World choice and the consistency gate

Parent: recursive-harness

You don't need deterministic dispatch/taste. With a Kripke / possible-worlds framing, non-determinism is fine — it just selects a world.

Taste = selecting a possible world; convergence is per-world

A taste/dispatch choice = picking a world w. Convergence is per-world: each world has its own fixed point; you pick one and converge within it. The non-determinism moves from "noise that breaks convergence" to "which convergent world to be in." This also re-solves convergence-needs-an-observable-target: don't converge to the one true (unobservable) target — choose a world by fiat and converge there; under-determination is resolved by choice, not observation.

The gate flips from "match truth" to "stay consistent"

If convergence = "stay consistent with the world you chose," the per-node gate is "are you self-consistent with the prior commitments?" — not "did you match the unobservable truth?" Consistency is far more checkable than truth. This is why dispatch needn't be deterministic: you don't verify it's right, you verify the path stays coherent with it.

"Then everything tightens" = accessibility contracts

Each choice shrinks the consistent-continuation set (the accessibility relation contracts); the constraint store grows monotonically → converges to a single world. Formal home: non-deterministic choice + monotone constraint store (constraint logic programming / Prolog choice-points / modal-logic tableaux building a model). Convergence = reaching a complete, consistent model (a fully-pinned world).

The hard part: wrong-world lock-in

Tightening helps convergence but hurts correction — an early wrong taste-choice gets locked in by later tightening. Mitigations:

1. Order choices: cheap/reversible early, locked-in late (real-options / staged commitment — don't bet big while the world is still wide and you're unsure).
2. Backtracking / belief revision (AGM, non-monotonic logic) to loosen a tightened constraint.
3. Human as world-acceptor (below).

The human's role gets lighter and truer

Not "specify the unique target" (impossible) but "accept / reject the coherent world the agent committed to, and redirect if the taste is off." That's exactly how creative/design collaboration works — react to a coherent proposal, redirect. A much lighter burden than full specification.

Open knobs (set the harness's strength)

• How mechanical is the consistency gate? If "consistent with prior commitments" can be written as checkable constraints (types/lint), convergence is guaranteed; if it still needs semantic judgment, it's high-probability.
• Who picks the world? Human at the top (/goal-style) or the agent samples its own taste — both converge per-world; agent-picked just needs the human-acceptor to catch off-taste worlds.