The science6 min read
Bias training is debugging the wrong layer
Escalation of commitment is not a character flaw. Two model copies deliberating as equals escalated at 99.2%; alone with no stake, the same model cut its losses.
The blog
Notes on deciding
Research, field notes, and the occasional strong opinion.
What the science actually says about choosing well, why structure beats vibes, and what changes now that the software has opinions too.
Second thoughts7 min read
Loud decisions are rarely the heavy ones
Stakes don't make a decision hard. Irreversibility and silence do. Amazon's one-way door framework names that failure mode, then buries it in a footnote.
The method9 min read
The evidence expires first
OpenAI stopped reporting SWE-bench. One paraphrase reorders model rankings 63% of the time. How to weight decision criteria by how fast each one decays.
Second thoughts7 min read
The weights are decoration
We sell weighted decision matrices. The evidence says the grid does the work and the weighting step adds almost nothing — which changes how you should run one.
Field notes5 min read
The model doesn't get a vote
Recent research says LLMs are superb at widening a decision and reliably sycophantic at judging one. Field notes on keeping the judgment human — with structure.
The method5 min read
The decision record renaissance
AI made text nearly free, and your memory was never trustworthy. Why decision records are the one document worth writing in 2026 — and what goes in one.
The science5 min read
Too many options is a design problem
The jam study got audited, the hungry judges got exonerated, and the fix for infinite options turns out to be deciding what matters before you look.