StandMeet × the eight controls

Read as a cybernetic machine, StandMeet is a control apparatus that destroys a stochastic LLM's variety and converges it onto one goal G = "answer as the owner would — grounded, safe, in voice." Mapped onto the eight controls (control-methods):

control	StandMeet
gate	✅✅ corpus ACL · capability gating · _meta trusted scope
isolation	✅✅ sandbox (bwrap, per-cap socket)
metering	✅ compaction · iteration/timeout caps · quotas
audit	✅ but shallow (transcripts; no policy-version / failed-branch capture)
incentive-shaping	〜 static 7b (persona / skills); no auto-loop
selection	〜 meta-layer only — best-of-n prompts in experiments, not per-answer
credentialing	〜 offline eval; not yet a promotion gate
script	✕ by design (free loop, not a workflow)

Central variety-destroyer = corpus grounding — retrieval + citations force the answer into the owner's actual content, not the LLM's freedom.

Shape: strong on the safety half (gate / isolation / metering — before+during), thin on the quality half (runtime selection / deep audit / shaping-loop — after+inside). Right order: cage first, improve second. Two designs fill the second half — agent-as-injectable-driver (landed: backend/agentcore/ Bridge/Driver; inject-and-launch still test-only) = selection + credentialing acting on shaping; the transcript ledger (eval-harness-on-prod-loop, still open) = deeper audit + the shaping loop.

The irreducible part: the controls only accelerate toward G; G itself is set by the owner's corpus + persona — the steering, the un-automatable half → intelligence-analog-civilization-digital-adc.

Same machine, three scales: vibe-linter (governs AI agents), StandMeet (governs the AI that represents you), and the vault holding these notes (governs your own note-taking, via its pre-commit gates) are all one pattern — mechanical constraint / gatekeeper / variety-destruction toward an owner-set goal. StandMeet's existing name for it: an institution → institutions-operate-on-the-joint-distribution.