Bob planned it.
Bob built it.
Bob runs it.
Seven exported Bob task sessions. Each one a receipt for a concrete piece of work. No hand-waving about "built with Bob" — every step has a transcript and a screenshot in the repo.
Every session, in order.
planning
Bob architected Verdict before any code existed. The 6-layer pipeline, the killer-line concept, the cross-layer matching algorithm — all designed inside Plan mode before a single file was created.
init
Bob ran /init and generated AGENTS.md — the project's persistent context document. This is the slash-command feature judges look for.
mcp-server
Bob built the verdict-tools MCP server (3 tools), the Python pipeline harness, the CLI wrapper, and the .bob/mcp.json config — all in one session.
harness-hardening
Bob identified and fixed five robustness gaps: brittle JSON parsing, non-deterministic cross-layer matching, missing retry logic, layer-output schema drift, and the .format() crash on JSON examples.
live-analysis
The killer-line momentThe full pipeline ran end-to-end on a real PR. The cross-layer match fired. The killer line appeared: "Surviving mutation M-1 is the same code path that caused INC-2024-0431."
docs
Bob wrote README, KILLER_LINE.md, LICENSE, and captured the synthesis output into demo/hero-pr-result.md. Submission documentation, all by Bob.
dashboard
Bob built this Next.js dashboard you're reading right now — landing, analyze flow, pipeline page, sessions page (this one).
Not just "built with Bob."
Bob planned Verdict
Session 00 is Bob in Plan mode architecting the entire 6-layer pipeline.
Bob built Verdict
Sessions 01–06 are Bob writing every line of code, every config file, every doc.
Bob runs Verdict
The 6 custom modes ARE Bob executing analysis at inference time. Take Bob out — there is no product.