Verification
Articles about how outcomes are checked and how verdicts update strategy weights.
- How to Build a Declarative Flow
A practical guide to structuring a declarative flow: defining the tools, setting the order, and establishing the verification contracts.
- How an Agent Generates Its Own Tools
A tool is generated as code, signed, verified against an expected outcome, and recorded. A failed tool is not trusted.
- Outcome Verification: Expected vs Actual
Verification compares the actual outcome of a tool against the expected outcome and records the verdict.
- How to Verify Agent Output
Verification: define the expected outcome, execute the tool, compare actual with expected, record the verdict, update the weights.
- Strategy Weights: How Outcomes Change Behavior
Strategy weights are updated after every verified outcome. A strategy that failed is deprioritized; the code does not change.
- Reward Between 0 and 1: Designing the Verdict Scale
A reward is a scalar between 0 and 1 computed from a verified outcome. It updates the strategy weights.
- Undeclared Tool Calls: When a Tool Is Invented
An undeclared tool call is a tool that is invoked without being declared in a flow. The kernel rejects it before execution.
- Unverified Automation: The Silent Failure
Unverified automation is a silent failure: the tool runs, nothing is compared, and the outcome is assumed. The kernel records a verdict.