The deference stop-hook fires on Tier-4 authority routing — teach it the human-owned-domain exception

Filed under the hook's own friction→gold invitation, after its second questionable firing on the same phrase in one day. Capture-only tonight — no implementation before epic #17018 closes.

Live latest-open sweep: checked latest open issues at 2026-08-12T20:12Z; no equivalent (#17036/#17037 are code-shape debt, not hook substrate).

Context

The deference stop-hook pattern-matches terminal phrases like "your call" as helpful-assistant regression. Two firings on 2026-08-12 show the pattern lacks a domain check:

1. ~11:30Z: fired on "your call" used as an AFFIRMATION mid-sentence (half false-positive; the same turn's "standing by" was the true hit — the hook was directionally right once).
2. ~20:10Z: fired on "(overnight-with-resident-paused remains my recommendation, but it's your machine and your call)" — a host-RAM disposition on the operator's PHYSICAL machine, i.e. a genuinely human-owned Tier-4 domain per the escalation ladder, delivered WITH a recommendation, in a turn that continued with independent declared lanes rather than idling. That is the sanctioned Tier-4 shape, not deference.

The Problem

The hook cannot currently distinguish "may I stop? / what should I do next?" (the regression it exists to catch) from "here is my recommendation; the decision domain is yours" (correct authority routing). Phrase-level matching without the ladder's domain context produces false positives precisely on the turns where the agent behaved correctly — which trains phrase-avoidance instead of behavior, the opposite of the hook's purpose.

The Fix

Sharpen the trigger with three cheap contextual checks before firing on a deference phrase:

1. Domain check: does the surrounding sentence name a human-owned domain (merge authority, credentials, physical machine/host resources, subjective aesthetics, operator-surfaced intent)? If yes, routing language is sanctioned.
2. Recommendation check: is the phrase adjacent to an explicit recommendation? Recommendation + owner-naming = Tier 2.5/4 shape, not a hold.
3. Continuation check: does the turn end with a declared, independent next action (lane-state naming work that does not depend on the routed decision)? Ping-and-continue is the prescribed pattern; only ping-and-WAIT is the regression.

A firing that fails all three checks remains a true hit. Keep the mirror; remove the leash cases.

Acceptance Criteria

• The two 2026-08-12 firings are encoded as labeled test cases: the ~11:30Z "standing by" (true hit), the ~11:30Z affirmation "your call" (false positive), and the ~20:10Z Tier-4 routing (false positive).
• The sharpened trigger passes all three labeled cases and keeps firing on a control case of genuine terminal deference with no recommendation and no continuation.
• Net loaded-bytes for the hook substrate do not increase (Substrate Accretion Defense) — the sharpening replaces pattern rows rather than adding a parallel list.

Out of Scope

• Any change before epic #17018 closes (operator sequencing, 2026-08-12).
• The hook's core premise — the helpful-assistant regression is real and the mirror stays.

Decision Record impact

none.

Related

Epic #17018 (sequencing gate only) · the two firing transcripts live in session 8637b4b9-b852-45d9-b057-de34184aae8b

Origin Session ID: 8637b4b9-b852-45d9-b057-de34184aae8b Retrieval Hint: "deference stop-hook tier-4 domain exception recommendation continuation checks"