Observations from building the dataset. These are descriptive measurements, not quality judgments; every item links to the raw data that produced it. Single prompt family, small N — read as signals, not conclusions.
For the same model at the same effort, cross-arm similarity between the plain api output and the CLI-harness output sits around 0.60-0.69 — far below the model's own within-arm self-consistency. The harness (system prompt, tool loop, file delivery) is a real variable of the drawing, not a transparent pipe.
Through the Claude Code harness, raising effort from low to max grows the generated card by roughly 14x (fable-5), 12x (opus-4.8) and 9x (sonnet-5) in output volume. The knob visibly changes how much gets drawn.
haiku-4.5 (Claude Code harness) and grok-4.5 (official api) produce statistically indistinguishable output volume across every effort tier we ran. grok-4.3 on the same official api does respond to the knob.
The official /responses endpoint returns 400 for effort=minimal on every gpt model we probed. Where a "minimal" tier appears through proxy or harness channels, it is mapped by the channel, not the model provider.
Effort=max is accepted by gpt-5.6-sol / terra / luna and rejected (400) by every earlier gpt model on the official api.
For the same gpt model and effort, the official api produced 5-8x more card HTML than the reverse-proxy dev channel. Channel identity materially changes what you get — one reason this dataset separates channel grades.
Several previous-generation models that answer fine over the plain api fail outright (server 500) when driven through the opencode agentic harness on the same subscription channel. Tool-loop operation is not implied by api competence.
On the P-q (qualitative) prompt variant, kimi-k3's reasoning volume ran about 10x its P-min behavior — the largest variant sensitivity we observed.