Building a knowledge-graph research vault

Every knowledge-graph wiki I’ve built — SignalTrace, the insider-signal research programme, this CV wiki — came out of the same loop. A research brief on one end, an Obsidian vault on the other, an iterative discovery pass in the middle that knows when to stop.

The shape

research-brief.md
  │
  ▼
┌─ loop ──────────────────────────────────┐
│                                         │
│  read goal + existing vault state       │
│         │                               │
│         ▼                               │
│  propose next entities / questions      │
│         │                               │
│         ▼                               │
│  pull evidence, write entity notes      │
│         │                               │
│         ▼                               │
│  update graph, update open-questions    │
│         │                               │
│         ▼                               │
│  convergence check: new evidence?       │
│         │                               │
│         └── yes ──► continue            │
│         │                               │
│         └── no, or iter_cap hit ──► done│
│                                         │
└─────────────────────────────────────────┘
  │
  ▼
vault/
├── index.md
├── People/
├── Organisations/
├── Events/
├── Concepts/
├── open-questions/
└── GRAPH_REPORT.md

  │
  ▼ (quartz build)
publishable wiki

The research brief

Plain markdown. No schema. States:

Goal — what the investigation is trying to answer
Seed entities — named people, companies, events, or concepts to start from
Source constraints — any stated source preferences (only public filings, only news after date X, etc.)

The brief is the contract. Drift from the brief shows up in the convergence check and prompts a re-read of the goal, not a vault rewrite.

Entity taxonomy — 5–7 folders, fixed up front

The decision that matters most early is the folder taxonomy. Once you commit to People / Organisations / Events / Concepts / Vehicles / Symbols, every wikilink in the vault references one of those categories. Changing the taxonomy two weeks in is a major refactor.

For financial / OSINT topics the SignalTrace taxonomy has held up well. For technical research — this CV wiki itself — the taxonomy shifts to Skills / Projects / Decisions / Notes / Playbooks / Influences / Open Questions. Different domains, same principle: fixed categories, consistent folder per entity type.

The iteration loop

Each iteration does three things:

Propose. Read the goal, the existing vault state, and the open-questions/ folder. Pick the next 3–5 entities or questions that would move the investigation forward. Write the proposal to the session log.
Pull and write. For each proposed target, search sources, extract evidence, write an entity note. Every note ends with a “Sources” block citing where each claim came from.
Update the graph. New entity notes get wikilinked from their neighbours. Questions that got answered move from open-questions/ to the relevant note. Questions that surfaced during the iteration get added.

The convergence check

A loop that never terminates is a loop that can’t ship. The convergence check at the end of each iteration asks one thing:

Did this iteration produce new evidence, or did it mostly re-tread existing ground?

If three iterations in a row answer “mostly re-tread”, the investigation has converged. Write the GRAPH_REPORT.md summary and stop. There’s also a hard iter_cap (typically 20–30) as a second-line kill switch.

Knowing when to stop is the non-obvious part. A vault that keeps growing forever is doing two bad things: diluting signal and burning research hours. Convergence is a deliverable.

Publishing through Quartz

Once the vault converges, Quartz 4 renders it as a static site. Wikilinks resolve natively. The graph view becomes the reader’s actual navigation surface — graph first, prose second.

scripts/build-all.mjs-style orchestration lets one repo serve many vaults if you end up with a series of case studies. On SignalTrace there are five case studies (sa-corruption, alfaromeo, strait-of-hormuz, paypal, attbid) built from one shared Quartz config, 308 markdown files in total.

Autonomous mode

For a brief that’s well-formed, the whole loop runs unattended. Input: a research brief file and an iter_cap. Output: a converged vault, a GRAPH_REPORT.md, and a open-questions/ folder of what remains open.

/vault-builder --auto --goal-file "research-brief.md" --iter-cap 25

The autonomous runs where the brief was too thin have a characteristic failure mode: the early iterations wander, and the convergence check keeps firing “new evidence” on tangentially-related entities. The fix is always to tighten the brief and restart, not to fight the loop.

What this is good for

OSINT / financial research — the SignalTrace shape.
Internal company knowledge — team wikis where the ontology is “people, projects, decisions, incidents”.
Case study sites — multiple vaults from one Quartz config.
Personal knowledge graphs — this CV wiki started as exactly this pattern.

What it isn’t

A general-purpose web scraper (the loop assumes a defined ontology)
An agent that keeps running forever (convergence is enforced)
A substitute for domain expertise (the research brief still has to be good)

DevAi