Scope: original public experiment. I built the application workflow and interface.
Use it to explore:
- Examine whether a revision responds to a specific criticism.
- Separate tool evidence from a model's own interpretation of its answer.
- Draft with tools
- Critique
- Revise
- Inspect
What happens inside
The draft phase can use Wikipedia search and a bounded arithmetic tool. A critique identifies possible gaps, and a revision responds to it. The stages are streamed so the change in the answer is visible.
Decisions and tradeoffs
Bound tool use
The draft has a finite turn budget. Tools retrieve public snippets or evaluate restricted arithmetic; they do not provide unrestricted code execution.
Expose the revision
A reader can compare draft, critique and final answer instead of seeing only the final wording.
Do not assume self-correction worked
The same model can repeat its errors. The revised result still needs external verification.
Illustrative example
An authored revision example
Draft: ‘A lower training loss proves a model generalizes.’ Critique: ‘That confuses training fit with performance on unseen data.’ Revision: ‘Use held-out evaluation to examine generalization, and account for distribution differences.’ This illustration is written to explain the pattern; it is not model output.
What this does not establish
Critique is not independent evaluation. A more polished response may still be wrong. The experiment does not establish measured quality improvement over a single answer.
Try a live run
Live requests send your prompt to OpenRouter and the selected model provider. Use public, non-sensitive material. Read how demo data is handled.
Checking live availability…
Technical walkthrough reviewed September 2026. Examples do not imply client adoption or production maturity.