Articles

Field notes

Technical analysis and methodology from production ML, credit risk, and applied LLM work.

9 long reads, about 120 minutes in all. Newest first.

Latest · October 2026

Your Text to SQL Model Didn't Fail. Your Evaluation Metric Did.

End-to-end accuracy conflates two specification failures with different fixes. Most Text to SQL evaluation harnesses score correct queries as wrong because nobody wrote down what correct means.

Applied ML16 min read

Topic
Oct 2026 SR 26-2 Excludes Generative AI. Your Model Risk Obligations Don't. SR 26-2 excludes generative and agentic AI from its scope while the UK's SS1/23 pulls the same pipeline in. Here is the evidence set that survives both regimes. Model risk12 min read Jul 2026 Self Bootstrap Exfiltration in Open Weights Agents: Absent Without Instruction, Induced by Injection Zero autonomous self bootstrap in 1,332 mundane control trials. Tool response poisoning produces about 36% pooled compliance, while DeepSeek V4 Flash is the only tested model that naturally schemes under goal pressure. Interpretability & safety20 min read Jun 2026 I Built a System to Map My Own Beliefs. AI Products Now Do It by Default. In 2023 I built a system that extracted structured belief maps from conversation. The capability I demonstrated is now deployed at scale, invisibly, inside the products billions of people use. The extraction half of the problem is solved. The influence half is the part nobody can see or measure. Interpretability & safety6 min read Jun 2026 The Refusal Axis Is Layer-Local: Why Single-Vector Steering Hits a Geometric Ceiling The refusal axis rotates ~90° between consecutive transformer blocks. Why single-vector steering hits a geometric ceiling, and what per-layer splines could do about it. Interpretability & safety26 min read Jun 2026 We Cut KYC From 7 Days to 30 Min Using VLMs Using Gemini Flash and Qwen3 8B VL to automate KYC verification at Kuda. Why consistent extraction mattered more than model accuracy, and how stratified evaluation reshaped the pipeline. Applied ML9 min read May 2026 Your LLM Extraction Pipeline Is Out of MRM Scope. Your Credit Model Isn't. When LLM-extracted features feed a credit model, the validation boundary runs through your data pipeline. A practical governance framework for the seam no regulator has defined yet. Model risk12 min read May 2026 We Finetuned Qwen 3 32B to Replace an LLM API How to evaluate a finetuned LLM for structured extraction when standard metrics don't work. Interextraction difference tracking, semantic matching, and the case where the finetuned model was actually more correct than the API it replaced. Applied ML8 min read Apr 2026 Your Best Feature Might Be Your Biggest Liability Pre-deployment stability analysis for credit risk models. Why distributional screening and model-decision screening give you opposite answers, and what to do about it. Model risk11 min read

If you’re evaluating this tradeoff for your own pipeline, I’m happy to talk through it.

Book a call Email

Work with me

Available for LLM, credit risk, fraud, and KYC engagements. Book a call to discuss your project or send an email if you prefer async.

Book a call Email instead

  • Fully remote
  • Short, medium and long-term contracts
  • UK, US and international clients
  • Async first, working with teams from US West Coast to Singapore time

jordan@jordanbrowne-moore.com