D231 - A chooser is slower than running everything it would choose between

decided · 2026-08-25 · measured, after building the thing this retires

The plan was an orchestrator: something that understands orbistoun's tools and picks the next experiment given the evidence. Two measurements, taken while building it, say that is the wrong shape.

	cost
one boot of the guest against a wall	~0.13 s
every argument of every import, exhaustively	276 boots, 50 s
every other diagnostic axis against one fault	6 boots, 1 s
one answer out of a local model, on the GPU	5-20 s

A boot is cheap because a wall is hit early. So the model is slower than exhausting the space it would be selecting from, and any prior it supplies saves nothing. This already retired orbistoun-propose::wall, which ranked which import and argument to suspect; it now also decides what the orchestrator is.

orbistoun-propose::turn is therefore a dispatcher, not a chooser. Every Gap the report can name maps to a fixed Step, and where the step is a sweep the sweep is exhaustive. A model appears in exactly one branch - NameAHash - and that is a measured placement rather than a preference: of the names a model earned against the hash oracle, not one exists whole in any module string and none of its words appears in any vendor-prefixed string, so the string harvester could not have found them. The two sources are disjoint, not redundant. only_the_branch_with_an_oracle_behind_it_proposes_anything pins it, so widening it is a deliberate act rather than a drift.

Three consequences worth stating, because each is a thing not built:

• No ranking. The report already ranks by how many calls a finding concerns, which is a fact about the run. Re-ranking here would be claiming to know better on no evidence.
• No implementation. Gap::Unimplemented maps to Step::Person, pinned by a test, because THE_LOOP is explicit that step 18 is a person writing code.
• A refusal states why. Step::Person carries a sentence naming what is missing. A dispatcher that quietly did something plausible with a gap it has no rule for is principle 3's failure one level up.

Measured end to end against a live title: eight findings, nine steps, everything mechanical done in 2.6 seconds, stopping where a person is genuinely required.