ADR-0015: Qwen-teacher verified distillation into Ouro — proposer, not imitation; execution is the teacher of record

Status

Proposed — awaiting approval from Alex Place. This ADR does not start a training run. It records how a Qwen→Ouro teacher/student path is allowed to exist at all under Σ₀, and authorizes only the offline, no-GPU proposer + verify front half (scripts/qwen_teacher_crystallize.py scaffold). The train/eval/promote back half stays gated on this ADR's approval and on ADR-0010.

Context

The owner's directive: use Qwen as the teacher and Ouro ("oura") as the student for the Σ₀ council — figure out how best to use Qwen to crystallize a model for Ouro, and make the unisona model update. Taken literally, that is knowledge distillation: copy a strong model's outputs into a weaker one.

Two repo facts constrain the literal reading:

1. Qwen was retired from the local coder lane (2026-06-28,

docs/SIGMA0-MODEL-ADAPTER.md; the registry test asserts !chain.includes("qwen2.5-coder")). It is no longer registry-managed but still pulls via Ollama (Qwen2.5-Coder-7B, ~4.7 GB @ Q4_K_M — fits theGB 3070). So Qwen is available as a tool, just not as a lead.

1. ADR-0010 forbids plain distillation.

The only sanctioned training signal is externally-verified-correct experience (code that passed tests; claims that are grounded). A soft-label teacher→student copy has no verification step, so it fails the External Reality Rule ("nothing accepted without evidence"). The existing flywheel (scripts/continual_ouro_pipeline.py) is built exactly this way: harvest → execution-verify → train → eval → promote, and its #1198 distillation branch only ingests cloud-teacher solutions that already passed the repo tests at capture time.

The tension resolves cleanly if we change what Qwen contributes. Qwen does not teach Ouro by imitation. **Qwen teaches by proposing; a compile+exec+assert gate is the teacher of record. Qwen becomes a high-throughput candidate generator; only its solutions that actually run green become Ouro training rows. This keeps the owner's teacher/student framing and** obeys Σ₀ — verification, not the teacher's authority, is what promotes a row.

Loop stages touched: Reason (a stronger proposer widens the candidate set the student learns from) and Converge (adapter-only learning from verified experience, already sanctioned by ADR-0010). Feature-gate check: this extends the existing continual-training flywheel and the one CSF/memory substrate with a new proposer front-end — it is not a new dream engine, a second memory system, or a parallel agent ecosystem. One Convergence Core; one flywheel; one more interchangeable proposer that happens to be Qwen.

Decision

We will allow a Qwen-teacher → Ouro-student path only in its verification-gated form, and implement it as an extension of the existing flywheel — not as a new subsystem.

1. Qwen is a proposer, never an oracle. Qwen2.5-Coder generates candidate solutions +

reasoning traces for decontaminated coding prompts. A Qwen output that is not execution-verified is discarded, not trained on. There is no soft-label / logit / KL distillation from Qwen — only its verified behavior survives.

1. Execution verification is the teacher of record. The same compile+exec+assert

green-subprocess gate that already governs continual_ouro_pipeline.py decides which Qwen-proposed rows exist. Provenance is stamped on every row (meta.proposer = "qwen2.5-coder", meta.verification = "green-subprocess"), so the corpus is auditable and a bad teacher is traceable and revocable.

1. The student is Ouro-1.4B (runnable now). The crystallized corpus trains an **Ouro-1.4B

QLoRA adapter** (~45 s/step on theGB 3070) — not the 7.6 B unisona.ai-Σ₀ PLT, which cannot train underGB and waits on the ≥24 GB box ([#1829]). "Crystallize a model for oura" targets Ouro deliberately.

1. Train/eval/promote is unchanged and eval-gated. Training reuses

scripts/train-qlora-ouro.py; promotion reuses the pure decide_promotion gate — a candidate promotes iff it beats the incumbent's held-out pass@1 by a margin, logged as a Convergence Record to data/eval/ouro-promotion-log.jsonl. The unisona model update is that promotion: the adapter swap + the local-model-registry.js entry flipping verified:false → true only because an on-box eval log backs it — never a vendor number.

1. Council is the trigger + follow-up authority, not a training input. The Σ₀ council frames

one Convergence Record (hypothesis: "verified Qwen-proposed traces lift Ouro held-out pass@1 above the incumbent by ≥ margin"; evidence: the baseline logs; follow-up: the Stage-D eval gate). Qwen and Ouro are the two seats — Qwen proposes, Ouro is the incumbent being challenged. The council decides whether the hypothesis converged; it does not hand-feed weights.

1. Decontamination + secret-scrub are mandatory front-gates. Prompts and rows pass the

existing 13-gram decontamination vs HumanEval + MBPP (scripts/decontaminate_training.py) and the secret-scrub regexes from pr_crystallize.py, before CSF packing. Public benchmarks stay read-only targets (ADR-0010 rule).

1. Offline, operator-gated, reversible. Every guardrail from ADR-0010 applies unchanged:

base frozen, adapter-only, collapse tripwire on promotion, content-addressed + CSF-archived, revertible, operator-initiated. The live request path never retrains.

Scope authorized by this ADR on approval: the proposer + verify + CSF-pack front half (no GPU). The train + eval + promote back half remains additionally gated on the operator running it on the GPU box under ADR-0010.

Consequences

• Positive:
  • Reconciles the owner's teacher/student directive with the External Reality Rule without

weakening either — a stronger proposer, a stricter gate.

• Reuses the entire verified flywheel (verify gate, QLoRA trainer, decide_promotion,

promotion log); the only new code is the Qwen proposer front-end. Extension, not addition.

• Runnable on the existingGB box today (Qwen inference + Ouro-1.4B QLoRA are both in-budget),

so it is not blocked on the ≥24 GB handoff.

• A stronger, licence-clean coder (Qwen2.5-Coder, Apache-2.0) widens Ouro's verified-corpus

coverage beyond what merged PRs alone supply.

• Negative / trade-offs:
  • Teacher bias leaks through the verifier's blind spots. Execution-verify proves a solution

runs and passes its asserts, not that it is idiomatic, safe, or non-degenerate. Mitigation: keep the asserts strong, cap per-prompt yield, and monitor the collapse tripwire on the trained adapter.

• Adds a model-serving dependency (a local Qwen endpoint) to the offline pipeline; mitigated by

keeping it opt-in and pointing at any OpenAI/Ollama-compatible endpoint via env.

• Distillation — even verified — is still the ADR-0010 last resort. This ADR does not claim

frozen-base levers are exhausted; it only makes the shape legitimate for when they are.

• Follow-ups (each gated by on-box evidence — none auto-promotes a model):
  • Stage A — Council scoping. One Convergence Record framing the hypothesis + baseline

evidence + follow-up gate.

• Stage B — Proposer + verify (this ADR's authorized front half).

scripts/qwen_teacher_crystallize.py → decontaminated prompts → Qwen candidates → green-subprocess verify → secret-scrub → 13-gram decontaminate → CSF pack. No GPU. Scaffold landed with this ADR; wiring the live Qwen call + a sampled row review is the first PR after approval.

• Stage C — Train. train-qlora-ouro.py on the verified corpus (8 GB box).
• Stage D — Eval-gated promote = the unisona update. eval_sigma0_adapter.py /

eval_humaneval_ouro.py; decide_promotion; registry verified flip + adapter swap; log the Convergence Record.

Alternatives considered

• Literal soft-label / KL distillation from Qwen into Ouro. Rejected — no verification step,

fails the External Reality Rule and ADR-0010's source gate. It would train Ouro to imitate Qwen's mistakes as readily as its correct answers.

• Re-instate Qwen as the local coder lead instead of distilling it. Rejected as a different

decision — it re-opens the 2026-06-28 retirement and ADR-0011's sole-coder stance, and does nothing for owning/adapting the Ouro substrate. Out of scope here.

• Keep only PR-crystallized data (pr_crystallize.py), no Qwen at all. Legitimate and is the

safe default if this ADR is rejected — merged PRs are already verified. Rejected as sole source because PR volume is finite and long-diff-skewed (only 530/1,178 unisona rows fit seq=1536, [[crystallization-grounding-corpus]]); a verified proposer cheaply widens coverage.

• Distil into the 7.6 B unisona.ai-Σ₀ PLT instead of Ouro-1.4B. Deferred — the 7.6 B student

cannot train underGB (#1829). The corpus this pipeline produces is model-agnostic and can feed the PLT student later on the ≥24 GB box.

Evidence