Gerra / Field notes
Work in
the open.
Research explainers, buyer guides, technical histories, and methods - tied back to the datasets and the underlying evidence.
How Git History Becomes a Code-Model Evaluation Dataset
A practical method for turning authentic commits, patches, and tests into point-in-time coding-agent evaluations without leaking the answer.
Agents Fail at Office Work Because the Data Never Existed
Models trained on the public web have never seen how organizations run. What consented, cross-tool company data adds to agent training and evals.
What Prediction Markets Know About Your Portfolio
How live event probabilities map to exposed securities with signed direction and sensitivity, and why point-in-time probability history matters.
Repository-Level Code Data for Coding Agents
Why complete repositories, cross-file context, tests, and change history are more useful for coding-agent training than disconnected source files.
Why Identical GPUs Rent at Wildly Different Prices
The same H100 rents for $1.20 to $22.02 per hour. What a 38-provider panel shows about GPU price dispersion, and why the equity signal is a null.
The Sim2Real Gap Costs Robotics Billions Annually
Simulation-to-reality transfer failures create multi-billion dollar losses across the robotics industry, with success rates plummeting from 90% in simulation to under 20% on real hardware, extending development cycles by 6-12 months.
The Global Humanoid Robot Supply Chain: Architecture, Economics, and Strategic Implications
Chinese manufacturing controls 63% of key humanoid robot supply chains while component costs plummet 40% annually. An in-depth analysis of the supply chain dynamics reshaping the $66 billion humanoid robotics market.
The QDD Actuator Renaissance
How a 2016 MIT breakthrough spawned the era of $3k humanoids – and why China now dominates the torque supply chain