slow-stack is a small lab for work that has to survive its own audit: preregistered experiments, frozen evaluation sets, and SHA-256 lineage from corpus to weights.

What lives here

• laya-nli-conflict-v12 — a 0.3B memory-conflict detection head (does a new statement conflict with, or stay compatible with, what was already stored?). The delivered head of a 12-round preregistered protocol; 14/14 acceptance axes passed.
• Earlier and failed rounds (v5 … v12) are kept as a research archive, each card marked "NOT a delivered model".
• Datasets: the frozen 66-case evaluation set (CC0-1.0), the training lineage (row counts + SHA-256 per round), and the current-round mirror.
• Collection: Laya NLI memory-conflict head — v4 & three-run protocol.

How to read the numbers

Thresholds and the aggregation rule were written down before the runs; three runs of the same configuration are reported by median, never picked. All data is synthetic — no real user data — and the evaluation set is asserted disjoint from the training corpus (leak audit 0/51). Code and claims pack: github.com/modusensus/laya.

Also here: ai-governance-compare — a claim-level EU↔China AI-regulation comparison with A/B/C evidence grading (conclusions may only rest on grade A).