Semiont boundary

Type: agentic-system-analyses/types/agentic-system-boundary.md

Boundary and evidence

This run analyses Semiont as a shared knowledge platform serving humans and model-based agents, not just its MCP adapter or an SDK library. Its intended use is building and maintaining source-linked knowledge that can later be retrieved and supplied to generation. The target is in scope as a memory/knowledge/context-engineering system: shipped services connect retained resources and annotations to context assembly, model calls and subsequent writes. The overall evidence tier is code-grounded; architecture prose remains declared design rather than proof of implementation or successful operation. SRC-1, README.md, docs/architecture/KNOWLEDGE-SYSTEM.md, apps/worker/src/semiont_worker/held_job.py.

The platform describes its purpose explicitly:

Semiont is an open, source-grounded semantic knowledge platform for building and maintaining trusted AI knowledge bases and context layers. It gives humans and AI agents a shared workspace and architecture to annotate, connect, enrich, and govern domain knowledge for accurate applications, agents, and workflows. --- README.md

The frozen evidence is the complete Git repository at a920067ac4ea63a57b1855b7d263485fbd22bab4. The checkout's HEAD was checked against the acquired revision. The applicability cutoff is 2026-10-15, the supplied run date, not a claim that a deployed service or later release has this behavior. The acquired source object is retained unchanged above. No additional repository, live documentation, release image, model output or external knowledge-base corpus is frozen.

Functional inclusions and traced entry points

All shipped paths at this commit remain eligible evidence under SRC-1. Functional inclusions are content admission and annotation; durable records and derived views; graph/vector maintenance and repair; retrieval and context assembly; queued model work and generated-content admission; human and SDK consumers; MCP integration; launcher/deployment configuration; shipped agent instructions, prompts and evaluators. Shipped callers are part of the target, not external applications merely because they use an SDK. SRC-1, Cargo.toml, package.json, docs/architecture/ACTOR-MODEL.md, docs/builder/skills/README.md.

Initial implementation inspection establishes a real shipped caller beyond the library API. The Worker service constructs inference clients from configured providers and models; an AgentWorker opens an authenticated SDK transport and claims jobs; claimed work reaches run_job. For a yield job, _yield invokes generation processing, uploads the artifact through client.yield_.resource, commits links/citations and completes the job. The Python SDK's YieldNamespace.resource forwards that upload to its content transport. These are included platform responsibilities, not behavior attributed to an unspecified enclosing agent. SRC-1, apps/worker/src/semiont_worker/service.py, apps/worker/src/semiont_worker/agent.py, apps/worker/src/semiont_worker/held_job.py, packages/sdk-python/src/semiont/namespaces/yield_.py.

made = await process_generation_job(inference, params, on_progress=reports.say, cancelled=job.cancelled)

--- apps/worker/src/semiont_worker/held_job.py

def resource(self, data: PutBinaryRequest) -> Upload:
    """Upload bytes as a new resource: the upload's progress, and the resource created."""
    return self._content.put_binary(data)

--- packages/sdk-python/src/semiont/namespaces/yield_.py

On the record side, yield_create registers the content, derives attribution, constructs the event payload and appends a YieldCreated event. The generation prompt builder consumes retained resource content, graph connections and semantic matches. The Librarian service wires graph/vector readers, embedding and inference clients, and listens for gather/match requests and projection-progress signals. The initial inspection does not yet trace every transport hop between these endpoints. SRC-1, apps/archivist/src/commands.rs, apps/worker/src/semiont_worker/generation/resource_generation.py, apps/librarian/src/semiont_librarian/service.py.

archivist.append(event(YieldCreated::NAME, Some(&resource_id), user, payload))?;

--- apps/archivist/src/commands.rs

The MCP executable is a separate included client: it constructs a SemiontClient over HTTP transports and exposes it through a stdio MCP server. This does not make Semiont itself merely an MCP tool. SRC-1, packages/mcp-server/src/index.ts.

Coverage and partial-loop limit

The boundary kind is complete artifact, partial loop, not whole-system. The repository tree and the shipped paths above were inspected, but material routes through gateway authorization, dispatcher admission, projection repair, Browser/desktop consumers and the full launcher have not yet been traced in implementation. They are not yet inspected, not unavailable and not deliberately excluded. That distinction prevents an initial claim of exhaustive whole-system loop coverage, global access-control enforcement or complete maintenance/recovery behavior. Later jobs may inspect these paths at the same registered commit; the source register's initial paths are not an allowlist.

External dependencies and evidence exclusions

The target's deployment description names container infrastructure and external services including Keycloak/PostgreSQL, NATS, Neo4j, Qdrant and optional local Ollama; its model configuration also permits configured provider endpoints. These are external dependencies, not additional frozen implementations. Their configured use and Semiont-side adapters are in scope; their internal correctness, uptime and security are not established. The launcher itself remains inside the target. SRC-1, apps/launcher/README.md, apps/worker/src/semiont_worker/service.py, apps/librarian/src/semiont_librarian/service.py.

No target execution was performed. The Librarian conformance documentation describes scripted model/embedding stand-ins and real surrounding services. It supplies evaluation design, not an observed run, a measurement of model quality or a causal experiment. External assistant runtimes that load skills or call MCP, third-party models, separately hosted knowledge-base repositories and deployment-specific human decisions are excluded; this prevents conclusions about their reasoning, corpus quality or end-to-end production outcomes. SRC-1, tests/conformance/librarian/README.md, docs/builder/skills/README.md, packages/mcp-server/src/index.ts.

Source register