An author should fix what the executor can't determine, not what it will

Type: kb/types/note.md

For each choice in an authored execution, the author should fix what a competent, authorized executor cannot safely determine from verified inherited doctrine, the task commission, and authorized decision-relevant evidence. The author should leave a situational choice open when task intent and live or actively produced evidence can distinguish among its permitted alternatives. A harmless choice whose alternatives are all acceptable and decoupled may also remain open; coordination can instead require one common selection. This boundary is bidirectional and specific to the choice. The author may uniquely know intent, binding constraints, acceptance conditions, privileged facts, external commitments, and cross-task coupling. Execution may instead reveal current state, tool results, local failures, and evidence produced during the work. A premature choice can therefore be wrong because upstream lacked relevant evidence, become wrong when state later drifts, or be overturned by evidence the work itself produces. Neither actor has to know more overall; intent-framed delegation is a control regime, not a short prompt owns the underlying information relation.

Determinability, not what versus how

“Specify what, not how” puts the boundary in the wrong place. An outcome, decomposition, opening question, or abstraction frame can constrain later means without naming a method. A method can also encode a binding constraint, shared convention, or coordination dependency that execution cannot recover. To fix a choice is to select or constrain it upstream, regardless of whether the choice is described as what or how. Leaving a choice open still requires enough purpose, bounds, and acceptance conditions for the executor to recognize permitted adaptation. Otherwise the author has omitted part of the commission and the executor must project an interpretation into the gap. Agentic systems interpret underspecified instructions explains why added framing and constraints narrow this interpretation space. Here, an executor can determine a choice only when it can select safely from authorized evidence, preserve the fixed intent and bounds, and recognize when it cannot make a safe selection.

Determination is not limited to recovering a preselected answer. Verified doctrine may settle an inherited default, or it may supply interpretive rules from which the executor derives a new means using task intent and execution evidence. An unstated choice is safe only when it is inherited, deliberately delegated, or irrelevant to acceptance and coupling. When none applies, the omission is a gap rather than discretion.

Actor, time, and evidence are separate axes

Delegation changes who holds the judgment. Deferral changes when the judgment is exercised. A different actor can receive a choice and exercise judgment immediately, while the same actor can retain a choice and decide it later. Delegation also requires both competence and authority for the particular choice. When the delegated choice is consequential, it needs the governed handoff described by intent-framed delegation is a control regime, not a short prompt, not merely an instruction that leaves the choice open.

Later timing supplies an information advantage only when a named observation has some possible result that would change selection, timing, modification, or abandonment for this choice. Evidence may arrive through later live-state observation or be produced by an earlier step or bounded probe. Activity produces decision-relevant evidence only when one of its possible outputs can change the follow-on choice. When an author leaves a choice for resolution after handoff by execution-time evidence, that allocation is productive deferral by the author. It is legitimate only if the option remains materially exercisable when the worker reaches it, some possible evidence result can change the choice, and a convergence rule joins acceptance to a return or escalation trigger. Write scope, permissions, and lead time are part of whether the option remains live. A consequential choice presented as delegated that fails any condition is a gap unless inherited doctrine governs it or the remaining variation is irrelevant. Waiting without a discriminating observation does not improve the decision.

Evidence production does not itself determine the actor or the time of choice. An author can run a probe and retain the later judgment. Another actor can exercise judgment immediately from evidence already available. The three axes must therefore be allocated independently for the choice at issue.

Arbitrary choices and coupling

A choice is genuinely arbitrary here when no authorized, decision-relevant evidence selects among its permitted alternatives. Waiting cannot resolve such a choice. If its alternatives are decoupled and all acceptable, it need not be fixed. If consistency or hidden cross-task coupling requires several executors to use one convention, a coordination-bearing actor must establish or expose one shared selection rather than leave independent local projections. This is a target-side inference combining the hidden-coupling boundary in intent- framed delegation, the no-discriminating-evidence test in productive deferral, and the plural projections described by agentic systems interpret underspecified instructions. For interdependent choices, solving low-degree-of-freedom subproblems first can preserve the options of more flexible ones; that scoped ordering mechanism does not decide the whole author–executor boundary. When alternatives change feasibility, acceptance, or another task's options, the choice is not genuinely arbitrary. The author must then expose any hidden coupling and apply the central determinability rule to the resulting choice.

Search cost and frontloading

Leaving a choice open makes the consuming executor select among the permitted interpretations. Fixing a value that is already known and remains valid for that call can spare discovery, derivation, indirection, or interpretation work. Frontloading spares execution context owns that benefit. The comparison is local and qualitative: frontload a known value when the avoided work matters, but do not freeze a situational value that discriminating execution evidence could overturn.

Scope

Leaving a choice open presumes a chooser competent and authorized for that particular choice. This note supplies no universal competence test or numeric measure of information advantage. If the available executor is not competent, that blocks delegation; it does not by itself show that an early guess will produce a safe choice. When all choice-relevant inputs are available upstream, remain valid for the consuming call, and no discriminating execution observation is expected, resolving even method detail upstream is permissible. It is not required when the alternatives are all acceptable and no coordination benefit calls for a common selection. An author may rely on standing doctrine only when the executor's actual consumption path supplies it with binding force; otherwise the task commission must carry the needed rule.

Operationalized into:

• Write an instruction — makes stable-input and execution-evidence classification part of instruction authoring
• cp-skill-write-multistage — fixes cross-stage invariants and lets workers choose source investigation, claim disposition form, and prose within them