MCP tools
Eighteen tools. Read tools are marked read-only for the client. Write and governance tools ask for confirmation in every supported client.
Reference
| Tool | What it does | Role |
|---|---|---|
| get_workspace_context | Workspace and ontology status. Call it first in a session. | member |
| resolve_decision_state | The decision in force for a question, with scope, time, sources, supersessions and conflicts. | member |
| get_decision_history | Timeline of a topic: approvals, supersessions, revocations and the current heads. An empty timeline says why. | member |
| find_conflicts | Incompatible claims, their sources and scopes, and what evidence would resolve them. An unrecognised topic filter says so. | member |
| ingest | Capture text or reviewed records with source metadata. Approved records require a governor under the organisation’s mode and verified evidence. | member (write) |
| get_ingestion_status | The persisted result of an ingestion job: records stored, skipped or queued. | member |
| list_pending_decisions | Candidates and conflicts awaiting review, with their reasons. | member |
| review_decision | Approve or reject a candidate, revoke an approved decision or retract a proposal, with a reason. | reviewed: admin/owner; peer: member |
| resolve_conflict | Resolve an open conflict with an approved winner, or dismiss it, with a reason. | reviewed: admin/owner; peer: member |
| start_ontology | Propose a small ontology (topics, scopes, source types, authority rules) from a description. | member (write) |
| get_ontology | The active ontology and its version history. | member |
| rate_answer | Rate a captured resolver answer by call_id; capture must be enabled. | member (write) |
| edit_ontology | Propose a change, approve a proposal or prepare a rollback. Approval is always a separate call. | approve: admin or owner |
| prepare_eval_dataset | Store a draft set of decision questions with the answers a person expects. Immutable once stored; preparing approves nothing. | reviewed: admin/owner; peer: member |
| review_eval_dataset | Approve the inspected version of a dataset, or retire it, with a reason. A retired dataset stays readable and admits no new run. | reviewed: admin/owner; peer: member |
| run_eval | Measure the resolver against an approved dataset and store one completed run with its summary. | reviewed: admin/owner; peer: member |
| get_intelligence | Datasets, their cases and review state, and the per-case expected/observed detail of a run. | member |
| export_eval_dataset | An approved dataset as JSONL: questions, expectations and references, declared evaluation_only. Never source text. | owner |
resolve_decision_state accepts an optional topic and scope, and a request_id used as an idempotency key bound to the question asked: repeating a call with the same request_id and question returns the same answer.
Measuring the resolver
The last five tools are an evaluation instrument for resolve_decision_state, not a training pipeline. The cycle is prepare, inspect, approve, run, read, export: a person writes the questions and the answers they expect, approves the exact version they read, and a run resolves every case against the current memory and records what each check found. Preparation never asks the resolver for an expected answer, and approving a dataset changes no decision and no ontology.
- Five checks per case: the epistemic state, the selected decision and its value, the scope, the sources the case requires, and whether the excerpts returned resolve to sources the answer also lists.
- A check with nothing to measure is reported as not applicable, never counted as a pass or a failure.
- A dataset is immutable once prepared and holds at most 20 cases; correcting it means preparing a new one and retiring the old.
- A run records the dataset hash, the scorer version and the ontology version. Re-running after an ingestion may legitimately give a different result: it measures the memory as it is now.
- The JSONL export is declared evaluation_only and carries questions, expectations and references — never source text, ratings or trace payloads. It is not a training dataset and nothing here claims it is fit for one.
- A passing run says the resolver returned what a person expected on those cases. It is not a measure of whether an assistant then respects that answer.
What an answer contains
- The state: verified (shown as Current in the web app), contested, insufficient_evidence or unknown.
- The decision in force, its scope and the date from which it applies.
- Who approved each decision and when; OIDA Cloud adds the approver’s name.
- Superseded decisions, listed as history.
- Open conflicts involving the topic.
- Evidence excerpts and source references, never full documents.
- Candidate topics when the question matched nothing exactly.