Open models, rebuilt in public.
Every week we take open models and try to bring them to a fully reproducible state — lineage registered on GLaaS, rebuild attempted, receipts published. Judged not by their paperwork, but by whether they actually rebuild.
every rebuild is a clean room: a fresh Claude Code instance, the published materials, and nothing else — all under roar · methodology & rubric ↓
| Model | Rebuild | Blocker | Patch | Discovery | Rebuild cost | Evidence | BOM |
|---|
Grades you can argue with.
rebuild statuses
- ✅ reproduced
- Byte-identical artifact from the published recipe. The strongest claim we make.
- 🟢 recreated
- Trained to comparable metrics; nondeterminism (or unpinned seeds) prevents byte-identity. Still a rebuild.
- 🟡 partial
- Some stages rebuild (e.g., data pipeline), others don't. The notes say which.
- 🔴 blocked
- Could not rebuild within budget. The blocker column says why — it grades the ecosystem gap, not the authors.
- ⏳ in progress
- Currently being attempted.
how the numbers are made
Clean room, every row. Each reproduction is performed by a fresh Claude Code instance — no memory of prior rebuilds — from the model's published materials alone, under the same versioned playbook and budget caps. Reproducibility means a competent stranger can rebuild it; so the operator is always a stranger. What each attempt learns is externalized to the public notes repo, never carried in the operator's head.
Derived, not self-reported. Attempts and burn come from the reproduction effort's own recorded lineage (every failed run is a job in the DAG — the effort has receipts too).
Smoke-test policy. Debugging happens at reduced scale or on CPU; full-scale compute is spent only on final verification. Burn = everything except the final green run.
Budget cap. A model is marked blocked when it exceeds the per-model cap (published per entry) or hits a hard blocker like unavailable data.
patches & forks
When a rebuild needs changes, we fork publicly and report the patch size — "recreated as published" and "recreated after 40 lines of fixes" are different claims. Fixes get offered upstream as PRs.
corrections
Built one of these models? If we got your row wrong — or your latest release fixes the blocker — email rebuilds@treqs.ai and we'll re-run it. Author-verified rows carry a ✓.
Want your model on the list? Same address.
Every row's receipts were captured by roar.
The lineage, the failed attempts, the rebuild command — all recorded automatically at runtime. Your models could have receipts too.