LLM-Wiki-v3 boundary

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

Boundary and evidence

LLM-Wiki-v3 is in scope as a memory/knowledge/context-engineering system. Its intended use is a domain-configurable Markdown knowledge store for human browsing and agent context retrieval. The frozen implementation issues model calls to extract proposed pages and supplies retrieval results from retained pages; it is not an enclosing general-purpose agent runtime. The boundary is the whole shipped repository system, including its operator CLI and curated seed driver, not only the extraction or retrieval APIs (SRC-1, README.md, src/wiki/cli/__init__.py, scripts/seed_corpus.py).

It serves two readers from one store:

1. Humans — browsing in Obsidian, a CLI, or any markdown viewer.
2. LLM agents — retrieving context via an MCP server. --- README.md

That passage establishes intent, not a shipped MCP service. The reviewed tree and CLI implement ingestion, review, rebuild, query and evaluation beyond the README's Phase 0 status. The Makefile still labels several later surfaces as stubs. This analysis distinguishes those declarations from executable paths rather than treating either status text as a complete implementation inventory (SRC-1, README.md, Makefile, src/wiki/cli/__init__.py).

The included state loop is source-byte retention → chunk extraction using the root ingest instructions → quote location and validation → pending proposals and manifests → interactive curated admission or rejection → namespace Markdown pages → index rebuilding and subsequent query/evaluator consumption. Ingestion can escalate between extractors after validation failure. The separate curated seed driver directly constructs, validates and writes retained pages and stores raw sources; it is part of the target, not an excluded caller. Evaluation includes retrieval adapters, saved baselines, source-span-based ingest metrics and model-generated candidate queries for human review (SRC-1, src/wiki/ingest/pipeline.py, src/wiki/extract.py, AGENTS.md, src/wiki/review.py, scripts/seed_corpus.py, src/wiki/retrieve/store.py, src/wiki/retrieve/rebuild.py, src/wiki/retrieve/query.py, src/wiki/eval/harness.py, src/wiki/eval/ingest_eval.py, src/wiki/cli/__init__.py).

def _system_prompt(root: Path) -> str: agents = root / "AGENTS.md" return agents.read_text(encoding="utf-8") if agents.is_file() else "" --- src/wiki/ingest/pipeline.py

    system=context,
    tools=[PROPOSE_TOOL],
    tool_choice={"type": "tool", "name": "propose_pages"},

--- src/wiki/extract.py

dest = _dest_for(root, page_path)
dest.parent.mkdir(parents=True, exist_ok=True)
page_path.replace(dest)

--- src/wiki/review.py

def load_pages(root: Path, include_superseded: bool = False) -> list[PageRecord]: --- src/wiki/retrieve/store.py

These passages anchor the classification and boundary: shipped instructions enter a model call, proposals have a retained-state admission path, and a later loader consumes the retained store. They do not establish that every prose policy is enforced or that the interactive review revalidates edited proposals.

Top-level tracked coverage

Every top-level tracked entry is included below. “Inspected” describes initial boundary work, not an assertion that every included file has been read. Uninspected implementation details remain available at the same pin for later analysis; they are not access gaps or exclusions. Paths in every row refer to SRC-1.

Dependencies, exclusions and evidence limits

The applicability cutoff is 2026-10-18; the sole evidence revision is 267541db838cbca757a112bd0cfa47fa61eb5fb2. No moving branch, later revision, external document or additional repository is admitted. HEAD matched that pin and the checkout was clean at inspection. The overall tier is code-grounded: implementation is available for the shipped loop, supplemented by doctrine and source-authored reports. No target code, tests, examples, model call or evaluation was executed.

External dependencies include Anthropic's client/service for extraction and candidate-query generation, Python parsing/validation libraries, Git for optional review commits and an optional editor for interactive review. Optional local dense retrieval/reranking dependencies are declared separately. Their external implementation, model weights and live responses are excluded; this prevents conclusions about provider reliability, model truthfulness, downloaded-model behavior or deployment performance. The target's dependency declarations and integration call sites remain included (SRC-1, pyproject.toml, src/wiki/extract.py, src/wiki/cli/__init__.py, src/wiki/review.py).

Human judgment and any downstream agent's reasoning after retrieval are outside this implementation boundary. The included CLI owns its interaction and returned retrieval fields, not an unprovided consumer's prompt construction, answer generation or long-term agent outcome. MCP, context compilation and autonomous admission are mentioned as later phases; they are not silently supplied by an imagined host. This prevents treating the intended agent-facing service or autonomous write loop as established shipped behavior (SRC-1, README.md, Makefile, src/wiki/cli/__init__.py).

The tracked raw-store placeholder, ignored raw bodies, absent derived indexes and empty audit log prevent reproduction of stored-source lineage, inspection of an actual index build or reconstruction of an ingestion/admission run from this checkout alone. These are artifact gaps, not missing implementation paths. The frozen baseline report remains attributed reported operation, not a causal experiment or an independently observed run (SRC-1, .gitignore, raw/.gitkeep, audit/log.jsonl, evals/baselines/baseline-20260529T144104Z.md).

Source register