SourceShift
Engineering notes from the SourceShift team. Post-mortems, LLM gateway scars, and the occasional working theory — drafted from real production fires by the engineers running them. No newsletter, no popups, no tracking.
- 2026jun 23The orchestrator that learns which model to trust
You're running five AI models behind one workflow and overpaying for the wrong one on every job, because nothing measures which is actually best. mini-ork is a CLI orchestrator that makes models compete, grades the results, and re-routes the next job to whoever won — so your routing gets cheaper and better on its own instead of staying frozen at whatever you guessed on day one.
- 2026jun 17Coordinating a fleet of LLM agents on one codebase
Point three coding agents at the same repo and they overwrite each other within minutes. The fix is not to make them talk; it is to give them an external referee. We built one using six ideas that are 30+ years old.
- 2026jun 15You can't align what you haven't measured
A Vapnik decision boundary, an RL exploration result, and a Hubinger sleeper-agent finding all say the same thing: training must include the failure mode. The agentic-era addition: train an agent to elicit it, because the human red-team budget runs out before the action space does.
- 2026jun 11Fable may be smarter. Show me the token bill.
A critique of Anthropic-style intelligence claims: if the gain comes from spending more test-time compute, publish the cost frontier.
- 2026jun 8The chapter that forgot why it existed
When an LLM agent generates text without a world model, it forgets its own goal mid-task. The fix is not more context.
- 2026jun 4Three LLM judges, but really 1.5: why a same-family panel collapses to noise
We needed to settle a disagreement between two LLM reviews of the same design doc. The clean answer was a 3-judge panel. The honest answer is that the panel we built is one rubric-design move away from being a beautifully-instrumented yes-man.
