Skip to content

/network/validation

Network Validation

A claim carries its source. The source decides how far it travels; a person decides the rest.

Stages
04
Status
Continuous
01

Agent Proposals

Agents read the graph and can propose a concept. Proposals land pending, never active

02

Source Classification

Every source text is classified, and the classification sets the trust score the promotion gate reads

03

Contested Claims

A claim that collides with another is marked contested rather than silently replaced. Both stay readable

From Source to Canon

  1. 01Ingestion

    Source text enters with its origin recorded on the record itself

  2. 02Classification

    The source is classified, and the classification sets its trust score

  3. 03Promotion Gate

    A draft below the threshold cannot become an active concept without an explicit override

  4. 04Curation

    A person accepts, contests or rejects. Accepted claims carry their evidence

Fixies in the Pipeline

Agents reach the graph through read-only tools and hold conversations with members. One tool writes, and it writes a proposal: a concept marked pending, waiting on a curator. An agent proposes; it does not promote, and it does not write to canon.

Agents run on a Letta server with persistent memory and a fixed tool set.

  • ·Named agents cover market questions, query building and risk reading
  • ·A knowledge submission goes to peer review, never straight to canon

The Promotion Gate

Classification, Not Voting

There is no agent vote. A source text is classified once, and the classification sets the trust score the promotion gate reads.

  • ·Origin recorded on the source text
  • ·Classification sets the trust score
  • ·Colliding claims are marked contested, not overwritten

Human Curation

A person accepts, contests or rejects a pending claim. Curation is the step that makes a claim canon.

  • ·A curator accepts, contests or rejects
  • ·Knowledge submissions go to peer review
  • ·Below-threshold drafts need an explicit override
  • ·Token incentives (pending emission)

Quality Assurance

What Is Watched

Canon claims carry validity dates and a status. A claim can be retired, superseded or marked contested when a later record contradicts it.

  • ·Validity dates on every canon claim
  • ·Known collisions recorded, not resolved silently
  • ·Retired claims stay readable with their reason

Trust Scoring

A source text's classification sets its trust score. The score gates promotion: a draft below the threshold cannot become an active concept without an explicit override.

  • ·Peer-reviewed and published research
  • ·Founder record: voice memos and direct messages
  • ·Member channels
  • ·Open web and public social

Self-Healing Network (Design Goal)

The network is designed to detect and repair inconsistencies over time, improving data integrity as validation matures.

Automatic Correction (Planned)

When conflicts are detected, the system is designed to apply corrections based on consensus and trust scores.

Adaptive Learning (Planned)

Validation algorithms are being designed to improve over time by learning from past validation decisions and outcomes.