Glossary

A small vocabulary with load-bearing meanings runs through these docs. This page collects it in one place: the platform terms that deserve exact definitions, and the workflow jargon that only the decision engine chapters introduce. If you add a term elsewhere in these docs, define it here too.

The terms

• Agent — in workflow vocabulary, one of your participants: a person, a service, or a model, registered by name. See the decision engine.
• Alignment map — the per-competency record of who co-sides with whom on the rounds where review panels split: similarity scores, the strict/lenient axis orientation, evidence counts, and cross-competency coalition correlations. Seat review panels whose members don't share one underlying bias. See the decision engine.
• Ask mode — the one-off question surface: answers against the project's memory, read-only; nothing written back, nothing stored in any conversation. The only generative capability available to viewers. See chat features.
• Assertion map — the default response map: every answer asserts "yes" or "no". Contrast the vector map. See response maps in the decision math.
• Attestation ledger — the one append-only record of review outcomes that the competency matrix measures through; its reference point is your 0–100 dial. See workflows: overview.
• Competency — one skill, named by you, e.g. "editorial" or "writing". Each agent carries one competency score per competency name, updated with every review outcome. See the decision engine.
• Competency interpolation — estimating a reviewer's score on a competency from their history on related competencies, when they have no direct history on it. Default on, opt-out per family. See competency interpolation.
• Competency matrix — the agents × competencies grid the decision engine maintains: one score per cell, each with its own history and trend. See tested results.
• Consensus mode — the dial at 0: no trainers. Confidence gates on whether reviewers converge, and how decisively, and the alignment map records who takes which side. Measures belief structure, never correctness. See workflows: overview.
• Contribution bar — the per-response breakdown of how much each perspective weighted the answer. See chat features.
• Customer account — the top-level account, addressed in the API as customer_uuid, that owns projects. See projects and teams.
• Dial — the 0–100 control over how much external reference the attestation ledger gets. Its position decides what competency scores mean: peer review alone can only measure belief structure; trainer benchmarks turn the same scores into calibrated correctness. Most deployments run the fluid mix between the ends. See workflows: overview.
• Entry (entity, place, thing) — a record in the memory. Entries come in three types — entities (people, organizations), places, and things — and carry a name, hint, characteristics, relationships, status, and events. See memory.
• Event — an immutable, hash-chained record that something happened: topic, status, prose summary, structured payload, related entries. It distinguishes when it happened (occurred at) from when Autopoetic learned it (recorded at). See memory: events.
• Family — in review workflows, a parent work unit plus its review children. Review rounds repeat inside the family until the review ladder resolves or halts; the response map is declared at family creation and the family is homogeneous from then on. See the review ladder.
• Fork — three different things: a conversation fork (a new conversation containing a copy of everything up to a chosen response), a project fork (copies the memory, agent templates, and each agent's personality; conversations, tasks, and API keys are not copied, and the fork gets fresh encryption keys), and an agent fork (a decision-engine agent inheriting the fork parent's track record as of the fork point). See chat features, projects and teams, and the decision engine.
• Generation — a numbered snapshot of the memory. Every write produces the next generation, every generation is retained (so the graph's time slider can scrub the full history), and writes built on a stale generation are rejected as conflicts rather than silently overwriting. See memory.
• Ground truth mode — the dial at 100: trainer answers are ground truth — infallible, pinning verdicts at confidence 1.0 — and competency scores calibrate to correctness. See the decision math → trainers.
• The Humanist — Autopoetic's model and the default chat agent: a subjective Mixture of Perspectives whose weights are computed per message and accumulate across the project. See what is Autopoetic?.
• Instrument — a vector response map plus its optional reviewer-facing prose (guidance map-wide, per-label definitions, per-scale-position anchors), pinned at family creation: the text rides the family's commit verbatim and keys reviewers' credit records, so an edited instrument is a new instrument whose history starts fresh. See instrument prose.
• Memory — the structured, queryable knowledge each project grows: entries, relationships, and events, built by one memory pass per turn and fully scriptable over the API and MCP. Encrypted at rest with per-project keys. See memory.
• Memory access — the per-conversation setting controlling that conversation's access to the project's memory: Full (read and learn), Read-only, or None. See chat features.
• Memory graph — the interactive explorer on the project page: entries as nodes, relationships as labeled edges, with semantic neighbors, a time slider, and deep-linked focused views. See memory → the memory graph.
• Memory pass — the single per-turn learning cycle: search the memory for what's relevant, merge in new and changed facts, commit, then compose the response against the already-updated memory. The "Reading memory…" and "Actively learning…" statuses are this pass at work. See memory → how it's built.
• Perspective — one internal viewpoint that evaluates every message in parallel before the response is woven. The Humanist thinks with four: Empiricist, Empath, Speculator, Cynic — woven by Identity. See what is Autopoetic?.
• Producer / reviewer — the two roles on a review round. The producer submits the work; reviewers claim review slots and answer. Reviewers' scores move with how their verdicts align with the outcome; producers' scores move with whether their work is trusted. See review workflows.
• Project — the strict data container: conversations, the memory, teammates, agent templates, workflow templates, and API keys all live inside exactly one project. Everything, including the encryption keys, is scoped per project. See projects and teams.
• Review ladder — the four verdicts after each round: CONTINUE (keep collecting), RESOLVE (confidence ≥ minimum_confidence), ADD_REVIEW (one more round), HALT (cap reached, unresolved). The same strings the answers endpoint returns as signal, lowercased as family_signal in run state. See the review ladder.
• Review slot — one seat on a work unit's review panel, claimed atomically: two concurrent takers cannot both claim, the loser gets a clean conflict error. See the decision engine → work units.
• Run — one execution of a template. Runs wait — for seconds or days, at zero cost and with no deadline — until external agents call back. See runs.
• Template — the ordered steps the engine runs: calls to your services, waits, branches, loops, and review clauses. See writing templates.
• Trainer — an agent marked infallible: its answers are your answer key, winning reviews outright and driving fast learning. Trainers are the dial's control. See the decision math → trainers.
• Vector map — a response map where answers are labeled unit-scale numbers, one per declared dimension ("quality", "safety", "tone"), instead of yes/no — optionally with an acceptance predicate and reviewer credit rules. See response maps in the decision math.
• Work unit — one piece of produced work plus its review slots, declaring minimum_reviews, minimum_confidence, maximum_reviews, the competencies, and the response map. Work units created with zero reviews are recorded and auto-approved (provenance, not review). See the decision engine → work units.