ADR 0079: improvement is a coach, not an audit

Amendments.

• Vocabulary: current spellings are standards (formerly ratchets) and improvement (formerly improve); current pointers reflect the suggestable-command surface that replaced the old Recipe name satellite; the decisions below are unchanged.
• ADR 0301 — lead priority: the lead-priority clause below narrowed — with the objective baseline clear, the first evidence-backed recommendation now leads next_action as the decide kind.

Status: accepted. Extends ADR 0029: its deterministic-rule and qualitative-review split remains; this decision changes the product frame and adds a priority above the report.

discern audit already separated facts the binary could score from reviews that needed an agent's judgement. Its output still looked like a compliance snapshot: a green Overall 100/100 headline followed by a dim parenthetical review count. The number was honest, but the presentation made an objectively complete baseline read as "nothing left to do." Users then had to turn the weakest-first report into their own priority queue.

The catalog also needs to teach practice quality, not reward file existence. Tests can exist but be coupled, serial, or shallow; docs can exist but drift or be impossible to navigate; guidance can exist but repeat code; a raw ceiling count can punish healthy growth. Those questions require judgement and must not be smuggled into the deterministic score.

Renaming a CLI verb and MCP tool is a hard-to-reverse public-surface decision. An alias would ease the transition, but it would permanently preserve the compliance-flavoured name in help, documentation, parity exceptions, and user muscle memory. The project is under the breaking-change policy of ADR 0009, so this is the point at which to choose one vocabulary.

Hard-rename the verb to improvement and its MCP tool to discern_improvement. There is no audit alias. The CLI verb registry remains the canonical set; the existing parity guards force the command registration, suggestable-command surface, MCP table, grouped help, schemas, and tests to move in lockstep.

Position improvement as a continuous-improvement coach:

• The score is presented as Automated practice health, followed by separate, non-dim lines for the number of objectively weak rules and the number of open improvement reviews. 100/100 therefore means only "nothing objectively weak."
• Every successful result carries one structured next_action, rendered near the top in human output and returned unchanged through --json and MCP.
• Objective gaps lead, ordered by recoverable weighted score credit. A failed weight-three rule outranks a failed weight-one rule; a failure outranks a partial at the same weight. Ties retain catalog order.
• When the objective baseline is clear, the first applicable qualitative review leads. Catalog order is the deliberate coaching priority.

Keep deterministic rules and qualitative reviews distinct. Broaden the latter to describe the shape of good practice—behavioural and isolated tests, actionable failure memory, current and navigable docs, non-inferable guidance, rate-normalized standards, isolated worktree resources, and executable skills—without moving those judgements into the score.

• A run answers "what should improve next?" before presenting the full report. Agents no longer need to invent prioritization from a list of findings.
• A perfect baseline remains useful as an objective signal while visibly coexisting with judgement work. The score is less likely to terminate the improvement conversation prematurely.
• CLI scripts and MCP clients using audit / discern_audit break and must move to improvement / discern_improvement. This is intentional; maintaining two names would keep the rejected frame alive and double the surface that parity guards must account for.
• The priority algorithm is intentionally simple and explainable. Deterministic weight is a proxy for value, and catalog order is editorial judgement for qualitative reviews; neither claims to know project-specific effort or urgency.
• Subjective reviews remain stateless. The agent lands its judgement as a real project change, then reruns improvement, preserving ADR 0029's model.

• Keep audit, change only the headline. Rejected: the noun still frames a periodic inspection, while the product behaviour is an ongoing coaching loop.
• Add improvement as an alias. Rejected: it avoids a short migration at the cost of two permanent public names, stale examples, and ambiguity about which tool an agent should prefer.
• Remove the score. Rejected: deterministic baseline health is a useful, enforceable signal through --min-score. The problem was presenting it as completeness, not measuring objective wiring.
• Score qualitative reviews. Rejected: the binary cannot render those verdicts honestly. Assigning points would turn judgement into fabricated precision and violate ADR 0029's central split.