Consolidating automated QA into a database
Consolidating QA data into a database makes complexity, structure and code health available to both the human and the agent.
Lynn Cole / 标签
experiments
A thread through the work.
Killer Instinct: the CEO model that surprised me
Training a tiny CEO model on ambition, corporate biographies and legal drama produced something unexpectedly charming.
A coding-agent hardening checklist
A coding-agent hardening checklist: model support, hooks, scheduling, tooling, skills and observability, with the remaining integration work still in flight.
Building the play-training dataset
Building a play-training dataset takes longer than the neat plan suggested. These are the challenge counts, solution trajectories and training signals so far.
When stronger training signals weaken self-assessment
A mentor improved the training signals but weakened Qwen’s self-assessment. The next corpus starts with code reasoning and software estimation.
Product hardening is finding the seams
Product hardening means finding the seams between features, models and workflows, while reducing the task-switching cost for the human.
A training harness with an internal evaluation step
A planned training harness combines coding datasets, recursive refinement, compiler verification, play exercises and persona-based reviews.
Pillowfort breaks under ambition
Pillowfort’s planner behaves differently when asked to build a company from scratch or augment an existing one. Ambition exposes the assumptions in the design.
Why I added an operator-reassurance layer
A long-running agent that works silently can look frozen. Progress narration became a necessary part of the product, even when the underlying work was going well.
Lokshen: ungoogle your noodle
A search engine for indie creators, with anti-SEO editorial rules, a trivia pirate, and a classifier and reviewer that argue over the intake.
After the activation injection: what can the model infer?
An experiment proposal separating activation injection, transcript conditioning, and what a model can infer after the intervention ends.
A composable Brainfuck compiler for LLMs
If people stop reading generated code, what should a language for the machine look like? A deliberately obnoxious compiler proposal.
Music models, control settings, and the Harmonifiddle
Musical failure states, control settings, impossible instruments, and why I started building my own music pipelines.
Qwench completes the Rain Simulator experiment
Five tickets, three worker personas, QA and visual inspection: a Rain Simulator built with Qwen3.5-9B and a 40k context window.