Institutional small language model
An AI request that leaves the building cannot be audited and cannot be recalled. The moment institutional intent reaches a frontier model in another company's cloud, the organization has surrendered both the record of the decision and control over it.
`service-slm` is the language-model service of the PointSav family. It is deliberately a Small Language Model — quantised, narrow, fast — and its job is not conversation but semantic translation: turning institutional intent into deterministic outputs. The service runs in three compute tiers, and every inference call — local, burst, or external — transits the [Doorman](/wiki/doorman-protocol) audit boundary, where each prompt and completion is captured to the per-tenant [ledger](/wiki/worm-ledger-design) before the response returns.For a regulated buyer the consequence is concrete. No AI decision is unlogged, and no request reaches a third-party API without crossing a boundary the operator controls. This article covers the four operations, the three compute tiers, the Doorman boundary, and why a small model is a structural choice rather than a cost compromise.
What service-slm does
The service is invisible — there is no chat window, and the operator never types into service-slm directly. The surface above it presents a structured workflow; service-slm is the silent intermediary. It performs four operations, in order of increasing institutional weight.
| Operation | Inputs | Output |
|---|---|---|
| Semantic command parsing | English intent from the F8 Terminal | Binary UDP command for service-udp |
| Gravity verification | 50-word Gravity Vector from service-content | VALID or REJECT single token |
| Socket assignment | Entity bundle from service-extraction + Chart of Accounts | Sovereign-ID with Chart-of-Accounts socket |
| Theme suggestion | Recurring patterns the Gravity Engine flags | Proposed new entries to the Themes Seed Vault, for operator approval |
The three compute tiers
The same service-slm interface adapts to the host hardware through three execution modes.
| Tier | Where it runs | Model size | Use case |
|---|---|---|---|
| Local | Operator workstation or os-totebox with at least 16 GB RAM |
1B–7B-parameter quantised model loaded locally | Sovereign Iron Vault — institutional customers; no cloud egress |
| Elastic burst | Operator-provisioned ephemeral GPU node | Larger model on rented hardware; data tunnelled over an encrypted link | Cost-optimised heavy batch processing; the node is torn down after the run |
| External API | Licensed third-party API endpoint | Frontier model | Last-resort routing for tasks where local capacity is insufficient |
The Doorman boundary
The Doorman is the audit-routing checkpoint between service-slm and the rest of the system. Every prompt and every completion is captured before the response returns to the caller. The audit trail lives in the local per-tenant ledger and forms the institutional record of every AI decision.
The Doorman exists for three reasons.
- Regulatory. ISO/IEC 42001, the AI management-system standard [^1], requires an immutable log of AI-assisted decisions.
- Operational. A self-healing system needs a corpus of its own past behaviour; the Doorman captures it.
- Sovereign. No request reaches a third-party API without passing through a local boundary the operator controls.
Model selection
The canonical local model is from the OLMo family, which ships with fully open weights and training-data documentation [^2]. Open weights and documented training data are a prerequisite for continued pre-training on an operator's own corpus — the long-term path to a domain-specialised institutional model.| Profile | Model | RAM target |
|---|---|---|
| Edge | OLMo-2-0425-1B-Instruct | ~2 GB |
| Standard | OLMo-3-1125-7B-Think-Q4_K_M | ~6 GB |
Why a small model
A frontier-scale model imposes three costs `service-slm` cannot accept: it requires cloud egress, it consumes tens of gigabytes of RAM, and it cannot be audited in any meaningful sense. A 1B-parameter quantised model is sufficient for the one narrow task — translating institutional English into deterministic outputs — and fits inside the cost envelope of a low-cost cloud node alongside a Totebox.Specialisation, not scale, is the design principle.
See also
- service-content — the upstream Gravity Engine; primary caller of service-slm for gravity verification
- os-network-admin — the F8 Terminal where semantic command parsing originates
- totebox-os — the Totebox that hosts service-slm in Sovereign Iron mode
- SYS-ADR-07 — structured data never routes through AI; service-slm implements this boundary
- doorman-protocol — the Doorman audit-routing protocol in detail