CGR FAQ
Common questions about Capability-Grounded Reputation (CGR) — the trust model in which agents earn reputation, per capability domain, from judgments that resolve against real outcomes. For the full treatment, start with What is Capability-Grounded Reputation?
What is Capability-Grounded Reputation?
Capability-Grounded Reputation (CGR) is a trust model for autonomous agents in which reputation is specific to an agent and a capability domain, is updated from judgments that later resolve against observable outcomes, and is always reported together with the amount of evidence behind it. Peer reviews are weighted by the reviewer's own empirically demonstrated calibration rather than counted equally.
How can an AI agent earn reputation from real outcomes?
By making judgment calls that later resolve against observable results. In CGR, a decision only moves an agent's standing when it can be checked against what actually happened — confidence, task volume, and self-description create no capability evidence. The method is independently reproducible: on real credit-default outcomes, CGR flags weak agents before most outcomes resolve (early-warning correlation −0.997 at 25% resolution in cgr-bench).
How do peer reviews of AI agents resist Sybil attacks?
Reviews are weighted by the reviewer's own demonstrated calibration, earned from verifiable outcomes, with a per-source cap. Unknown reviewers fail closed, so fresh identities do not arrive with influence — in cgr-bench, a coordinated review farm shifts the score by zero at every tested threshold, while the same attack moves an ungated average substantially. Manipulation remains a threat-model problem, documented openly.
Why report evidence mass with the score?
Because a 0.90 built on five resolved outcomes and a 0.90 built on four hundred are not the same claim. CGR reports both, and an evidence-gated ceiling keeps thin records from presenting as proven ones until enough outcomes resolve.
Does a high CGR score mean an agent is good at everything?
No — reputation is per capability domain. A strong contract-review agent earns nothing in security auditing. Cross-domain transfer of judgment is empirically weak, so CGR never treats capability as global.
Is CGR reproducible?
Yes — cgr-bench reproduces its properties from source with one command: cold-start and Sybil-resistance behavior asserted in CI, plus validations on real public data (real credit-default outcomes; ~1,900 real human forecasters). Where a reconstruction lands below an originally recorded value, the benchmark reports the reproduced number.
Is this a live reputation network?
Not yet. What exists today is the documented, reproducible method and the governed evidence substrate that captures decisions and outcomes. Reputation becomes more valuable as real resolved outcomes accumulate — that accumulation is the point, and it cannot be faked retroactively by anyone, including us.
References
- What is Capability-Grounded Reputation? — the pillar definition and model.
- CGR overview — the documented scoring model and endpoints.
- cgr-bench — one-command, independently reproducible benchmark of CGR's properties.
- GMP v0.2 specification — the Grafomem Memory Protocol behind the evidence substrate.
Scope: a validated method on public data — not a live outcome network.