README.md

---
title: README
emoji: 🧪
colorFrom: indigo
colorTo: blue
sdk: static
pinned: false
---

<p><strong>slow-stack</strong> 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.</p>

<h3>What lives here</h3>
<ul>
  <li><a href="https://huggingface.co/slow-stack/laya-nli-conflict-v12">laya-nli-conflict-v12</a> — 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.</li>
  <li>Earlier and <strong>failed</strong> rounds (<code>v5 … v12</code>) are kept as a research archive, each card marked "NOT a delivered model".</li>
  <li>Datasets: the <a href="https://huggingface.co/datasets/slow-stack/laya-nli-conflict-eval">frozen 66-case evaluation set</a> (CC0-1.0), the <a href="https://huggingface.co/datasets/slow-stack/nli-conflict-train-lineage">training lineage</a> (row counts + SHA-256 per round), and the current-round mirror.</li>
  <li>Collection: <a href="https://huggingface.co/collections/slow-stack/laya-nli-memory-conflict-head-v4-and-three-run-protocol-6ac698ed9a92a16917804f5b">Laya NLI memory-conflict head — v4 &amp; three-run protocol</a>.</li>
</ul>

<h3>How to read the numbers</h3>
<p>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: <a href="https://github.com/modusensus/laya">github.com/modusensus/laya</a>.</p>

<p>Also here: <a href="https://github.com/modusensus/ai-governance-compare">ai-governance-compare</a> — a claim-level EU↔China AI-regulation comparison with A/B/C evidence grading (conclusions may only rest on grade A).</p>