KIP 2.0

English | 中文

This glossary describes the current KIP 2.0 draft. The Specification, Cognitive Consistency contract, Cognitive Memory Profile, and Memory Interface define the contracts. Historical terminology is retained in v1/.

Core state and trust

Concept: A semantic entity in a MemorySpace. Types and fields come from versioned Schema Packages.

Proposition: A truth-neutral semantic statement. Its existence alone establishes neither truth nor acceptance.

Assertion: An actor's stance toward a Proposition, with mode, confidence, valid time and supporting Evidence. Assertions can compete without overwriting the historical record.

Evidence / Activity: Evidence records observations; Activity records how cognitive artifacts were generated or changed. Repeated derivation from one source does not create independent corroboration. Outcome Evidence comes from instrumentation or review; an actor's own account remains an agent_statement.

Belief / Epistemic Projection: The accepted view computed over eligible Assertions, conflicts, evidence and dependencies at a query's scope and time. Belief is not a mutable truth flag on a Proposition. A cached projection needs a validated computation basis.

Governance: The protected authority controlling access and permitted operations. Evidence, confidence, imported signatures, memory usefulness and a fired Watch never grant authority.

Experience and memory

Event: A time-bounded occurrence recording what happened. It may supply evidence for an Experience without itself encoding a goal-directed trajectory.

Experience / ExperienceStep: A subject's goal-directed trajectory and its ordered observation, decision, action and feedback records. Step order is structural; a caused_by claim is a semantic Proposition plus Assertion supported by Evidence. Earlier does not imply causal.

Memory: Past state participating in future retrieval, prediction, decisions or actions. Storage alone does not establish useful memory.

Knowledge / Insight: Knowledge names durable regularities warranted by evidence rather than a universal Core kind. Insight is a declarative lesson derived from experience or evidence. An Insight is not an executable Skill or a verified learning result.

Action Briefing: Recall organized around a future action: relevant goals, constraints, evidence, analogous Experiences, Skills, risks and commitments. Retrieval alone does not establish that a Skill was used.

Confidence: Epistemic support carried by an Assertion. Changes should reflect evidence and reasoning, independently of recall frequency.

MnemonicState: A Profile Facet separating memory_strength (future accessibility), salience (importance/noteworthiness), and utility (expected decision value). These are not truth probabilities or permissions. KIP 2.0 uses this Facet rather than the v1 metadata.memory_strength field.

Procedures and learning

Skill / SkillRevision: Skill is the stable procedural identity. SkillRevision freezes its behavior, task family, applicability, procedure, success criteria and recovery under a behavior digest. Decisions, trials, grading and procedural authority refer to the exact revision.

Attempt: One application under recorded conditions, assigned to a trial before its outcome is observed. Multiple observations of an attempt do not create additional independent attempts.

DecisionRecord: The immutable record distinguishing what was retrieved from what actually influenced a decision, including the exact Skill revision and its attempt linkage.

TrialRecord / EvaluationRecord: The immutable comparison basis and replayable verdict retaining exact rules, inputs and artifacts. Sharing a task family selects candidates; it does not establish attribution or baseline membership. TrialState and GradingState are current caches of those records.

Skill lifecycle: proposed → trialed → adopted → revoked, governed by validated evaluations and authorized policy. Only comparative evidence promotes trialed behavior; imported Skills begin unproven. Descriptive feedback alone is not verified improvement.

Dependency validity: Derived cognition retains its provenance and computation basis. Changes make affected derivations reviewable; reuse requires checking their basis. A stale stored summary cannot prove its own validity.

Interfaces and durable work

Memory Interface: The business Agent's five intents: observe, recall, revise, feedback and forget. A Brain Adapter interprets them through existing KIP state operations. This optional interface neither adds a Core kind nor requires another LLM.

Processing receipt: Tracks intake from durable recording through processing disposition to recall availability. An after barrier accounts for specified inputs; a fresh Space snapshot alone does not prove processing completion. It is distinct from a transaction receipt.

Watch / WorkingState: Watch is durable attention state evaluated against changes or a silence deadline. Firing creates attention, not permission to act. WorkingState summarizes resumable cognition with its basis_seq so a Brain can resume from validated state plus a delta.

Cognitive Capsule: A portable state artifact with declared dependencies and integrity. A signature proves origin and integrity, never truth or destination authority. Import is a destination-governed transaction.

Protocol, reliability and learning evidence: Language tests, contract models, engine conformance, fault/replay tests and behavioral evaluations answer different questions. Passing structural checks does not prove that a real Brain learned; see Brain Evaluation.