The scan — reading before believing
Silex starts by reading your repository locally: parsing every source file, extracting classes, methods and the relationships between them — calls, inheritance, HTTP routes crossing language boundaries, raw SQL joins. No code leaves your machine.
The result is a local graph: nodes are code symbols, edges are relationships found in the source, each one traceable to a file and line.
The community map — the real shape of the system
Silex clusters the relationship graph (Leiden algorithm) into communities: groups of classes that actually work together — regardless of folder layout. This is the architecture as it is, not as the src/ tree claims it is. Click a node to see its real member classes.
Ordering domain core
49 classes · cohesion 0.067The Order aggregate's contract surface: repository, domain events (cancelled/shipped/paid), and the integration-event handlers that glue the order saga together.
Members (real class names)Description written by an AI agent reading this community's real code pack — the same client-orchestrated flow you get.
The sagas show up as their own communities — the graph found the choreography steps (stock confirmation, payment) without being told what a saga is. Folder layout says "Ordering.API"; the graph says "five independent slices and two saga legs".
Evidence anchoring — no anchor, no fact
Every claim Silex makes carries an anchor: the file and line range that proves it, with the quoted snippet verified against the source. A claim that cannot be anchored is not stored as a fact — it is flagged as a gap.
hallucinated anchors across 328 agent submissions in a real 520-class enterprise engagement. The gate is enforced at the door: an agent-submitted rule without provenance is rejected.
Confidence — honesty as a number
Confidence expresses how strongly the evidence supports a claim — and it carries provenance: whether the claim came from static analysis, from an LLM pass, or from your own AI agent. Numbers are never invented: an agent-proposed rule shows n/a until a human reviews it, not a made-up score.
Unknown territory — what Silex can't prove, it tells you
The inverse of confidence: the parts of the codebase where evidence was too thin. Instead of guessing, Silex marks them. For a migration or an AI-assisted change, knowing where the map ends is as valuable as the map.
eShop reference scan: 14% flagged as unknown territory by gap_analysis.
State machines — recovered from code, not documentation
Where state lives in the code (enums, status fields, guarded transitions), Silex reconstructs the machine — every transition anchored. The real order lifecycle recovered from eShop (Order.cs:99-131): Submitted → AwaitingValidation → StockConfirmed → Paid → Shipped, with Cancelled reachable from the early states.
Cross-check against section 02: the saga steps (stock confirmation, payment) appear as communities, and the transitions they drive appear here — two independent reconstructions agreeing is exactly what evidence-based archaeology looks like.
Vendor & generated code — classified, never silently deleted
A real codebase carries vendor bundles, minified assets and generated migrations. Silex classifies them out of the way by default — out of the map, out of the packs, out of the cost estimates — but never hides the fact.
One rule protects you: a vendor-looking file that your team has edited (local commits in git history) is treated as your code and stays on the map, marked. Forked jQuery from 2014 with business rules inside? It stays visible.
Human review — the 1.5 layer
Between raw extraction and a canonical domain model sits the review layer: every LLM- or agent-proposed fact is a candidate until a human confirms it. Your AI agent does the legwork through MCP tools — reading community code packs, proposing rules with anchors — and you stay the authority. The community descriptions in section 02 were produced exactly this way.
Generated from a real Silex scan of eShop (MIT) with v0.11.35. The graph, the tables and the state machine are data, not illustration.