Sucralosis: Reward Without Realization
When a system chases a stand-in for what it wants, it finds every way to hit the number without the real thing. Three questions that catch it early.
Read the chapter »
Everything here comes from real builds. Build logs with the test counts and the failures left in. Long-form lessons on evals that lied, agents that ignored their rules, and the gates that caught them, written so you can use them on your own work. And a field guide for spotting when a system, human or machine, starts optimizing the wrong thing. New pieces land here first. Who’s writing this

state/repo-state.md, last full run at commit f861f98: 839 TypeScript + 57 Python, 0 failing. Measured. Pulled 2026-09-25.
git log, Sadyr repo: first commit 2026-06-18, live 2026-06-30, 63 commits on 8 working days (claims ledger C-01). Measured.
state/manifest.yaml B-03 / B-08, claims ledger C-08. Measured, 2026-09-08 to 2026-09-09.
Game development-stats report, 2026-09-20 (git + line count). Measured. Written by Claude Code under Todd's direction; the game is not shipped.
state/repo-state.md, filesystem inventory across five capability folders (claims ledger C-11). Measured. Pulled 2026-09-25.
The multi-agent operating system I run Adroit on: agents do real work, inside rules the tooling enforces.
An evidence-first content pipeline: discovery, knowledge base and drafts, with a human approving every change.
A revenue agent that answers only from approved sources and hands a sales rep a briefed lead.
An action RPG in Unreal Engine, built with Claude Code. Playable build under playtest. Not shipped.
A proofread-by-ear desktop app I built because I'm dyslexic. Fully local. Shipped to an audience of one.
An ADHD accessibility system: a local knowledge base structured so an AI agent can navigate it.
Twelve years of marketing and operations work: brand guides, case studies, websites and dashboards.
A field guide to diagnosing systems that drift: what a system really optimizes, and whether it can still be corrected. Software, AI models, companies and the people inside them fail in the same shapes.
When a system chases a stand-in for what it wants, it finds every way to hit the number without the real thing. Three questions that catch it early.
Read the chapter »Some systems can't see their failures. Others turn their own correction machinery against the fix. How to tell bit rot from a rootkit, and what each needs.
Read the chapter »Signs of health are cheap to fake. The markers worth trusting are the ones only real correction machinery can produce, read now, before the outcome.
Read the chapter »Specification gaming, sycophancy and eval-gaming: the proxy failure in machinery with no self. Two senses of corrigibility, and instruments for builders.
Read the chapter »This month I'm finishing Sadyr's meeting-booking flow, rebuilding this site, and queuing 32 LinkedIn posts for October and November. More »
Follow on LinkedIn