Statewave Guide
An open-source project for AI code understanding and agent verification. By constructing an application graph and a Feature Scope mechanism, it addresses the verification blind spot in which an AI assertion is technically correct but functionally irrelevant, ensuring that verification is grounded in the correct functional context.
Timeline (last 90 days)
The Statewave Guide team used the anthropic-opus-5 model to run hundreds of tests on its in-house application, and the results showed zero fabricated routes reaching the production model, zero fabricated permissions, and zero unknown topics erroneously introduced.
In validation testing, the model refused all 27 requests for prohibited capabilities and all 23 requests for unsupported operations.
The model has an 89% probability of finding the correct evidence, but only a 20% probability of turning it into a verifiable product assertion; 71% of features ultimately had no verified capability description, and the workflow recovery rate was only 18%.
After introducing Feature Scope, the verified assertion output rate rose from about 20% to 95%, erroneous feature assertions dropped to 0, and model restraint reached 30/30.
Even after the assertion output rate rose to 95%, 45 of the 63 features still ultimately had no verified capability description.
Feature Scope is an intermediate layer introduced by the Statewave Guide team to fix the verification blind spot: the graph first determines the feature currently being analyzed and its corresponding nodes and edges, and then generates candidate assertions only from within that subgraph.
Statewave Guide's design flow was changed from "model → fabricate assertions → verifier" to "graph → deterministic assertion opportunities → model adds meaning → verifier," downgrading the model's role from fact producer to expresser.
Statewave Guide is built on top of Statewave, which is an open-source, self-hosted AI agent memory runtime emphasizing governance and provenance, under the Apache 2.0 license.
The verifier only answers "whether this statement is supported by real evidence," not "whether the evidence belongs to the current feature," which leads to technically correct but completely irrelevant answers.
All Facts (9)
The Statewave Guide team used the anthropic-opus-5 model to run hundreds of tests on its in-house application, and the results showed zero fabricated routes reaching the production model, zero fabricated permissions, and zero unknown topics erroneously introduced.
50%UnverifiedIn validation testing, the model refused all 27 requests for prohibited capabilities and all 23 requests for unsupported operations.
50%UnverifiedThe model has an 89% probability of finding the correct evidence, but only a 20% probability of turning it into a verifiable product assertion; 71% of features ultimately had no verified capability description, and the workflow recovery rate was only 18%.
50%UnverifiedAfter introducing Feature Scope, the verified assertion output rate rose from about 20% to 95%, erroneous feature assertions dropped to 0, and model restraint reached 30/30.
50%UnverifiedEven after the assertion output rate rose to 95%, 45 of the 63 features still ultimately had no verified capability description.
50%UnverifiedFeature Scope is an intermediate layer introduced by the Statewave Guide team to fix the verification blind spot: the graph first determines the feature currently being analyzed and its corresponding nodes and edges, and then generates candidate assertions only from within that subgraph.
50%UnverifiedStatewave Guide's design flow was changed from "model → fabricate assertions → verifier" to "graph → deterministic assertion opportunities → model adds meaning → verifier," downgrading the model's role from fact producer to expresser.
50%UnverifiedStatewave Guide is built on top of Statewave, which is an open-source, self-hosted AI agent memory runtime emphasizing governance and provenance, under the Apache 2.0 license.
50%UnverifiedThe verifier only answers "whether this statement is supported by real evidence," not "whether the evidence belongs to the current feature," which leads to technically correct but completely irrelevant answers.
50%