Control methods & their direction

Eight ways to govern a non-orchestrable, autonomous, possibly-cheating executor — and the single direction they all push.

The direction (why they're one family)

All eight do one thing: accelerate the convergence of uncontrolled/stochastic behavior toward a target outcome G. Formally they're operators on one stochastic process (7,8 on the generator; 1,2 on the support; 3,4 on the cost structure; 6 on sampling; 5 on cross-round dynamics). Quantitative ancestor: Ashby's Law of Requisite Variety (1956) — "only variety can destroy variety"; so convergence rate = entropy-destruction rate, upper-bounded by the controller's capacity. That ceiling is human verification bandwidth (verification-is-the-new-latency) — which is why 30 rounds of the conversation could only converge there.

Caveat (the other half): these are all propellant; none answers what G is. Setting/rewriting G is steering, grounded in preference measure = the wall → intelligence-analog-civilization-digital-adc.

The eight methods (by timing)

before — script · gate · credentialing
during — metering · isolation
after — audit · selection
inside the executor — incentive-shaping

vibe-linter today = gate + audit. Best neighbor to annex: isolation (a sandbox makes "the door is the only exit" physical). The same eight = civilization's toolkit for governing humans, entered from the other end (law=gate, price=metering, prison=isolation, court=audit, market=selection, education=shaping, license=credentialing, credit-score=reputation). Nearest live research: Redwood's AI Control.

Applied case, 2026-07-17: the utter-manual build scored against all eight — script/gate/audit/isolation built to depth, selection/metering/shaping partial, credentialing the sharpest unused (auditor track records vs the observed auditor-depth-variance failure mode). Record: control-methods-applied.

Up: vibe-linter