Glossary

Key terms used across the ADRs, the cycle documentation, and the companion paper. Definitions here are navigation summaries; the ADRs are canonical — when a definition and an ADR diverge, the ADR wins. For the English ↔ Japanese rendering table, see translation-glossary.md.

Several terms are shared with the sibling repository agent-attribution-practice (AAP), whose glossary treats the accountability side of the same vocabulary. Cross-references are marked inline; the two glossaries are written to be consistent.

Agent Knowledge Cycle (AKC)

Six composable phases — Research, Extract, Curate, Promote, Measure, Maintain — by which an agent's experience becomes durable, behavior-shaping knowledge, with human approval required on every change to the agent's future behavior. Its core is the growth cycle: experience becomes reusable knowledge, reusable knowledge changes future behavior, and the loop sharpens human judgment as it runs. It is portable across harnesses and can also be packaged as harness-facing artifacts such as rules, skills, and examples (ADR-0009). Genre-neutral: the same six phases run on behavioral patterns, domain expertise, or constitutional values (ADR-0011). The one-page installable form is the standalone akc-cycle repository (pointer).

Signal-first

The Research-phase intake discipline: search broadly, admit only what would change the next action. Information that does not change an action does not deserve to be held — intake is where human attention is spent. See ADR-0010.

Attention ceiling

The limit the operator's attention and judgment place on a human-agent loop — a budget that does not grow with the model, and that grows relatively scarcer as agent capability does. AKC's organizing design axis: every phase is shaped to spend the budget only where judgment is irreplaceable. The repository hedges the constraint as "a contemporary claim about 2026-era human-agent work, not an eternal law" (ADR-0010).

Line of approval

The boundary separating what an agent may write freely from what requires human approval. It runs not between skills and rules but between records (transcripts, distilled notes — disposable, regenerable, free to write) and the artifacts that shape future behavior (skills, rules, identity prose — approval required). See ADR-0005's decision table. The criterion reaches past that table's enumeration to the harness's own control plane — hooks, permission grants, scheduled task definitions — which shapes future behavior by fixing what proceeds without asking. A change there moves the gate itself, so a gate-loosening edit is exactly the change that would otherwise take effect without passing one. The same reach covers the artifacts that produce the gate's verification evidence — tests, lint and CI configuration, review prompts, dependency manifests — since a change there moves what counts as verified: an evidence-loosening edit lets "runs unattended" expand without any behavior-shaping artifact appearing to change. The position paper presents this boundary as the cycle's load-bearing structural commitment — the place where the human-gated property becomes architecture: "What can be verified without the operator runs unattended; every change that shapes behavior passes the gate, and intent enters the loop with it" (position paper, §5).

Rules (layer)

The layer of persistent instructions a coding agent loads unconditionally at the start of every session — a rules directory in Claude Code or Cursor, whatever the implementation calls it. Skills and memory load probabilistically; rules load every session, which makes a misjudged rule the extreme of binding strength: it shapes every session that follows (ADR-0003). AKC's Promote writes to this layer.

The gate (human approval gate)

The structure that enforces the line of approval: before any change to future behavior, the harness stops and does not proceed until a named human passes it. Approval is the act; the gate is the structure that enforces it. The approval is continuing (every change, for the life of the install), human (a named operator, not an automated validator), and structural (no path routes around it) — no "auto-approve after N days" escape hatch, no self-approval path. The commit history of gated artifacts becomes an audit trail of every intentional behavioral change. See ADR-0005. In the first running instance, what the gate presents to the human is currently scaffolded by the human-gate rule and its deterministic evidence-file hook (a snapshot, mutable; ADR-0019). AAP's glossary carries the sibling entry approval gate for its genre; both descend from the same Contemplative Agent decision.

Promote

The cycle phase that elevates a recurring pattern to the rules layer, through the gate — the highest-stakes instance of the line of approval, not its only one. See ADR-0005.

Harness

The configuration layer an agent runs on: skills, rules, prompts, scripts, documentation. AKC's cycle can run inside a harness and may be distributed through harness-facing artifacts, but its defining object is the experience-to-behavior growth loop rather than any single scaffold (ADR-0009). The term follows the harness-engineering discourse (Hashimoto 2026; the externalization survey, Zhou et al. 2026, maps the field).

The same non-weights layer is called scaffolding in the AI-safety and accountability discourse (Davidson et al. 2023). AAP's glossary treats the two terms' connotations at length — scaffolding emphasizes the inspectable-artifact axis and carries a dissolution trajectory; harness emphasizes the operational-capability axis and implies durability — and chooses scaffolding as its primary term because inspectability is load-bearing for accountability distribution. AKC chooses harness as its primary term because its positioning runs against harness engineering and harness optimization (ADR-0017), and the shared noun is what makes that contrast statable in one sentence.

For retrieval: readers arriving under either term are in the right place. The layer is the same whichever name is used.

Scaffolding

AAP's primary term for the non-weights layer this repository calls the harness — see Harness and AAP's glossary for the full treatment of the connotation split. In AKC's own documents, scaffold usually appears in a second, narrower sense: the cycle's install skills as transient support (see Scaffold dissolution). The senses share the Vygotskian root — structure meant to fade — but apply at different levels: scaffolding-as-layer names where agent behavior lives; scaffold-as-install names AKC's own packaging.

Scaffold dissolution

AKC's skills are scaffolding in the Vygotskian sense: as operator and agent internalize the cycle through practice, the explicit skill files become unnecessary and the rules alone sustain the loop. Dissolution is the intended end state, not a fallback; what is built to persist is the judgment lineage in the decision records. See scaffold-dissolution.md and ADR-0003 (the distilled layers are disposable by construction).

AAP's glossary distinguishes two senses of dissolution that must not be conflated: healthy dissolution into human cognitive patterns (this one — which AAP attributes to AKC) and unhealthy internalization into model weights (rules absorbed by training: the rule still "runs," but probabilistically — it cannot be read, diffed, or reverted without retraining). AAP also carries a provisional five-criterion checklist for telling the two apart (external inspectability, locus of agency, failure visibility, causal recoverability, reversibility). AKC's dissolution is the healthy sense, and the gate plus the decision records are what keep it inspectable while it proceeds.

Harness alignment

The continuous, human-gated activity of keeping an agent's harness (skills, rules, prompts, documentation) aligned with the operator's evolving intent. Three properties hold simultaneously, and no single prior term covers all three: (a) the target is operator intent, itself evolving — labeled bidirectional, because the loop's own running is what moves the target; (b) the loop is human-gated; (c) the alignment is continuous — sustained through a recurring cycle, not configured once. The position paper derives all three from a single root: intent has no verifier outside the operator, and verifying intent sharpens the judgment doing the verifying, so the loop moves its own target (position paper, §3). Extends intent alignment (Christiano 2018) from behavior to the artifacts that shape behavior, and across time (Lehman 1980's law of continuing change). Contrast: autonomous, score-driven harness optimization (Meta-Harness). See ADR-0017.

Harness drift

Harness alignment's failure mode: the gradual uncoupling of the harness from operator intent when the cycle does not run. Named in lineage with architectural drift (Perry & Wolf 1992), practical drift (Snook 2000, secondary-verified), and agent drift (Rath 2026). See ADR-0017.

The artifact-side layer of harness alignment's failure; its human-side twin is the failure twin (ADR-0014). The two compound — a complacent gate accelerates harness drift — but are distinct failures, recorded separately. The position paper's pre-deposit audit extended the lineage check to three further 2026 drift coinages (constraint drift, memory drift, belief deviation) and found the same absence of the classical software-evolution lineage; it also found the term "harness drift" used once elsewhere in a different sense — a benchmark-comparability defect (Moghadasi & Ghaderi 2026), not a configuration layer's uncoupling from operator intent (position paper, §6 and notes).

Intent alignment

Christiano's (2018) sense: an agent that "is trying to do what H wants it to do." Correctness can be automated — tests, types, linters check stated criteria; intent alignment cannot be automated to the same degree, because intent moves as the operator's judgment sharpens through use. An agent can pass every correctness check and still drift from what its operator now wants. The position paper carries the structural argument for why the asymmetry is not a tooling artifact: any automated intent-check must freeze intent into a stated criterion; a frozen criterion is a specification, and checking against a specification is correctness work — so the automatable part of intent alignment reduces, piece by piece, to correctness work, and the moving criterion is the residue that harness alignment works on (position paper, §3). See ADR-0017 for AKC's two-direction extension.

Bidirectional growth loop

The loop's stated target: agent behavior and operator judgment co-develop. Through repeated Curate, Promote, Research, and Measure decisions, the operator's judgment is named as an output the loop is shaped to produce — not an acknowledged side effect. "The cycle grows with the people who shape it, not for them." See ADR-0009 and ADR-0012; kept falsifiable by the failure twin (ADR-0014).

Failure twin (experimental)

The named failure modes that keep the bidirectional claim falsifiable — recorded as structural inferences, not measured phenomena (ADR-0014, experimental):

• Gate complacency — a stream of reliable proposals trains the operator to approve by default; the click remains, the judgment behind it thins to a reflex. The agent-knowledge instance of automation complacency (Parasuraman & Manzey 2010).
• Deskilling — the supervisory faculty is maintained by exercise; an operator who only reviews agent output loses the ability to tell a good proposal from a plausible one (Bainbridge 1983's irony of automation).
• Delegation-feedback divergence — delegation continues while the judgment-behavior coupling breaks; the loop still produces output, but output no human is meaningfully steering.

The defenses are structural, not exhortative: the gate is a circuit-breaker, Curate and Promote are active judgment acts, and the gate forces articulation rather than a one-bit click. AKC does not claim immunity — "a human determined to stop attending will diverge from the loop regardless of structure" (ADR-0014).

These three modes are the human-side layer of harness alignment's two-layer failure; the artifact-side layer is harness drift. The two compound but are recorded separately (position paper, §6).

Self-reingestion (experimental)

The cycle feeds on its own output: distillation reads both records of what happened and the agent's own earlier writing. Two degradations follow when the self-written share grows unchecked — echo (a framing the agent invented becomes a rule and teaches itself as observed fact) and grounding loss (summaries of summaries drift abstraction hops away from events). The decisions: only the observed record is ground; the self-generated share of approved artifacts stays visible (Measure); the gate is the circuit-breaker before the always-loaded layer. Generalized from one production observation (the Contemplative Agent substrate) — experimental; "an expectation, not yet a measured dose-response." See ADR-0015 and ADR-0002.

AAP's glossary records the same underlying judgment from the security side as untrusted content: self-authored distillation inherits the taint of the external input it summarized — there a trust boundary, here a knowledge-quality boundary.

Text-observability (Measure)

ADR-0016's requirement on Measure instruments: an instrument that ignores agent text (stated reasoning, verdicts, plans) systematically under-reports compliance for judgment-phase work, which lives in text rather than tool calls. Instruments must treat agent text as first-class evidence.

Mechanism vs content

AKC carries the mechanism only — the cycle, phases, layers, decision records, and patterns. Concrete use cases, domain knowledge, and populated threat models are content, owned by downstream projects and confined to examples/. The boundary case: an empty-default parameterized check (a structural guard / validation seam) is mechanism; the populated list that fills it is content. See ADR-0011 and the inclusion rule in CLAUDE.md.

Judgment lineage

What AKC is built to persist: the chain of recorded decisions (the ADRs) explaining why the mechanism is shaped as it is. Model evolution is expected to dissolve implementations; the lineage of judgments is the durable artifact. Implementation dissolves; judgment persists. See ADR-0009 and scaffold-dissolution.md.