Can You Trust the Probability? Testing a Decision Model on 77 PubMed Papers
A decision model's probability is not how sure it is. On 77 PubMed papers, see how to check calibration and set a threshold from coverage and accuracy.
All articles
One Forward Pass, Then Probabilities: How Decision Models Work and Are Trained
A decision model reads the logits of listed options after one forward pass and turns them into probabilities. Jev discloses RLCD but not its loss; open replicas publish their heads.
Our Own AI Memory System: Ghost In Shell 5.2 Illustrated
This is an AI memory system we designed ourselves from a year of actual use. It does not use a vector database. It uses plain-text files plus a trigger-phrase index. Six figures show how it works: three roles, trigger phrases written as symptoms, the startup index character budget, the propose → judge → apply flow for merging memory, a hit-rate measurement over 301 sessions, and the trade-offs of this approach.
AI Remembers but Fails to Recall: A Year of Lessons from Our Own AI Memory System
Several AI CLIs shared a memory system we designed ourselves, in one workspace, for a year. It accumulated about a thousand episodic memories and more than five hundred knowledge notes. What caused the loss was not that the AI forgot. It was that the AI trusted something it had not checked. This piece covers the three fixes in Ghost In Shell 5.2, and six pitfalls we hit ourselves, each with the check we use now.
On one 16 GB card, the accuracy winner and the speed winner are not the same deployment
On one RTX 4060 Ti 16GB and one frozen set of 700 items, Ridge, OrcaSAQ-2, and Ternary Bonsai 2 scored 78.4%, 77.6%, and 74.6%. Wall clocks were 47.55, 62.59, and 34.01 minutes. This compares three deployments, not one set of weights.
What Is a Decision Model? Jev, Clef and the AI That Only Chooses
A decision model returns a probability for every pre-listed option in one forward pass and writes no text. Here is how it differs from LLM JSON output, where Jev and Clef stand, and when not to use one.
Ternary Bonsai 2 27B on a 16GB GPU: why multi-layer bakeoffs promoted it to daily default
On an RTX 4060 Ti 16GB seat, Prism Ternary Bonsai 2 27B (PQ2_0) was scored against a same-box Qwen3.8 dense daily model across short functional gates, long-context needles, and tool loops before a stable default alias pointed at the ternary engine.
Nine billion DNA letter swaps, pre-drawn as a map: what AlphaGenome Atlas ranks
DeepMind’s AlphaGenome Atlas precomputes molecular effects for about nine billion possible single-letter DNA changes and ranks them with AVI scores. A research traffic map, not a clinical diagnosis.
Letting Agents Assign Tasks to Each Other: A Practical Build from tmux to Handoff
How to turn manual copy-pasting across CLI agent terminals into auditable mutual task assignment using dynamic discovery, capacity allocation, and file-first handoffs.
One Month of Local Model Benchmarking: I Overturned My Own Rankings
From acceleration knobs and extreme quantization to a nine-model comparison, a large-sample retest at month's end reshuffled the entire leaderboard. A record of a broken yardstick, small-sample error bars, and the current division of labor between pi and local models.