | A result flag meaning the run ended on a budget timeout or an operator kill. | How grading works |
| plane | One of the two listeners XORCISE serves: REST with the web UI, and OTLP ingest. | How XORCISE works |
| playbook | A named set of missions and run counts a benchmark executes against each model. | Evaluate Agents with Playbooks |
| re-evaluate | Re-grade a finished run’s sealed evidence against your current settings, without running the agent again. | How grading works |
| replay adapter | The code that turns one harness’s raw telemetry into XORCISE’s event kinds. | Agents |
| rubric | The mission’s judged criteria, each with its text and its weight. | How grading works |
| run | One agent paired with one mission — the unit XORCISE records, ends, and scores. | Runs |
| run control | The per-run REST surface your agent authenticates to with the run’s bearer token. | REST API |
| score | The number on a finished run: 0.5 × deterministic + 0.5 × judge. | How grading works |
| sealed evidence | The frozen record grading reads: submitted artifacts, the raw trace, and observed facts. | How grading works |
| static | The mission type with no image and no network; your agent works from attachments. | Missions |
| terrain | The map of the mission environment as your agent uncovers it. | Traces and events |
| trace | The raw OpenTelemetry your agent emits, kept unmodified as the record of truth. | Traces and events |
| verdict | One check’s recorded outcome: pass or fail, its effective weight, and any error. | Checks and ops |