os-orchestration: The Stateless Aggregation Layer
editorial(infrastructure): de-narrate os-orchestration-stateless-hub (Track-B) — re-verified os-orchestration/src/lib.rs is still a 2-line scaffold, app-orchestration-slm still owns real Tier B metering (allocation/fleet/metering/yoyo_proxy.rs, port 9180); dropped 3 narration blocks, stated facts directly per register-documentation.yaml; fixed missing index_group (fleet-and-edge-deployment); EN+ES
@@ -10,14 +10,13 @@ bcsc_class: forward-looking language: en paired_with: os-orchestration-stateless-hub.es.md category: infrastructure index_group: fleet-and-edge-deployment status: active quality: complete last_edited: 2026-08-03 --- **Correction (2026-08-02, verified against canonical `origin/main`):** the elaborate capability-broker/federation/statelessness-enforcement architecture described below doesn't exist. The real `os-orchestration/src/lib.rs` is a 2-line placeholder scaffold (`"SYSTEM EVENT: os-orchestration scaffold verified."`) — no capability brokers, no federation model, no commercial-tier Ring logic. Its own README even names a different aggregation crate, `app-interface-command`, not the `app-orchestration-exchange`/`app-orchestration-market` names this article uses elsewhere. The "Yo-Yo GPU Broker" section is closer to accurate — `app-orchestration-slm` is real (`crates/orchestration-slm/src/{allocation,fleet,yoyo_proxy}.rs`, port 9180) — though it's framed differently in real code (a chassis connecting `service-slm` Doorman instances, not "a pool of GPU capacity from PPN + external providers"). **Resolved 2026-08-03**: the whole article is re-hedged to planned/intended language below; the Commercial Model section's Tier B revenue/metering claim — which this correction had not yet flagged as a separate defect — turned out to be misattributed to `os-orchestration` entirely and is corrected to name `app-orchestration-slm` (the real crate that actually implements per-tenant usage metering, `crates/orchestration-slm/src/metering.rs`, confirmed on `origin/main`) as the system that owns that model. The PointSav platform is designed around a deliberate architectural boundary: an aggregation layer intended to coordinate work across [[totebox-archive|Totebox Archives]] while holding no customer data, storing no keys, and writing nothing to any [[worm-ledger-architecture|WORM ledger]]. This layer is planned as `os-orchestration` — today a registered but unimplemented scaffold crate, not a running system (see the correction above). The PointSav platform is designed around a deliberate architectural boundary: an aggregation layer intended to coordinate work across [[totebox-archive|Totebox Archives]] while holding no customer data, storing no keys, and writing nothing to any [[worm-ledger-architecture|WORM ledger]]. This layer is planned as `os-orchestration` — today a registered but unimplemented scaffold crate, not a running system. Understanding what `os-orchestration` is intended to be requires first understanding what it is designed not to be. It is not planned as a database, not a credential store, and not a custodian. Every archive in the PointSav network is designed to maintain its own isolated state — its own WORM audit trail, its own key material, its own DataGraph segment. `os-orchestration` is planned to sit above that layer as a coordinator: routing requests, enforcing capability boundaries, and brokering cross-archive work without ever touching the underlying data. @@ -41,11 +40,11 @@ Practically, this design intends for `os-orchestration` to carry no obligation u The commercial structure of the PointSav platform is intended to follow a Rings model. The first two planned Rings of capability — the foundational services every Totebox would require to operate — are intended to be provided free of charge with each Totebox provisioning. This is meant to cover the local inference tier (Tier A), basic DataGraph access, and the standard service stack. **This section previously misattributed a working revenue and metering system to `os-orchestration`. It does not have one — `os-orchestration` is an unimplemented scaffold (see the correction above). The real Tier B usage-based revenue and metering model belongs to a different, real crate: `app-orchestration-slm`.** `app-orchestration-slm` is a working "Yo-Yo broker chassis" (confirmed on canonical `origin/main`, port `:9180`) connecting Totebox Archives' `service-slm` Doorman instances to a shared Yo-Yo GPU fleet. It implements real per-tenant cost metering (`crates/orchestration-slm/src/metering.rs`): each inference request's cost is computed from measured inference time and a configured hourly USD rate, and recorded to a per-tenant ledger that a Tier B customer is billed against. Tier A — the free, local inference tier — is unaffected by this and requires no metering. If `os-orchestration` is ever built as an aggregation/federation layer, it may in principle route requests toward `app-orchestration-slm`'s Tier B service, but the metering, billing, and revenue-accrual logic itself lives in `app-orchestration-slm`, not in `os-orchestration`. The Tier B usage-based revenue and metering model belongs to a running crate, `app-orchestration-slm` — not to `os-orchestration`, which is an unimplemented scaffold and has no revenue logic of its own. `app-orchestration-slm` is a working "Yo-Yo broker chassis" (port `:9180`) connecting Totebox Archives' `service-slm` Doorman instances to a shared Yo-Yo GPU fleet. It implements real per-tenant cost metering: each inference request's cost is computed from measured inference time and a configured hourly USD rate, and recorded to a per-tenant ledger that a Tier B customer is billed against. Tier A — the free, local inference tier — is unaffected by this and requires no metering. If `os-orchestration` is ever built as an aggregation/federation layer, it may in principle route requests toward `app-orchestration-slm`'s Tier B service, but the metering, billing, and revenue-accrual logic itself lives in `app-orchestration-slm`. ## The Yo-Yo GPU Broker `app-orchestration-slm` is the real, working [[yoyo-compute-substrate|Yo-Yo GPU broker]] chassis (see the Commercial Model correction above), implementing on-demand GPU allocation and per-tenant metering today. When a Totebox Archive's `service-slm` Doorman submits an inference request that exceeds local Tier A capacity — either because the model is too large for the local hardware or because concurrent load has exhausted available compute — the request is routed to `app-orchestration-slm`'s chassis. `app-orchestration-slm` is the real, working [[yoyo-compute-substrate|Yo-Yo GPU broker]] chassis, implementing on-demand GPU allocation and per-tenant metering today. When a Totebox Archive's `service-slm` Doorman submits an inference request that exceeds local Tier A capacity — either because the model is too large for the local hardware or because concurrent load has exhausted available compute — the request is routed to `app-orchestration-slm`'s chassis. The chassis connects to a Yo-Yo GPU fleet and routes the inference request, returning the result to the originating archive. The archive does not need its own GPU hardware for Tier B workloads; the chassis provides elasticity, and records the cost of each request against the requesting tenant. @@ -61,6 +60,6 @@ For the same reason, `os-orchestration` is designed never to be the system of re ## Design Anchors The intended design of `os-orchestration` draws on several architectural positions. The planned commercial structure, including the Rings model, sets the frame within which `app-orchestration-slm`'s real on-demand GPU allocation and metering mechanism operates today — see the Commercial Model correction above for how that split actually works. The data marketplace provisions, if built, are intended to govern the per-org capability boundaries that would prevent cross-organizational data leakage. A WORM audit obligation is intended to apply to archives and to deliberately exclude the aggregation layer. The browser-facing orchestration surfaces (`app-orchestration-exchange`, `app-orchestration-market`) are intended to operate as the customer-visible face of the aggregation tier, once built. The intended design of `os-orchestration` draws on several architectural positions. The planned commercial structure, including the Rings model, sets the frame within which `app-orchestration-slm`'s real on-demand GPU allocation and metering mechanism operates today, described above. The data marketplace provisions, if built, are intended to govern the per-org capability boundaries that would prevent cross-organizational data leakage. A WORM audit obligation is intended to apply to archives and to deliberately exclude the aggregation layer. The browser-facing orchestration surfaces (`app-orchestration-exchange`, `app-orchestration-market`) are intended to operate as the customer-visible face of the aggregation tier, once built. Together these positions describe a platform where coordination is intended to be separated from custody. The design goal is for the aggregation layer to be effective precisely because it does not hold what it routes — though today, only `app-orchestration-slm`'s piece of this picture is real.