Consistency Verification

Generating one on-style image is achievable. Generating twenty on-style images is where consistency verification earns its keep. The critic agent evaluates every output against the style specification — not against the reference images. The specification is the contract; the references were inputs to the contract.

The consistency score has three components. Dimensional Adherence: for each of the 7 DNA dimensions, does the output fall within the tolerance range specified in the style document? Scored per dimension, minimum 7/10 per dimension to pass. Cross-Output Consistency: compared against the last 5 approved outputs, does this new output look like it belongs to the same family? Measured by visual feature comparison — palette distance, layout similarity, typography matching. Target: consistency score above 0.85 on a 0-1 scale. Intent Fidelity: does the output depict the content described in the prompt? An image that is perfectly on-style but shows the wrong subject is a style success and a content failure.

The three scores are independent. An output can pass dimensional adherence but fail cross-output consistency if it technically meets every spec but "feels" different — which usually means a semi-invariant element was not specified tightly enough.