Compounding substrate
Taxonomy ratification 2026-05-09: split architecture/ into substrate/+patterns/+architecture/; retire company/+help/
@@ -0,0 +1,144 @@ --- schema: foundry-doc-v1 title: "The Compounding Substrate" slug: compounding-substrate category: substrate type: topic quality: complete short_description: "The Compounding Substrate is the architectural pattern PointSav builds and stewards, combining five structural properties to produce a platform where every operational interaction generates training signal that compounds across all tenant deployments." status: active bcsc_class: public-disclosure-safe last_edited: 2026-04-30 editor: pointsav-engineering cites: - ni-51-102 - osc-sn-51-721 paired_with: compounding-substrate.es.md --- Every operational interaction on the PointSav platform generates training signal that compounds across all tenant deployments — producing an AI system that improves continuously without any tenant surrendering ownership of their data. **The Compounding Substrate** is the architectural pattern that makes this possible: open, forkable platform code; a deterministic data layer that functions independently of any AI compute; and AI added as an optional layer any tenant can compose in or out. A curator — PointSav — periodically rolls accumulated signal into improved base models that flow back to all deployments without disrupting customer data ownership. The pattern applies the open-source foundation model (Apache Software Foundation) combined with the commercial distribution model (Red Hat) to AI substrate: the platform becomes open commons, and value migrates up to operations, integration, and a federated marketplace. This article describes the pattern, names the five properties, and explains the value-chain inversion that makes the model durable. ## Substrate Definition A Compounding Substrate is an AI-substrate architecture where: 1. The substrate code is open and forkable. 2. The deterministic data layer functions independently of any AI compute. 3. AI is added as an **Optional Intelligence Layer** that any tenant can compose in or out. 4. Every operational interaction generates training signal that compounds across the substrate's deployments. 5. A curator (PointSav) periodically rolls accumulated signal into improved base models that flow back to all deployments without disrupting customer data ownership. ## Structural Platform Properties Each property is a structural claim. Each names a specific reason hyperscalers cannot replicate it without dismantling their own business model. ### 1. Customer Stack Custody Every customer owns their full stack: data, compute, adapters, and deployment composition. The substrate (code plus base model) is open under permissive licence. Data and adapters are the customer's intellectual property. Hyperscaler structural gap: their business is monetising the substrate as rented service. Substrate ownership erodes lock-in — the foundation of their billing model. ### 2. Decoupled Intelligence Tier The data and deterministic-processing rings function fully without the AI ring. Customers, community members, regulated buyers, and air-gapped sites can deploy a fully-functional PointSav data platform with zero AI compute. AI is additive value, not table stakes. Hyperscaler structural gap: their AI products tightly couple AI compute to data services. Decoupling them eliminates AI-compute revenue from any deployment that opted out. ### 3. Dynamic Compute Routing `service-slm` is the platform's sole access-control gateway (the Doorman) — transparently routing among three compute tiers: local OLMo 3 7B on the customer's machine, multi-cloud burst (Cloud Run / RunPod / Modal / customer GPU), and external API (Claude / Gemini / GPT). The customer does not pick the tier; request shape and budget caps do. Hyperscaler structural gap: each tier in their world is a separate billing relationship; their ecosystem does not span competitors' frontier models. They cannot abstract this routing. ### 4. Privacy-Preserving Federation Customers opt in to a federated LoRA marketplace (privacy-preserving aggregation per the SDFLoRA / FedEx-LoRA / HeLoRA research lineage). Every customer's improvements lift the substrate. The customer's own data never leaves; only adapter weights and KV cache blocks (without source data) flow into the federation. Hyperscaler structural gap: per-tenant billing and compliance posture make cross-tenant pooling structurally illegal in their model. They cannot operate a true federation. ### 5. Curated Substrate Advancement OLMo 3 base flows to a PointSav continued-pretraining variant, released as the substrate for subsequent deployments. Each year's curated commons feeds the next year's base. By 2030, the federation-trained base is intended to be competitive with frontier proprietary models on the federation's domains. Hyperscaler structural gap: they cannot let customers' data train a base model the customer subsequently owns. That destroys the lock-in that justifies their margins. ## Structural Value-Chain Inversion Hyperscalers' value chain depends on the customer remaining on the vendor's substrate. The Compounding Substrate's value chain depends on the customer compounding *off* the vendor's substrate. The two business models are mathematically opposed; one cannot adopt the other without dismantling itself. This is the asymmetry that makes the pattern durable. A hyperscaler that copied the substrate-ownership property would erode its own lock-in. A hyperscaler that copied the optional-intelligence property would lose AI-compute revenue on every deployment that opted out. A hyperscaler that copied the federated-compounding property would breach its own per-tenant compliance contracts. ## Platform Stewardship Role Not vendor. Not gatekeeper. **Steward.** - Steward of the protocol (governs the Doorman specification, runs the Constitutional Convention process aligned with the fundamental physics of 2030 hyperscaler infrastructure). - Steward of the base model (publishes the continued-pretraining variant, contributes upstream to OLMo when relevant). - Steward of the marketplace (operates the federated LoRA pool, takes a percentage of revenue-share LoRAs). - Operator-of-record (sells appliances plus integration plus support). - Reference customer (the PointSav development environment plus Woodfine Management Corp. — proof the pattern works). The substrate is open commons; value migrates to operations, integration, and the LoRA library marketplace. ## Continuous Compounding Cycle Every action produces data; every data produces knowledge; every knowledge improves future actions. The loop runs continuously, in every tenant deployment, federated through the commons. For an operator evaluating the platform: each month of production use makes the AI layer materially better — without additional investment or data sharing beyond what the operator chose at onboarding. ``` operator + assistant does work ↓ produces git commits + file edits + session logs + conversation turns ↓ ingested by service-fs[tenant] ← WORM ledger ↓ parsed by service-extraction[tenant] ← deterministic ↓ writes structured to service-content[tenant] ← knowledge graph ↓ indexed by service-search[tenant] ← full-text index ↓ queried by (when AI active) service-slm ← Doorman; routes among 3 compute tiers ↓ trains (periodically) LoRA adapters ← per-tenant skill packs ↓ contributes (opt-in, federated) federated LoRA pool ← commons benefit ↓ rolls into (annually) base-model continued-pretraining ← curated by PointSav ↓ ships in appliance update ← every customer benefits ↓ used by operator + assistant in next session ← loop closes, compounded ``` ## 2030 Operational Trajectory Per `[ni-51-102]` continuous-disclosure language and the forward-looking discipline of `[osc-sn-51-721]`, the trajectory described below is `planned` and `intended`, not declarative-future. The shape is in place; the operational throughput matures over time. By 2030, the Compounding Substrate aims to produce: - A base model competitive with hyperscaler frontier on regulated SMB tasks. - A federation of one hundred-plus customers, each owning their full stack, each contributing to and benefiting from the commons. - A protocol stack versioned twice through Constitutional Convention and validated in production. - A market position where regulated SMB industries — small clinics, mid-sized law firms, regional financial advisors, real-estate operators — have standardised on this pattern because their compliance posture requires it and the substrate's design delivers it. The pattern does not displace hyperscalers in volume; they retain the unregulated long tail of cloud AI. The trajectory captures the regulated SMB market that hyperscalers cannot economically reach. ## See Also - [[apprenticeship-substrate]] - [[3-layer-stack]] - [[worm-ledger-architecture]] - [[sovereign-ai-routing]] - [[language-protocol-substrate]]