Articles ​
Each article tells the story of one line of experiments, for a wide technical audience: what was tried, what failed, what held and what it does not prove. The reference stays What we bring: every claim of the project in one table, with its measurement and its evidence, and Results for the numbers. An article repeats neither. Newest first.
- An end-of-turn detector trained on real conversations: it learned them, and did not get better with voice agents, 20 November 2026. The third version of the 632 KB end-of-turn detector: real two-speaker conversations labelled from the other speaker's channel, the labels checked against a second detector before any training, then a result read once: better on held-out human conversations, worse than its predecessor on real speech addressed to an agent, a pattern whose meaning its pre-registration had written down before the run.
- What worked, what failed, on one laptop: the skycmd experiments as of 19 November 2026, 19 November 2026. An overview of every line of experiments: five models that passed a gate written before their test, each with the limit stated with the pass; the failures, published as results; the versions that stopped at their checks before training; the fine-tuning that could not run on a 16 GB laptop; and the method that held it together.
- A 752 KB model that remembers, and still fails its gate: three versions of a fast-weight memory, 17 November 2026. A fast-weight memory with sleep passes for a tiny ternary decision model: a first version whose memory was never used, a second stopped by its own CPU checks before any GPU time, and a third whose memory is used and fades with the delay, while sleep adds nothing measurable and two games lose.
- A 632 KB model that knows when you have finished speaking: one failure, one harness bug, then a pass, 17 November 2026. An end-of-turn detector for voice agents: a first version that failed three of its four gates, a padding bug that had been hiding the reference model's true score, then a second version designed before training around one job, a teacher to distil and a stopwatch-chosen size, which passes, with its weakness on real conversations stated with the pass.
- A 250 KB secrets verifier: rules propose, the model judges, and the gate still says no, 16 November 2026. A tiny model that judges the candidates of a rule-based secret scanner: a first version that passed on its own generated data and failed on real files, then a second, trained and tested on real repositories, that beats every rule scanner tested and still misses its pre-registered bar by 0.027.
- A 3.7 MB tool router built on a MacBook: two failures, then a pass and its limits, 16 November 2026, with a closing section of 19 November 2026. A small model that picks an assistant’s tool: two failed versions, published as failures, a method written down before every test, and a third version that passes, without hiding what that pass does not prove; then four versions that tried to retrain it without a shortcut, the last of which fails on tools of a new family and closes the line.