Said Mohaddes Sadeqi

What I decided

• Write the physics myself. A statevector and density-matrix engine in TypeScript, a parameter-shift VQE, Richardson-extrapolated ZNE — so the browser is the lab bench and nothing is a recording.
• Make honesty mechanical. CI fails if a rendered research claim drifts from the evidence record; every arcade card names what computes it; the pet physically cannot state a number the engine didn't produce.
• Let the research correct itself. I set out to prove a gap, found QUTest and Benchpress, narrowed the claim in public, and then built the one tool the surviving gap actually called for.

How it was built — honestly

I direct AI coding agents the way a lead directs a team: I set the rules (nothing fake, every claim sourced, tests before merge), review every decision, and keep the physics and the research judgment mine. The commit history says Co-Authored-By where an agent typed — and the verification discipline (unit + end-to-end suites, evidence-integrity CI, primary-source-only research) is exactly what makes that collaboration trustworthy. Ask me why the parameter-shift rule works or why the creature uses a semi-implicit Euler integrator; those are my answers, not the agent's.