EvoForge

EvoForge is a coding workspace that turns AI-written changes into results developers can inspect and verify. It addresses a practical problem: plausible code and a confident agent conversation are not evidence that a change works. Muse Spark 1.3 is the principal Builder. GPT-6.1 Sol and Gemini 4 Argon provide independent implementation plans; Claude Opus 5.5 is the specification-first Critic. A separate Runner executes tests, PostgreSQL preserves scoped state and cost reservations, and Ed25519 certificates make recorded result bytes checkable. The workspace shows EvoForge and a single Spark call side by side, with model costs and disclosed allowances. Judges can try the server-funded defaults without entering API keys. Live product: https://evoforge-web-230852112787.asia-northeast3.run.app/en Recorded four-model run: https://evoforge-web-230852112787.asia-northeast3.run.app/en/share/job_306819b0b688 Verify: https://evoforge-web-230852112787.asia-northeast3.run.app/en/verify/job_306819b0b688 The recorded run passed 13/13 hidden checks in both lanes: a tie, not a superiority claim. Model API usage was 10 cents for EvoForge, 3 cents for the single call and 13 cents for the shared Judge; Runner compute is excluded. Separately, an earlier same-model, single-attempt experiment improved Spark from 22/36 to 35/36 tasks by changing only the patch harness, with non-overlapping 95% Wilson intervals. Historical measurements are not presented as results of the new four-model configuration. Hackathon contributions include the deployed workspace, model orchestration, execution and evaluation paths, durable evidence and signed-proof workflow. Earlier design documents are background, not claimed as newly written implementation. The video is an edited, narrated walkthrough of recorded production evidence, not a live screen recording. Current limits: one shared project and a 20-run site-wide daily quota. Signatures authenticate certificate bytes; adversarial verifier hardening remains unfinished.