A surfaced choice is a triage, not a page

Status: concept note + doctrine + one shipped leaf (internal/choicetriage) and its first consumer (internal/operatorbrief). This note is a rung under the no-babysitting umbrella: that note bounds how a genuine human page behaves (rare, cheap, evidence-carrying, bounded silence); this note asks the prior question — was the page genuine at all? It is the decision-side dual of internal/waiting: waiting bounds how long a real human hold may sit; choicetriage decides whether the hold was ever a human’s to make.

The operator’s ask

Look at how we surface choices. Most of the time we stop and ask a human to pick one of two or three options. But that framing centers the human by default. Decenter it. Almost every “choice” is really: one obvious option an agent should just take; or an evaluation that wants a fresh context window at the best model to work it end-to-end; or a piece of work too big for one step that should become a GH ticket. The genuine “a human must decide this” case is the small remainder, not the default.

The thesis

A narrating fleet manufactures decisions. It reaches a fork, prints two or three options, and waits. The act of printing options makes the fork look like a human’s to resolve — but the framing is the bug, not the fork. Fold any surfaced choice honestly and it lands in exactly one of four places:

Disposition	What it really is	Who moves it
TAKE_OBVIOUS	one sane move; the “choice” is a to-do	an agent, now, best-effort
FRESH_CONTEXT	knowable but not obvious	a fresh context window at the top model tier
FILE_TICKET	real work, too large for one context	the fleet, via a new/existing GH ticket
HUMAN_RESIDUAL	a genuine policy / auth / release / priority / trust call	a person — the only case that waits

The load-bearing inversion: the default is FRESH_CONTEXT, never HUMAN_RESIDUAL. A human decision must be earned by a real signal that names authority a person holds (approve/release a build, grant an auth, set a priority, make a policy/legal/budget/trust call). A choice that names none of those is, by construction, not a human decision — so it routes to a clean evaluation, an obvious action, or a ticket. The human residual shrinks toward zero as the other three absorb what was only ever pretending to need a person.

Why this is the dual of no-babysitting

Babysitting is polling — a human running a busy-wait loop over a system with no interrupt controller. The no-babysitting note converts every poll into a rare, typed interrupt. But an interrupt is only cheap if it is warranted. A fleet that pages a human for every fork has a working interrupt controller wired to a smoke alarm that trips on toast. choicetriage is the filter on the interrupt line: it keeps the “seven watches” honest by refusing to raise a human-decision interrupt for something that was an obvious action or an evaluation. Green means witnessed, not quiet; and quiet here means the human is not paged for a decision the fleet could make itself.

What shipped

• internal/choicetriage (tier 1, pure, stdlib-only, imports nothing internal): the closed Disposition vocabulary above plus Triage(Signal) Verdict, a deterministic fold — state in, disposition out, no I/O and no clock — so any layer (the operator brief, a dispatch loop, a stop hook) can route a surfaced choice through the same taxonomy. The authority test is a small, conservative, declared token list (mirroring waiting.BlockedReasonHints); it is deliberately hard to trip, because the whole point is to make HUMAN_RESIDUAL rare.
• internal/operatorbrief now folds every Choice through Triage and carries Disposition / Resolve / NeedsHuman as the load-bearing fields. A missing-report page whose fix is a runnable command (generate \fak cadence –json`…) becomes TAKE_OBVIOUS — regenerate it, page no one; a release go/no-go stays HUMAN_RESIDUAL. The legacy Default/Options` “pick one” fields are kept for back-compat and marked for retirement.

Rungs (what this note does NOT yet ship)

• R1 — more consumers. The dispatch stop-hook and notify path still phrase some forks as “pick one” without folding through choicetriage. Route them.
• R2 — retire Default/Options. They encode the old human-centered framing; Disposition/Resolve supersede them. Deprecate then remove.
• R3 — the deeper rename. worktype classifies work; choicetriage classifies choices. There is a latent shared idea — a “program” of work the fleet owns end-to-end vs. the residual a human owns — worth unifying so the two taxonomies do not drift. Tracked as #3137, not done here.
• R4 — a residual counter. Count HUMAN_RESIDUAL dispositions per witnessed unit of work, the decision-side analogue of the babysitting touch counter, and ratchet it down. A rising residual rate is a regression in the doctrine.

The falsifiable property

Over a soak window, the fraction of surfaced choices that reach a human is bounded and trending down, and every choice that does reach a human names an authority token a person actually holds. A human paged for an obvious action, an evaluation, or ticket-sized work is a bug against this note.