Populi work-type placement policy matrix
This page is the canonical policy matrix for first-wave personal-cluster placement boundaries. It expresses intent aligned with ADR 017, ADR 018, and ADR 009. Shipped behavior may lag this matrix until roadmap phases complete; for current wire semantics use mens SSOT and unified orchestration.
Matrix
| Work class | Local single-node | Trusted LAN personal cluster | Overlay-WAN personal cluster |
|---|---|---|---|
| Agent task (non-GPU critical) | Allowed (default) | Allowed (gated) | Allowed (gated, conservative timeout) |
| GPU inference task | Allowed | Allowed (lease-gated) | Allowed (lease-gated, latency caveats) |
| GPU training long-run | Allowed | Allowed (explicit profile and checkpointing) | Not default; pilot-only explicit opt-in |
| Distributed collectives | Optional local/LAN only | Pilot-only with strict topology constraints | Out of scope by default |
Meaning of columns
- Local single-node: default developer and single-container flows; no Populi required.
- Trusted LAN personal cluster: nodes under a single operator or agreed trust domain, reachable on a private LAN with stable RTT; TLS/mTLS and bearer policy per ADR 008.
- Overlay-WAN personal cluster: user-owned nodes joined across the public internet via VPN/wireguard-style overlay or equivalent; control-plane reachability may be decoupled from bulk artifact paths (see overlay runbook).
Policy notes
- Hosted donation or multi-tenant public GPU marketplace remains out of scope for this wave (ADR 009).
- Cloud provider dispatch (
vox mens train --cloud, provider nodes) is a separate execution surface from Populi mesh until an explicit convergence ADR merges them; see Mens cloud GPU strategy. - Promoting WAN distributed training to a default supported path requires a new ADR and updated matrix row(s).
Gating vocabulary
- Gated: requires explicit config / policy / feature enablement; not implied by joining a cluster.
- Lease-gated: requires authoritative lease semantics per ADR 017 once implemented; until then treat remote GPU paths as experimental only.
- Pilot-only: documented rollout and kill-switch validation required before production reliance.
Related documentation
- Populi GPU mesh implementation plan 2026 — phased delivery (roadmap); Phase 5 tasks
p5-placement-policy,p5-queued-capacity-rebalance,p5-gang-nccl-pilotcover unified placement, queued replanning on capacity changes, and collective pilot bounds. - Protocol convergence research 2026 — transport and delivery-plane context.