The graph should answer questions, not create an aura
A glowing node connected to five other nodes looks important. The useful question is much less theatrical: what does each connection mean, and which source supports it? OpenPlanter is worth studying because it brings an investigation workspace and a knowledge graph into the same interface. For an FLLC research desk, the interesting design problem is making every relationship easy to challenge.
The OpenPlanter repository describes a recursive-language-model investigation agent with a Tauri desktop application and a separate Python command-line interface. It documents heterogeneous dataset ingestion, entity resolution, web research, file operations, shell execution and delegated subtasks. Its desktop graph uses Cytoscape.js, with category filtering, alternative layouts and source documents accessible from graph nodes. Those are documented capabilities, not results of a product benchmark performed for this article.
Give every edge a verb and a receipt
The following is a proposed FLLC publication design, not a claim that it is already implemented in OpenPlanter or Sky Ops. Each visible edge should have a specific verb: published, awarded, supplied, amended, or cited. Avoid an unlabeled line that invites readers to invent its significance. A company mentioned in a document is not necessarily a contractor, and a contractor is not necessarily implicated in misconduct.
Use a fictional procurement example. Example Transit Authority publishes award record EX-104 naming Example Sensor Works as its supplier. A second document amends the delivery date. The graph can contain two organizations and two documents without making any claim about a real person.
| Element | What the proposed record should preserve | | --- | --- | | Organization | Stable source identifier, source spelling and jurisdiction | | Document | Publisher, canonical reference, publication date and retrieval time | | Relationship | Explicit verb, supporting passage and relevant date range | | Review | Accepted, unresolved or contradicted, with a short explanation |
Selecting the award edge should reveal the award record. Selecting the amendment should show what changed. Neither action should require trusting an AI-generated paragraph that has lost its references.
Matching is a decision, not a visual effect
Entity resolution becomes consequential when records use abbreviated or inconsistent names. Suppose one synthetic file says Example Sensor Works and another says ESW. Shared initials are a candidate match, not a completed match. A proposed review should compare stable registry identifiers first, then examine jurisdiction and the source's own explanation of the alias.
Keep unresolved candidates separate. If two records are merged prematurely, every later relationship inherits the mistake. A cleaner graph is not worth erasing ambiguity. The interface should let a reviewer see why a merge was accepted, undo it and identify the dependent connections that now need rechecking.
This also changes the meaning of confidence. Rather than inventing a percentage because two names look similar, record the evidence: matching official identifier, explicit alias statement, conflicting jurisdiction or missing corroboration. Those reasons help a human decide what the source actually permits.
Separate the working investigation from the public exhibit
A public-facing graph and an agent's working directory should not be the same surface. OpenPlanter's documented access to files, shell tools and external services makes that separation especially important. The repository describes an investigation agent, not a promise that arbitrary visitors can safely operate it inside an unrelated public website.
For FLLC, a safer proposed publishing path has three boundaries. An isolated research workspace handles permitted source material. A review step accepts a small set of supported claims. The public site renders that accepted export without inheriting the agent's credentials or execution tools.
That separation creates useful failure behavior. A stalled research run does not erase yesterday's approved article. A malformed document does not become a command. A publication can be corrected without exposing private working notes. These are proposed engineering requirements; this review did not deploy a worker or execute OpenPlanter.
Design the overview around uncertainty
A top-down global view can still be visually compelling without pretending to know everything. Give the reader a region filter, document timeline, category legend and a direct route from the map to the supporting record. A selected node should make its provenance more prominent, not merely animate the camera.
Do not turn an organization's reference location into a claim about a person's whereabouts. Do not infer criminal affiliation from proximity, a shared supplier or a repeated name. For reporting on public allegations, attach the publisher and legal status to the statement. A charge, a conviction and an analyst's open question are different kinds of information.
A useful overview makes unknowns visible. Unmapped records can remain in a list. Unresolved relationships can remain outside the published graph. A country label can identify reporting context without depicting a group's territory or live operational location.
A first evaluation that produces something concrete
Start with the fictional four-record example, including one ambiguous alias and one correction. The expected deliverable is an inspectable export, not a claim that a real investigation has succeeded. Ask a second reader to reconstruct each accepted relationship from its cited record and identify the unresolved match without help from the author.
Then change one source field. The reviewer should be able to see which assertion changed, which remained supported and which requires another decision. An attractive graph that cannot explain that difference is not ready to publish as evidence.
Source and publication scope
OpenPlanter upstream repository, documentation reviewed September 22, 2026. OpenPlanter was explicitly requested for FLLC coverage; current GitHub star membership was not established. This article proposes an integration boundary and evaluation exercise. It does not claim a completed FLLC integration, a live investigation, independently verified entity matches or access to private records.