Status: BINDING (operator policy, recorded 2026-10-14; reaffirms the gate already enforced in TRAINING.md). Supersedes any contradicting statement in planning docs.
| Asset | Disposition | License |
|---|---|---|
| Engine + all code (this repo) | Open, always — it is the product and the adoption surface | Apache-2.0 |
| Model weights (own trained arms) | Open, gate-gated — publishes only after both release-gate legs pass | Apache-2.0 |
| Training corpus (merged-v5 + predecessors, teacher-facts, disputes ledger) | Withheld, private | n/a (not distributed) |
| Extraction/consensus machinery (VIVERE lane, multi-family consensus tooling, per-kind calibration fit, slice-probe/parity discipline) | Process knowledge — described in docs, artifacts withheld | n/a |
No weights publish until suite ≥ 0.68 AND JevBench ≥ 0.6494 — the arm must beat its base and the community checkpoint on both surfaces. First arm through: e1-e (0.8083 / 0.6840). Quantized variants inherit the gate only with their own replayed numbers, like everything else.
- The corpus is the moat, not the weights. Weights alone are inspectable and (with a comparable corpus) replicable; the corpus — 183k typed decision labels with two-family consensus, the per-question disputes ledger, and teacher-facts-v1 — is not cheaply reproducible by a single-model distill. Publishing weights without corpus keeps the training recipe's input private while making the result useful.
- Open weights are the advertisement for the engine. The engine is only demonstrably good when a real model rides it; the corpus keeps the next generation (and any training-service offering) ours.
- Accepted trade: Apache-2.0 weights let anyone build on them, including competitors — including distilling from our released model. That is the price of adoption and consistent with the local-first, no-telemetry, honest-claims brand ("Claims we refuse to make"). A one-teacher distill of our model cannot reconstruct the multi-family consensus corpus.
- Honest disclosure: the published model card states plainly what is withheld (data + machinery) and why. No "fully open" claims.
- Explicit opt-in pull only; the engine never auto-downloads weights at build or run time (PLANNING.md §73). Base binary stays useful at zero ML.
- Model cards carry: base-model lineage, gate numbers per surface (one number per surface, never a universal claim), calibration artifacts shipped (per-kind T), and the license.
Every teacher/judge contributing to a corpus that touches a released arm must be permissively licensed. Current judges: qwen-line teacher (Apache-2.0), strands-decider-E2B (Apache-2.0) — clean. Before adding future judge families (e.g. GLM, Kimi), verify their license permits corpus-derived derivative training; corpus stays private regardless, but keeping the lineage clean preserves the option to release it later.
Experimental tier (operator direction, 2026-10-14). Separately from the corpus of record, probe corpora may be built from ANY teacher family — clearly labeled experimental artifacts, purpose: measure whether teacher family/quality moves the intelligence-transfer curve (RESEARCH.md §16). Rules: probe corpora never feed a gated or released arm, never merge into the corpus of record, and are discard- or-archive-only. The corpus of record stays permissive-clean; the experimental tier is where non-permissive-teacher options get tested before any such decision is reflected here.
- Training/fine-tuning services on the withheld machinery.
- A future corpus release under a negotiated/non-commercial license.
- Next-generation weights (gated on their own legs, not auto-published).
- Enterprise support, certification, and integration work on the Apache-2.0 engine.