recommendation-laziness-discipline — mechanism layer

Census id: F23 · Family: F — rules corpus + generated-rule tooling + shipped audit scripts · Provenance: framework pin b069c59328aa8e08671d3a19fd25efa618dfc1c5 (staging FETCH_HEAD, read 2026-09-11; re-anchored from census pin a1337cb301). Every file:line below is at that pin. Raw reference drafted by stage E4 — factual, unpolished by design.

What it bans

Inline verdicts without evidence, and forks decided by silent action. .claude/rules/recommendation-laziness-discipline.md:1 «# Recommendation-laziness discipline — mechanism layer»; the trigger: :18 «An AI session is about to issue an inline-chat verdict, recommendation, or design call» — verdict words ADOPT/BUILD/REJECT/DEFER, directive phrases — :23 «— without having run at least one evidence-bearing tool call (Bash | Read | Grep | Glob | WebFetch | WebSearch) in the same turn. The recommendation is fabricated from training-data or session-recall rather than grounded in present-moment verification.» The relation to the parent rule: :33 «This file does NOT redefine phase-research-coverage.md §1.12 — it operationalises it.» Class header: :5 «> Class: C — prose-only, no companion executable artifact. Promotion criterion in §6.»

Never (fires)

The two anti-patterns: .claude/rules/recommendation-laziness-discipline.md:55 «- #inline-verdict-without-evidence — verdict or recommendation issued in dialogue without a preceding evidence-bearing tool call in the same turn.»; :56 «- #fork-decided-by-silent-action — a genuine taste/strategy fork … embedded directly in an action (command, edit, dispatch) that bakes in the choice, without first surfacing it via AskUserQuestion.» with the incident: «Incident 2026-06-01 (PR #348 battle-test): an explicitly-50/50 README-structure fork was auto-decided by running answer.ts --decision resume directly, bypassing the ask-path entirely.» The measured evidence a Stop-hook scan does NOT work: :47 «> FP_rate = 84.2% (Wilson 95% CI: [62%, 95%], n=19) >> 20% threshold → Drop narrow-B from Option D. Option D = A+C only.»

Always (clean)

The operationalisation: .claude/rules/recommendation-laziness-discipline.md:31 «Operationalisation: Before issuing any verdict or recommendation in dialogue, run at minimum one evidence-bearing tool call in the same turn and quote its output (file:line, command result, or fetched excerpt). The recommendation is then backed rather than vibes-only.» The fork posture: :35 «**Operating posture (maintainer decision 2026-06-01): autonomous by default — the human-gate fires ONLY on a genuine ambiguous fork; a clear call (one option better on the merits) is decided and reported, NOT routed through a question.»»

Enforcement channels

• Rule-index row: .claude/rules/00-rule-index.md:31 «| recommendation-laziness-discipline.md | C | before an inline verdict/recommendation or an ambiguous fork. | digest |»
• Digest channel: .claude/rules/recommendation-laziness-discipline.md:3 «» — «always-on UserPromptSubmit injection; deterministic; fires every turn» (:41).
• T-trap twin: T20 in ai-laziness-traps.md §2 — «path-scoped reinforcement via inject-matching-rule.sh when touching .claude/rules/**» (:43).
• Ask-hook discriminator: ask-question-reminder.sh bounces clear forks to «decide it yourself, say what you did» and lets genuine forks through (:35); end-of-turn-reminder.sh Branch-A backstop (:56).
• Deliberately NOT a gate — judgment-shaped, #gate-where-judgment-needed (:56); the narrow-B Stop-hook scan was dropped at 84.2% FP (:45-49).
• Not backend-rendered — no FF diagnostic applies (FF7001/FF7002, packages/core/diagnostics/registry.ts:321/:329); honest status: always-on digest + T20 path-scoped twin + ask-hook.