# Glossary of the aigile method

Every term of the method, one sentence each, grouped by kind. For terms
that also appear in the working paper's glossary, the working paper
remains canonical; this file adds the terms introduced after it
(standards and domain files, intent weights, the readiness gate, the
slicing vocabulary, findings) and is the complete current register.
Links point to the full treatment at https://goaigile.net.

## Written artifacts

| Term | Meaning |
|---|---|
| Constitution | The one-page law of a project: numbered clauses, each of which would justify stopping work if violated. Read by every agent, every session. |
| Clause | One rule of law (C-01, C-02, ...), carrying the version it entered at. |
| Amendment | The only way the constitution changes: a diff with written rationale (change, reason, rejected alternatives, what becomes invalid) and a semantic version bump. |
| Constitution stamp | The line on every piece of work recording which constitution version it was built under. |
| Decision record (ADR) | The reasoning behind a clause or a choice; read on demand, not by default. |
| Standards file | Technical conventions for one area (naming, error handling, testing), one screen each; loaded only when a slice declares them. Convention, not law. |
| Domain file | Cross-feature business facts (domain rules, fixed lists, how the customer operates), human-written, loaded per declaration. |
| Collaboration corpus | Written norms about working together, learned at retrospectives: wording that works, slicing rules, calibration values. |
| Idea | A backlog entry; free to capture, carries no obligations, and is not yet an intent. |
| Feature intent | The human-owned statement of purpose: why, success in user terms, scope, and known requirements. Has a weight class. |
| Weight class | Light (purpose, success, scope, knowns) or full (adds scenario, decisions, sketch); full applies to new user-facing surfaces, expected constitution contact, or larger features. |
| Known requirements | The intent's fact list under the provenance rule. |
| Provenance rule | Facts the human actually holds may be written at any length; nothing may be invented to make a document look complete. |
| Primary scenario | Full-weight intents only: an owner-written walkthrough of a named user doing the real thing; writing it forces the behavior decisions. |
| Behavior decision | A choice about how the feature behaves, explicitly made by the owner in the intent. |
| Deferred decision | A choice the owner has explicitly left open, bound to the demo that will answer it. |
| Sketch | An owner-drawn mockup for a new surface; binding for structure, not for pixels. |
| Slice list | The feature's decomposition into slices, with ordering rationale and risk flags. |
| Slice spec | The one-screen specification of one slice: summary, acceptance criteria, Standards and Domain lines, a "not specified on purpose" list, and the stamp. |
| Acceptance criterion | A single checkable statement, tagged [test] (must become an executable test) or [prose] with a justification (checked at a demo). |
| Deviation report | Written when work stops mid-build: what the spec assumed, what is actually the case, two or three options with consequences, and the recorded decision. |
| Deviation log | One line per implementation-detail deviation; raw material for the retrospective. |
| Atomic PR | The single reviewable unit: code, tests, and any spec change together, with the stamp and the deviation log. |
| Verdict | The recorded outcome of a demo: validated, misaligned, or intent evolved, with who drove and what artifacts resulted. |
| Finding | One recorded contradiction between a spec and the system, with detection and resolution dates; classified descriptive, normative, or constitutional. |
| Archive | Sealed specs of validated slices; read-only, and stale by design. |
| Derived document | A description of current behavior regenerated from code on demand; never maintained by hand. |

## Structural concepts

| Term | Meaning |
|---|---|
| The loop | The working shape: intent, slice, specify, build, verify, validate, learn, with three backward paths. |
| Backward edge | Any defined channel through which learning flows against the direction of production and changes the plan. |
| Durable layer | The small, law-like document layer (constitution, ADRs, invariants); maintained deliberately, read every session. |
| Ephemeral layer | Per-slice working documents (intents, specs, criteria); archived after validation, not maintained. |
| Derived layer | Documents regenerated from the source of truth at read time; cannot go stale. |
| Layered truth | The precedence rules: code and tests decide what the system does; the constitution decides what it should do; a contradiction between the two is a discovered defect. |
| Slice | The unit of work, defined by the user question it answers; called increment in the papers, and the two words mean the same thing. |
| Walking skeleton | The mandatory first slice: the thinnest end-to-end cut, ugly but real, demoed before further slices. |
| Spike | A timeboxed investigation of an open question whose deliverable is a written answer; it never ships code and never has acceptance criteria. |
| Sizing rule | The binary size test: one human can review the change in one sitting, and one agent can build it in one session. |
| Split catalog | The preferred seams for splitting an oversized slice: scenario, data variation, scope, operation, rule depth, fidelity; never layer or activity. |
| Detail budget | Specifications state what must stay true, not what happens to be true, because descriptive prose goes stale. |
| Readiness gate | The adversarial check before slicing: a full-weight intent whose owner cannot walk the scenario goes back at the cost of a conversation. |

## Protocols and events

| Term | Meaning |
|---|---|
| Deviation protocol | The rules for a mid-build discovery that the spec is wrong: classify observably, stop or log, never diverge silently. |
| Observable classification | The sorting question: what must change? Tests, interfaces, or the constitution moving means behavioral; internals only means implementation detail. |
| Andon | The stop-the-line event for behavioral deviations, named after the Toyota practice. |
| Consortium | Whoever the authority matrix designates to decide a stopped deviation, within the latency budget. |
| Latency budget | The agreed maximum time for an andon decision, with a named default if it passes. |
| Drift | Divergence between specifications and the system, discovered after the fact; treated like bugs, not like shame. |
| Drift pipeline | The four stages of handling drift: prevent, detect, triage, learn. |
| Conformance sweep | The check that reads specs against code and reports contradictions; a retro ritual first, a standing agent later. |
| Graduation trigger | The evidence threshold for moving from ritual to standing agent: the ritual keeps finding real drift. |
| Drift half-life | The health number: the time from a gap appearing to the gap being fixed. |
| Skeleton demo | The human-driven demo of the walking skeleton, before further slices are built. |
| Feature demo | The guaranteed human-driven walkthrough before a feature seals. |
| Risk-pulled demo | An on-demand demo of a slice that touched the constitution, stopped the line, or sits in a drift-prone area. |
| Brains-eyes-hands rule | A demo watched as a recording or read as a summary does not count; the human operates the software directly. |
| Retrospective | The learning event after a validated feature, with three routed outputs: system learnings, collaboration learnings, drift findings. |
| Comprehension audit | The retro check that each human's validation authority is backed by recent hands-on contact. |
| Gardening | The scheduled pruning of the durable layer: consolidate, delete, promote, demote, with no lost enforcement. |
| Promotion / demotion | Moving a repeatedly violated fact into the constitution, or moving over-specified prose out of it. |
| Impact census | After a major amendment: querying the stamps for all work built under older versions and dispositioning each. |
| Honest rollback | Reverting an amendment with the stamped list of work built on it in view, so the cost is enumerable. |

## Authority concepts

| Term | Meaning |
|---|---|
| Core invariant | Humans always own the why; automation only ever moves who owns the check. |
| Authority matrix | The per-layer assignment of checking work between humans and agents, recorded in the constitution. |
| Fixed row | Feature intent and validation, which never transfer to agents in any configuration. |
| Graduated autonomy | Authority moves toward agents on a recorded track record and returns on incident, like a licensing system. |
| Mandate | The written, human-held, revocable grant under which a lead-architect agent may decide amendments. |
| Human sampling | The share of agent-reviewed work that humans still review; never zero, serving as tripwire and comprehension spend. |
| Correlated verification failure | The hazard of agents checking agents of the same model family: shared blind spots approve the same error every time. |
| Scribe | The agent role at demos that turns spoken feedback into spec-diff proposals for the human to approve in session. |
| Goodhart anchor | The rule that self-improving agent processes are judged only against human demo verdicts, never against internal numbers alone. |
| Rulebook | The numbered rules R1 to R36, defined across the deep-dive papers. |