Glossary

A/B/C Levels. Verification automability categories. Level A: agent executes autonomously with binary pass/fail. Level B: agent executes, judgment evaluates. Level C: human evaluates. See Automability Levels.

Bootstrap. One-time project setup using michi-bootstrap. Surveys the project, recommends a documentation structure, scaffolds Michi artifacts interactively.

Campaign. One dated run against a charter in an expedition. The learning-mode analog of a milestone. Opens with a mandate, executes as a run, closes with a review.

Charter. A durable mission or theme in an expedition — revised as new data lands, not redone. The learning-mode analog of an epic. The name signals a standing line of inquiry with no fixed end.

Context compaction. Claude Code’s mechanism for managing the context window in long sessions. Earlier content is compressed, which can evict file read caches and degrade the agent’s recall of early decisions.

Dark factory. A manufacturing plant that runs with the lights off — no human presence. The software analog: an autonomous agent completing work with minimal human involvement. Michi’s name comes from this metaphor.

Debrief. Post-session review using michi-debrief. Assesses delivery, reviews decisions, captures learnings, calibrates trust for the next session.

Michi. The toolkit itself — process, skills, and documentation structure.

Entrusted mode. The agent has wider initiative within established scope. You review at gates and exceptions. Earned through demonstrated reliability. See Modes.

Epic. A named unit of work — anything substantive enough to track. Organized in docs/epics/ either as a flat file for short-arc work or as a subdirectory containing specs, plans, verification, and journals for multi-milestone work. Multiple concurrent epics is normal; the “main” thrust is signaled in STATUS.md and README.

Expedition. Michi’s learning-mode lane — open-ended, spiral exploration where the end isn’t known, run with michi-expedition. The sibling of the production lane (planning → session → debrief). See Expeditions (Learning Mode).

Given-When-Then. Format for decomposing verification scenarios into executable steps. Given (preconditions), When ( action), Then (expected outcome).

Holdout tests. Verification tests the agent never sees during development. Run post-completion to catch assumption errors. Layer 4 of the 5-layer strategy.

Iron Law. “No ‘done’ without running full verification.” The non-negotiable rule in michi-session.

Iteration cycle. Explore → Brainstorm → Plan → Execute → Verify → Document. Applies at every scale. See The Iteration Cycle.

Mandate. An expedition campaign’s contract: scope, budget, when to stop, and what a finding must pass before it’s trusted. Bounds the open-endedness from outside so the agent neither runs forever nor parks every few minutes.

Milestone. A scoped increment of work within an epic. Each milestone produces something testable and has its own plan doc.

North Stars. The four governing principles: What Over How, Minimize Latency, Sustain the System, S-M-L-XL. See Principles.

Paired mode. Human and agent work in a tight loop. Human present, engaged, driving. The mode for learning, new codebases, and risky work. See Modes.

Plan doc. A markdown document that serves as the implementation contract for a milestone. Contains steps, acceptance criteria, scenarios, and sections for decisions, notes, and discussion that the agent fills during execution. See Plan Docs.

Portrait. The accreting picture in an expedition — the best-current understanding, promoted from verified findings. Replaces “deliverable” as the thing the spiral aims at: blobs of color sharpening toward a photograph.

Progressive detail. Three levels of documentation depth: Principles (always in context), Guidance (loaded per phase), Checklists (referenced during execution). Matches the S-M-L-XL scale.

Rule of 3. When thinking becomes circular — same ground, same fix, same hypothesis — stop after three rounds. Come up for air. Three iterations without progress is a signal to step back.

Ruminations. The open, still-murky companion to an expedition portrait — what’s unresolved and worth mulling, not yet actionable. Distinct from the backlog (actionable, tracked) and the portrait (becoming clear).

S-M-L-XL. The relative scaling framework. Applied to effort, benefit, risk, documentation depth, process rigor, and verification depth. See Principles.

Scenario. A story about a user getting a benefit, decomposed into Given-When-Then steps and categorized by automability level (A/B/C). See Scenario Testing.

Smoke test. A real API call with real data through the full stack. Layer 3 of the 5-layer strategy. The layer most commonly missing from agent-assisted development.

Spiral Iteration. The property that iterations progress forward, not in circles. Each cycle produces a deliverable AND richer context for the next cycle.

Sustainability. The principle that each iteration should leave the system where the next one is at least as productive. Assessed by michi-sustainability at scaled intervals.