blog: Rolling Out Action Authority on New Surfaces

[amavashev]

Summary

Operational counterweight to the five theory posts shipped this session. Extends the existing shadow-to-enforcement cutover decision tree per-surface for the action-authority extensions covered in PRs #648–#652.

Structure: four-week playbook + rollback tree

• Week 1: Inventory — surface-by-surface template with current gate, blast radius tier, reversibility, frequency, existing audit
• Week 2: Shadow-mode instrumentation — per-surface dry-run table (memory / merge / click / voice)
• Week 3: Per-surface gate primitives + calibration — explicit caps for each surface, drawn verbatim from the sibling posts:
  • Memory: per-tenant write quota, TTL on unverified facts, per-write provenance, scope isolation
  • Merge: distinct-approver, requires_human_approval, per-session promotion budget, deploy-gate
  • Click: requires_fresh_screenshot, cross-tenant nav deny, target-intent risk schedule, risk-denominated session budget
  • Voice: predictive reservation, wall-clock cap, tier-aware gating, per-turn-boundary re-check
• Week 4: Cutover, lowest-false-positive-cost first — memory → click → voice → merge
• Rollback decision tree — denial-rate thresholds (>2× shadow rate for 1h flips to shadow; >5× for 15min kills the gate), with concentration-heuristic-labeled per-tenant and per-agent rollback signals
• Monitoring (first 72h / first month / ongoing)
• Short runbook template — 7 entries per surface

Author: Albert Mavashev
Date: 2026-05-22
Word count: ~2,850 body

Reviews

• Internal cycles 1–3 (scorecard 9.4/10)
• Glossary auto-linker applied 7 contextual links
• Codex external review: round 1 REVISE-MINOR (8 findings, all applied / 0 pushed back), round 2 SHIP

Codex verified all per-surface gate primitives match the sibling PRs #648–#652 verbatim or near-verbatim (15 of 16 caps directly cited).

Notable changes through review

Cycle 1 caught and fixed:

• Terminology drift: "reserve-to-actual ratio" → "reserve-to-commit ratio" (canonical corpus term)
• Tag: rollout → adoption (matches cutover-decision-tree precedent)
• Click "wall-clock cap" reframed as "session budget denominated in risk, not count" (matches the click sibling's actual framing)
• Cutover-order metric clarified as cost-of-false-positive-denial, not action blast radius
• R/A ratio row in rollback table scoped to voice-specific
• 20%/40%/30% concentration thresholds labeled as starting heuristics
• Week 3 calibration window made explicit ("at least a week"; parallel-surface adoption requires extending Week 3 to 2–3 weeks)
• Two rhetorical closers cut
• "What Action Authority Adoption Is" section tightened from anaphoric essay to operational summary

Codex round 1 caught additional:

• L36 synthesis quote "the lifecycle is the stable layer" — the actual H2 in PR blog: What Four New Surfaces Taught Us #652 is "Reserve-Commit Is the Stable Layer." Replaced with prose paraphrase.
• "Risk order" / "lowest-risk" framing throughout aligned with "false-positive-cost order" / "lowest-false-positive-cost" (the actual ranking metric).
• L103 calibration band absolute "the quota is wrong / not constraining anything" softened to starting heuristic framing.
• L125 "Most shadow weeks produce a clean bimodal distribution" hedged.
• L138 cross-surface "reserve-to-commit ratio" claim scoped: voice has a true R/C ratio; other three surfaces use cap-fire rate vs shadow baseline as the analogue.
• L152 ">85% intended denials" labeled as minimum triage bar; sensitive surfaces target higher fractions.
• L187 "Reserve-to-actual ratio per surface" rewritten to scope voice R/C separately from the other three surfaces' cap-fire baselines.

Per-dimension scores

Dimension	Score
Factual accuracy	9.5
Credibility	9
Cross-links	9.5
SEO (title 44/51, desc 155/160)	9.5
Code accuracy	10
Structure & flow	9.5
Terminology	9.5
Tone & style	9.5

Overall: 9.4 / 10

Test plan

• `npm run dev` and verify post renders at `/blog/rolling-out-action-authority-on-new-surfaces`
• Verify post appears on `/blog/` index sorted to top (date 2026-05-22)
• Click through all internal links — five sibling links depend on PRs blog: Agent Memory Writes Are Actions, Too #648–blog: What Four New Surfaces Taught Us #652 being merged first
• Confirm date/author/tags/reading-time header renders above body
• Confirm Prev/Next post navigation works
• `npm run build` succeeds with no broken-link warnings

Dependencies

Depends on PRs #648–#652 being merged first. This post is a practical playbook layered on top of the five-post arc; it references all five siblings extensively in the intro, the Week-3 calibration sections, and Next Steps. Required merge order: #648 → #649 → #650 → #651 → #652 → this PR.