Technical note
Execution and verification are different states.
A technical note on the bounded arithmetic workflow implemented in this website preview.
The question
How should an interface represent an action that completed but did not produce the requested result? This note examines one deterministic arithmetic example, not an AI-model evaluation or a general reliability benchmark.
The fixture
The preview bundles three invented daily order counts: Monday 18, Tuesday 22, and Wednesday 15. The requested result is their sum. The fixture is intentionally small enough for a person to inspect without trusting the interface.
Separate the state transitions
The sequence is Problem, Evidence, Proposal, Authority, Execution, Verification, and Proof. Advancement stops at Authority until the visitor approves this local calculation. Approval grants no account, file, provider, or connected-service permission.
What execution records
The execution step records the proposed result. With the mismatch option disabled, it records 55. With that option enabled, it records 54. Both branches reach the verification check; neither receives proof merely because a number was recorded.
What verification checks
The check recomputes 18 + 22 + 15 and compares the expected sum with the recorded result. Equality reaches Proof. A mismatch enters Could Not Verify and stops the workflow without a verification seal.
Interruption and reset
Pause prevents advancement while preserving the current local state. Resume returns to the same stage and permission boundary. Reset removes the sample approval and recorded result and returns to Problem. State is kept in browser memory rather than persisted to storage.
A bounded conclusion
This demonstrates how one checkable result can remain separate from its execution state. It does not establish factual accuracy, successful external action, dependable citations, model performance, security certification, or production Merism availability.
Questions for larger workflows
A broader workflow needs a check aligned with its real goal, authorized evidence access, independent enough observations, explicit failure handling, and a policy for checks that are unavailable. Showing a receipt is useful only when its contents support the conclusion.