Eigenius boundary

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

Boundary and evidence

Eigenius is in scope as a memory/knowledge/context-engineering system: a typed knowledge platform with model-assisted prose formalization and an agent-facing tool surface. It is not classified here as an enclosing autonomous agent runtime. Its declared intended use is verifiable AI in science and engineering, built on a typed, versioned knowledge graph (SRC-1, README.md). That statement is intent, not evidence of scientific reliability:

An open-source platform for verifiable AI in science and engineering, built on a typed, versioned knowledge graph. --- README.md

The executable source supports the knowledge-system classification. The shipped CLI constructs EncodingFormalizer, supplies a persistent RocksDB backend when --db is set, and injects the formalizer into the kernel server (SRC-1, cli/src/main.rs). The formalizer reloads a document branch, obtains recorded model decisions or live decision arms, runs the encoding pipeline, emits an artifact, and commits recorded draws. Thus retained content can influence a later formalization; this is not merely a stateless returning computation (SRC-1, crates/eigenius-encoding/src/formalize.rs):

let branch = doc_branch_head(&backend, &req.doc_id)?; --- crates/eigenius-encoding/src/formalize.rs

let draws_committed = commit_draws(&backend, &req.doc_id, doc_layer, draw_resources_all)?; --- crates/eigenius-encoding/src/formalize.rs

Model calls are part of the shipped implementation, not just an external agent's possible usage. The kernel's Anthropic client constructs a model request containing a user prompt and schema-constrained output, then sends it to the provider. Separately, the Deno entry point registers real or mock CompleteText and CompleteJson components (SRC-1, kernel/src/dcg/anthropic_client.rs, orchestration/src/main.ts). The provider call itself is inspectable:

let resp = reqwest::Client::new() .post(API_URL) .header("x-api-key", api_key) .header("anthropic-version", ANTHROPIC_VERSION) .header("content-type", "application/json") .json(&body) .send() .await .map_err(|e| format!("anthropic request failed: {e}"))?; --- kernel/src/dcg/anthropic_client.rs

Functional inclusions and inspection extent

The target is the Eigenius platform shipped in this repository, not just its Rust library interface. The CLI, MCP server, SDK, browser notebook, drivers, ontologies, maintenance machinery and evaluators remain in the functional boundary. A shipped caller is not treated as an external host. The entire Git repository at the registered revision is available for later source inspection; the paths below record initial coverage, not an evidence allowlist.

The boundary kind is complete artifact, partial loop. The complete tracked tree was inventoried, and material entry points were traced into formalization, persistence/replay and caller-controlled admission. However, the initial inspection has not traced every maintenance, retrieval, institution and evaluation path. These are uninspected available sources, not access gaps or deliberate functional exclusions. Consequently this boundary does not certify whole-system operational coverage.

One consequential caller distinction is already supported: generation is not identical to assertion into the serving chain. The notebook obtains the emitted artifact and invokes load only when its land setting requests it (SRC-1, notebooks/src/runtime/notebookStore.ts):

if (cell.land) { const outcome = await eigen.load(artifact, { contentType: "application/x-esl", autoCommit: true, }); --- notebooks/src/runtime/notebookStore.ts

This optional path and the CLI's construction of the served formalizer are included platform responsibilities. Treating them as excluded applications would incorrectly narrow the target to the kernel API.

Frozen evidence and exclusions

The frozen evidence is repository https://github.com/eigenius/eigenius at 836498ebd99733067d45aa6f7b09d4ff5035dd99. The acquired checkout's HEAD matched this pin, and git status --short reported no changes. The analysis cutoff is 2026-10-15; later commits, live documentation changes and external linked papers are not evidence in this run. The overall tier is code-grounded. Source inspection establishes implemented paths and declared contracts, not their successful execution.

External model services, independently installed agent hosts, scientific libraries, container engines, external datasets, image registries and deployed databases are dependencies, not additional frozen sources. Repository-side adapters, configuration, workers and importers remain included. Provider behavior, external library correctness, actual image contents, live deployment state and real data coverage cannot be established without separately registered evidence. No such evidence was supplied or acquired. The source mentions a lexicon DB snapshot and WordNet dictionary as prerequisites for the inspected comparison test (SRC-1, crates/eigenius-encoding/tests/served_formalization.rs); that test's source is evidence of an evaluator contract, not a supplied snapshot or observed passing run. Target code, tests and demos were not executed.

Source register