Writing/Fig. 06 — ink, finding its shape
What worked, what failed, and what the evidence changed.
117 investigations of injected failures, four AI models, three days — the results, the wrong answers, the bugs that were ours, and the limits of what it shows.
The first version, part by part: what it does today, what it deliberately doesn't do, and what comes next.
A realistic forecasting estate, broken fifteen ways, and six investigators — from a bare language model to the full Lumis loop — asked to find each cause.
On why production incidents need the creative jump language models struggle with — and why the answer is a model of the system, evidence for every claim, and people in control.
Every ingredient for a self-healing data and AI pipeline already exists. What's missing is a trustworthy way to put them together.
Benchmarking XGBoost, LSTM and Transformer architectures against XAU/USD session structure — what held up, and what was just a flattering backtest.
Write-ups by MinoHealth AI Labs about systems I worked on. Not my byline — included because they document the work.