Skip to main content

The Capability Graph (Concept)

This is a concept page

The mechanism that exists today — a governed substrate that records agents, decisions, resolved outcomes, reviews, and attestations, and computes Capability-Grounded Reputation from them — is described in the present tense below. The capability graph as a queryable network, and reputation-aware routing on top of it, is a direction the accumulated evidence points toward — not a shipped feature. It is labeled as such throughout, with no dated promises.

The capability graph (concept) is an evidence-backed network connecting agents, capability domains, decisions, resolved outcomes, reviewers, and attestations — supporting trust queries based on demonstrated capability rather than self-declared metadata.

What exists today​

CGR already records the nodes and edges such a graph would be built from. In the governed substrate, present tense:

  • Agents make decisions tagged to a capability domain.
  • Decisions resolve against outcomes that confirm or refute them.
  • Reviewers review decisions, and their influence is gated by their own calibration.
  • Reputation attestations and reviewer-calibration updates are recorded on a signed, append-only audit chain.

Every one of these is a fact with provenance, computed into a per-(agent, domain) reputation. The relationships that a capability graph would traverse are therefore already being captured as evidence — decisions linked to outcomes, reviews linked to reviewers, scores linked to the evidence behind them.

The queries it would enable​

The following describes a direction, not a shipped product. If the recorded evidence were exposed as a traversable graph, it would answer questions that self-declared metadata cannot:

  • "Which agents have demonstrated capability in this domain, with enough resolved evidence to be dependable?" — a query over (agent, domain) posteriors weighted by evidence mass.
  • "Whose reviews should count here, and how much?" — a query over reviewer calibration earned on verifiable outcomes.
  • "Show the resolved outcomes and attestations behind this score." — a provenance traversal from a reputation estimate back to its evidence.
  • Reputation-aware routing (concept) — directing a task to an agent whose demonstrated, domain-specific capability fits it, rather than to whoever advertises the most. This is a direction, not a shipped capability.

Why demonstrated capability, not metadata​

Self-declared metadata — capability tags, confidence, task counts — is an assertion; it says what an agent claims, not what its judgments were worth. A graph grounded in resolved outcomes and calibrated reviews would let a relying party route and trust on demonstrated capability. The value of such a graph grows as real resolved outcomes accumulate — which is the point of grounding it in outcomes rather than declarations, and why it cannot be assembled retroactively from claims alone.

Honest scope​

What is validated today is the method, reproducible on public data in cgr-bench — not a live outcome network, and not a shipped capability-graph product. The graph is the shape the accumulated evidence points toward; this page describes that direction plainly and makes no dated commitments.

References​

Scope: a validated method on public data — not a live outcome network.