What builders changed with Dorsal

1. A misleading AI metric corrected

   A builder used Dorsal to uncover an evaluator that blamed the model for text it hadn’t written. Correcting the measurement changed the reported error rate from 2.1% to 0.8%.

   Builder’s field report · September 2026

2. A research approach improved

   Dorsal surfaced alternative architectures for separating characters from noisy backgrounds. The researcher reported the share of outputs without spurious characters rising from roughly 80–82% to over 90%.

   Researcher’s feedback · August 2026

3. 30 decisions or briefs changed

   A paying founder’s agent recorded 30 changes over 27 days. Changes still in use included sealed test cases, focused reviewers and recording production AI calls for controlled replay.

   Customer-agent report · October 10–November 5, 2026

4. A proposed rewrite rejected

   A proposed Excel-to-PDF path turned one workbook into 1,591 pages, including 117 blank pages. Dorsal’s evidence and the agent’s local tests helped keep Excel as the primary path.

   Builder’s decision test · August 2026 · See the experiment

5. Guardrails and tests strengthened

   A solo builder’s report traced Dorsal’s contributions across seven areas, including process signaling and authorization races. Outside examples informed guardrails and tests; local tests established behavior.

   Builder’s repository report · November 2026

6. An independent production review

   An experienced engineer installed Dorsal without help. His report describes how counterevidence validated locally found issues and made the proposed fixes specific.

   Engineer’s first-use report · August 2026

Based on user feedback and agent-generated reports, not a controlled benchmark. Some advice was rejected or reverted. Updated November 9, 2026.