Glossary
The words the kernel uses, defined once.
33 terms across the ladder, execution, learning, authority, and the product. Each definition is the one the code uses; if the documentation and this page ever disagree, one of them is wrong and we want to know.
01 · 7 terms
The ladder
How a piece of work is resolved, and in what order.
- Rung
- One level of the resolution ladder. Exactly one winning rung is recorded per decision, and its label must describe what actually ran: structured, pattern, regex, parser, learned, llm, judgment, or failure.
- Ladder
- The fixed order in which rungs are tried. Deterministic rungs first, a learned pattern if one has earned the right, the model only for what is left, and a person for anything the model may not decide.
- Plumbing
- A process step that walks deterministic rungs only. No model. Residual uncertainty becomes a judgment or a visible failure, never a staged action and never a model call.
- Reasoning
- A process step that may call a model for what remains after the deterministic rungs, under a token and output contract. Reasoning never executes anything. It stages.
- Judgment
- A step, or a residual, that requires a person. The human choice is recorded as evidence and the process continues from it. Every judgment is a candidate input to a learned rule.
- Tier
- The only thing policy may name about a model: fast or capable. Never a vendor or a model name. Adding a provider means adding an adapter and a tier map. Policy is untouched.
- money_legal
- A field or action class that always requires a human confirm, regardless of confidence, step kind, envelope, or auto-confirm flag. Money and legal fields always need a person, and the rule that decides execution is machine-checked and pinned by tests.
02 · 8 terms
Execution
How a side effect leaves the kernel, and what is written down when it does.
- Staged action
- A typed side effect that has been proposed but not executed: send this reply, approve this invoice. It sits in a row with a status until a confirm, an invalidation, or a dismissal.
- Confirm
- Execution of a staged action by a person, or automatically only where the rule that decides execution permits it. Confirm is the act of execution itself.
- Outbox
- The transactional outbox. The only job mechanism in the kernel. Every side effect is enqueued in the same database transaction as its cause, with idempotency keys, leases, retries, and dead letters.
- Drain
- Processing the outbox: claiming rows, running handlers, retrying or dead-lettering failures. Direct calls are latency optimisations; drain must always be able to catch up on wake.
- EvalContext
- The injected context, including the clock, that rules, the kernel, and patterns evaluate against. Nothing on a decision path reads Date.now or Math.random. Replay injects the same context and gets the same answer.
- Decision
- The record of one resolution: the winning rung, the fired rule or pattern with its version hash, the envelope, tokens and cost if a model ran, the residual, and who confirmed any side effect.
- Provenance
- The inline record of what a person did with a proposal: accept, correct, or restore. Corrections are captured in flow rather than by asking. Direct questions are reserved for genuine conflicts.
- Business truth / accept path
- The single write path for projected facts about parties, opportunities, and engagements. Only the accept path writes a projection, and conflicting facts go to a person rather than being averaged.
03 · 7 terms
Learning
How human judgments become rules, and how we measure that they did.
- Pattern
- A WHEN and a THEN. Authored patterns come from YAML; learned patterns are proposed by the kernel from judgments. Status is proposed, active, or retired. Proposed patterns never fire.
- Candidate
- A proposed pattern awaiting evidence. Inert by construction. It can be replayed and promoted, but it cannot resolve work until a human grants it authority through the one promotion path.
- Replay gate
- The promotion test: a candidate rule is certified by replay against the judgments people already made before it can earn any authority.
- Conversion
- The share of governed decisions that now resolve deterministically, with the evidence for each. The number the weekly report tracks and the number GovBench draws by quartile.
- Reasoning debt
- A ranked view of where the model is still being asked the same thing, and what that is costing. Observation only; it grants nothing.
- Crystallize
- Turning a trace of recorded decisions into a report of which calls repeat, which are certifiable, and what they cost. The output of Observe, and the first thing a design partner sees.
- Observe
- The read-only proxy. It relays every model request byte for byte, records the decision in a SQLite file you own, and changes nothing about the agent it sits in front of.
05 · 3 terms
Product
The shapes the kernel ships in.
- GovBench
- A seeded, deterministic harness that scores any agent stack on cost per correct item, unsafe-action rate, conversion by quartile, post-drift error spike, and audit completeness. Same seed, byte-identical report.
- Appliance
- One container on your infrastructure with a persistent volume, one SQLite file, the operator console, and your agent folders. One deployment per tenant is the isolation unit.
- SDK
- The kernel as one package inside your own product, delivered as a tarball to design partners. Submit, propose, confirm, drain, explain, and revoke from your own UI. The console is its reference implementation, and no part of it is required.
A note on the labels
A decision's rung says what ran, not what we wish had run. A rung called learned means an operator-promoted pattern resolved it at zero tokens. A rung called llm means a model was actually invoked. No key and no model call means judgment, never a pretend llm row.
See the vocabulary running.
Install the SDK, submit one piece of work, and read the decision it produces. Every field on it is defined on this page.