The selector — the strategy nobody set, chosen from the task's own fields

The bottom rung of the resolution, never a replacement for it:

explicit on the task → role → team → project → selector

The owner's existing choice outranks a better default — the selector speaks only where every rung above it is silent, which is the ordinary case and the intended one.

It reads fields, never vibes. The shape of the step comes from what the task already declares — its type, its skill, its stage — and the tier from the resolved model. The model applies this table; it does not reason about it — the choice is table-driven so that a script could resolve the same fields to the same row; none ships yet, and the eval scenarios are the current check — and a shape the table does not positively recognise falls to standard rather than to judgement.

The tier words map to the corpus's own (one home for this, cited elsewhere): reasoning = a model that chains its own thought — the top tier, and most of the mid tier today; light = the cheap fast tier that does not. The fallback chain's strong → mid (dispatching) and the cost file's top/mid/cheap name rungs of capability; this table asks only one property of them. In doubt, read the tier as reasoning — the cost of that error is a skipped cot prefix, the cost of the opposite is paid noise.

Shape of the step (from fields)	Light tier	Reasoning tier
mechanical — a templated edit, a move, a rename	standard	standard
a writing or analysis deliverable	cot + self-refine	self-refine
a decide with real consequences	self-consistent	self-consistent
anything unrecognised	standard	standard

Cost gates the silence (patterns §24, §21). Up to ~2× a single pass the selector acts silently. N-fold strategies (self-consistent) switch on silently only for the classes the project agreed to in config — the stock one is decide with consequences — and anywhere else they are offered in one line with the price, needing a nod: automation that quietly triples a bill is the leak the audit exists to find.

Every resolution is recorded on the run with its source — strategy: self-consistent (auto: decide-with-consequences) · (explicit: task) · (cascade: role) — because why did it cost 3× must be answerable without archaeology (§27). The session is not a dispatch: the advisor's own turn takes no strategy, the same law as it cannot re-tier itself.

This table is a locked surface. It changes by proposal with evidence — an eval that shows a row choosing badly, a cost slice that shows one not earning its keep — never by a quiet edit, and never by the run working under it. Finer judgement than the fields carry is what the explicit rungs are for; a selector that grows opinions is drift wearing initiative.