Home / Patterns / Stepper & folds
Stepper & folds
The learning loop: step rail, evidence folds, model cards and promotion prerequisites.
Step rail + folds
1 · Collect corrections
42 human corrections since the last build — 18 more recommended before promotion.
2 · Train a candidate
Candidate v3 trained; 96% agreement on holdout.
3 · Promote to live routing
Requires the guard-threshold check and your approval.
How agreement is measured
Source / label agreement on a held-out slice, split by category.
What promotion changes
Only routing for categories the model owns; rules still win first.
numbered steps with states (now/ok/todo); folds carry the depth so the page stays scannable
<div class="steps"><div class="step now"><span class="dot acc"></span><div><b>…</b> <div class="sub">…</div></div></div></div> <details class="fold" open><summary>…</summary><div class="fold-inner">…</div></details>
Model cards + inline progress
Category learner
live72% across 14 categories
Agreement
96%
Last build
2d ago
In live routing
5 of 14
Promoter candidate
paused38% of prerequisites met
Needs 18 more corrections before promotion unlocks.
.progress-inline is a light inline progress bar (its 3px radius is nulled by the global radius rule — see AUDIT Q5)
<div class="card-grid">
<div class="card"><div class="card-h"><h3>Model</h3><span class="badge ok">live</span></div>
<span class="progress-inline"><i style="width:72%"></i></span>…</div>
</div>Prerequisites are shown before promotion, not after failure. Source / label agreement, required corrections, and what live routing changes are all stated up front. The learning loop compiles repeated LLM reasoning into small auditable models — the UI has to make that trustable.
Triage Design System v0.13.1-experiment — the central, app-agnostic UI design authority.
Source of truth: /tokens/tokens.json. Showcase chrome only:
site/.