ADR-0145 — Attention economy and decision rails

Context

Agents are good at investigation, mechanical execution, and evidence collection. Humans are the scarce authority for intent, material tradeoffs, waivers, irreversible actions, and acceptance. Current rules spend that authority poorly: optional economy guidance permits repeated prose, bare questions permit guesses, review rows withhold assessments agents can make, and repeated raw output obscures the evidence it is meant to expose.

Decision

1. Disrespec governs Markdown artifacts. It maximizes clarity per token: one fact once, no filler, repeated source material, empty section, or chat conclusion that restates a successful artifact. Chat returns only clickable artifact links. Blocked decisions, failed creation, incomplete verification, and irreversible-action confirmation remain explicit. Source code, raw evidence, commit messages, and repository-native PR formats are exempt. Its full method is single-sourced; every artifact writer carries one byte-identical spine checked for drift. Level: toolable through family gates.

2. Discover facts before asking. Agents decide reversible, convention-bound details. Material behavior, public contracts, security tradeoffs, costly choices, conflicting authority, and irreversible actions require human selection. Prefer the native picker: three genuine options by default, two for binary choice, recommendation first, one-sentence tradeoffs, and Other. Without a picker, render the same numbered choices plus Other; never ask a bare question. Batch only independent decisions. Deferral leaves the artifact draft or pending and blocks dependent work.

3. Specs contain only useful sections. Intent and Requirements are the minimum. Non-goals, Open questions, Affected areas, and source-drop sections are conditional. C005 and C006 retire without ID reuse. C007 still blocks a ready spec containing a blocking unresolved decision.

4. Agents assess; humans decide. Review coverage is ID | Assessment | Evidence with Supported, Unsupported, Unverified, or Blocked. Frontmatter carries decision: pending | accepted | changes-requested | deferred; an agent leaves it pending, then presents a state-aware human picker. Only that selection changes the decision. Accepted reviews with Unsupported or Unverified rows list their IDs under waivers. Summary, task status, suggested decision, and empty attention sections retire.

5. Evidence is written once. Short output appears once. Dominating output moves to adjacent evidence-<slug>.md. A receipt contains stable E-NNN anchors, command, working directory, state identifier, exit status, and untouched raw output. Governing artifacts carry a decisive verbatim excerpt and receipt link. One command may support several claims when each maps to its anchor. Existing verify info strings remain command-result records; bodies may carry the excerpt and receipt link. Without an artifact, chat may carry compact command, exit status, and decisive output.

Narrowed decisions

• ADR-0058: fixed empty sections retire; requirement records stand.
• ADR-0083: verify blocks stand; Pass becomes Supported.
• ADR-0097: C016 becomes supported-needs-evidence; C005 and C006 retire.
• ADR-0101: structured choices stand; their section is conditional.
• ADR-0109: optional economy becomes Disrespec for artifact writers.
• ADR-0119: independent assessment and human acceptance coexist.
• ADR-0121: receipts replace evidence repetition.

Consequences

• Human attention lands on decisions, evidence gaps, waivers, and irreversible acts.
• Agent assessments become explicit without granting agents acceptance authority.
• The checks contract advances to 0.17.0.

Status

Accepted (2026-07-12).