Self-Discovery & Publishing Loop

Self-Discovery & Publishing Loop

A living intelligence loop for domains, projects and corpora.

EigenVertex continuously improves the knowledge system around a domain. It detects gaps, watches for change, discovers candidate sources, queues evidence for review, refreshes memory and publishes grounded outputs.

Most systems wait for a question. EigenVertex watches for what changed. A corpus should not become stale the day after ingestion — so EigenVertex pairs every corpus with a loop that finds what is missing, what is new, and what is worth publishing.

Discovery finds what is missing. Publishing turns validated knowledge into usable outputs.

Canonical Loop

The operating model

This is the operating model of EigenVertex as an active domain intelligence system.

01 Corpus, projects & sessions The living material EigenVertex already knows and works from.
02 Detect gaps and changes Find what is missing, weak, contradicted or out of date.
03 Discover candidate sources Search for sources, entities and claims worth adding.
04 Score, deduplicate and queue Every candidate explained, not just ranked.
05 Human or agent decision Accept, reject or watch — under explicit policy.
06 Ingest selected evidence Accepted sources enter the corpus through the same connector.
07 Refresh vectors, graph & temporal memory Only the impacted layers are updated.
08 Publish grounded outputs Briefs, digests, decks and pages — then measure feedback and repeat.

Passive RAG

  • Answers questions
  • Depends on existing sources
  • Becomes stale
  • Rarely explains what is missing

EigenVertex Discovery Loop

  • Detects missing coverage
  • Discovers candidate sources
  • Tracks changes over time
  • Refreshes memory selectively
  • Publishes grounded outputs
  • Keeps humans in control

Discovery

Discovery — find what is missing and what has changed

EigenVertex separates three responsibilities: a connector knows where to search, an agent knows what to search, and a skill knows how to reason in a domain. Ten agent families cover source discovery, gap detection, contradiction tracking and validation — only the families needed for the mission at hand are active.

Scout Tracks recent signals and newly available sources.
Gap Hunter Detects missing, weak or contradictory coverage inside a corpus.
Alternative Hunter Surfaces competing approaches and missing diversity.
Entity Expander Explores an entity’s neighbourhood, aliases and categories.
Source Prospector Finds new sources worth reusing across missions.
Contradiction Hunter Finds claims that conflict with existing knowledge.
Change Detector Compares two temporal states and proves what differs.
Bridge Builder Finds links between domains, projects or workspaces.
Validator Verifies a candidate before it is allowed into the corpus.
Triage Editor Deduplicates, scores and prioritises the candidate queue.

Candidate queue & scoring

Scores are explainable. A priority number is not enough — EigenVertex keeps dimensions, reason codes, provenance and decision history for every candidate.

CandidateSourceReason codeRelevanceNoveltyCoverage gainCredibilityDecision
missing_coveragerecent_updatecontradiction_signalnew_sourcevalidation_neededurgent_announcementhigh_authority_sourcedigest_candidatealert_candidate

Refresh memory selectively, not blindly

After evidence is accepted, EigenVertex updates only the impacted layers — then produces a before/after change report.

Compiled corpus memoryQdrant vectors and citationsFalkorDB GraphRAGGraphiti temporal memoryProject, session and briefing projections

Human-in-the-loop

Automation with control

By default, ingestion is human-reviewed. Automation is possible only under explicit policy, trusted sources, thresholds, budgets and kill switches.

Human review Every ingestion requires a decision.
Semi-automatic High-confidence items can be accepted; ambiguous cases go to review.
Controlled automatic Only for trusted sources, explicit policy, budget limits and an audit trail.

No silent ingestion. No invisible automation. Every decision remains traceable.

Publishing

Publishing — turn validated knowledge into grounded outputs

Validated knowledge becomes documents, presentations, digests and pages — every one keeping provenance, citations, version and distribution trace.

Document outputs Briefs, reviews, reports and digests on a common, sourced contract.
Presentation outputs Real, editable .pptx decks — not flattened slide images.
Recurrent publishing Hourly, daily or weekly updates, only on significant change.
Distribution App, file export, email, private web, RSS and webhook.

PowerPoint outputs are real editable .pptx files: templates, masters, layouts, text, tables, charts and notes stay native objects — never a flattened slide image.

Publish only when something meaningful changed

EigenVertex can produce hourly, daily or weekly updates, but only when a significant delta is detected — with schedule, budget, suspension and a full replay trace of every attempt.

  • Schedule and watch policy
  • Delta detection before publishing
  • Budget, suspension and replay
  • A trace of every attempt, successful or not
Daily market intelligence digestWeekly project risk updateCustomer service issue digestRegulatory watchCompetitor updateScientific literature watch

Watch Policies

Watch policies — controlled domain monitoring

A WatchPolicy defines what EigenVertex watches, where it watches, how often, with which budget and under which intake rules.

Web & websitesRSS / AtomSitemapSearch engineEmail newslettersX / public signalsYouTubeAcademic sources (arXiv)Private drive sources

Architecture

Architecture built for traceability

A connector knows where to search. An agent knows what to search. A skill knows how to reason in a domain. These three responsibilities are never fused.

Postgres Missions, candidates, decisions, publishing runs and operating state.
Qdrant Vector retrieval, chunks and citations.
FalkorDB GraphRAG Documentary relations across entities and decisions.
Graphiti + FalkorDB Temporal memory for projects, sessions and briefings.
LangGraph Multi-step orchestration for discovery, refresh and publication.
LLM matrix Provider selection per phase, Cerebras-first where speed and quality allow.

A connector knows where to search. An agent knows what to search. A skill knows how to reason in a domain.

APIs

Build discovery and publishing into your own agents

Every step of the loop is exposed through the API — from profile and missions to candidates, queue decisions and published artifacts.

Discovery
  • Discovery profile
  • Discovery missions
  • Discovery candidates
  • Discovery queue
  • Accept / reject / watch
Live Discovery
  • Live discovery source profiles
  • Live discovery items and queue
Publishing
  • Publishing request
  • Publishing run
  • Published artifact
  • Distribution event

Backlog

An executable product program

Discovery, Publishing and Corpora Constellation are tracked as a clear program of epics — not a loose feature backlog.

DISC Discovery Detect gaps, discover sources, queue candidates, refresh memory.
PUB Publishing Generate grounded documents, decks, digests and recurring publications.
CONST Corpora Constellation Connect corpora, projects, sessions, sources, entities, news and outputs.
APP Beta v1 App Make RAG, Projects, Sessions, Discovery, Publishing and Observability operable.
This is the loop that turns corpora into living intelligence systems.

Human control is the default. Automation is explicit, bounded, observable and reversible.