0099 — Heretic v2 of GPT-OSS 20B is the sovereign default model

Status: proposed · Date: 2026-11-04 · Cites: sub-doctrines 8.a, 8.d, 8.j, 10.f, 12.c · Related: ADR-0027, ADR-0038, ADR-0064, ADR-0093

Owes: the advertised ADR count in CLAUDE.md, AGENTS.md, GEMINI.md, README.md and docs/index.md (tests/meta/test_adr_counts.py), a nav entry in properdocs.yml, and a fragment in changelog.d/. All are included in the change that carries this record.

Context

ADR-0064 established gptossloop on GPT-OSS 20B (gpt-oss:20b) as the sovereign default engine, on by default and preferred first for BUILD (ADR-0038), and as the default model behind the sovereign DESIGN and DECOMPOSE providers (ADR-0027).

In an autonomous coding loop, a model must inspect arbitrary source trees, execute shell commands, analyse security properties, and repair failing checks. Stock gpt-oss:20b refuses 98 of 100 prompts on the Heretic refusal benchmark. When a coding prompt or repository text triggers a refusal reflex inside an unattended DESIGN, DECOMPOSE or BUILD turn, the loop stalls for a reason that has nothing to do with the operator's authority or the safety of the work. In vibey, safety is enforced by deterministic bounds — human gates, per-cycle spending and turn caps, worktree and container isolation, environment scrubbing, and an append-only PostgreSQL ledger — rather than by a model's alignment reflexes.

The Heretic project publishes two abliterated builds of GPT-OSS 20B:

  • v1 (p-e-w/gpt-oss-20b-heretic): 58/100 test prompts refused, KL divergence 0.96 from the original model.
  • v2 (p-e-w/gpt-oss-20b-heretic-v2, built with Heretic v1.1.0): 49/100 test prompts refused (against 98/100 for stock), KL divergence 0.1248 from the original model, under the Apache-2.0 license.

Ollama's Hugging Face GGUF integration serves p-e-w/gpt-oss-20b-heretic-v2 directly via hf.co/mradermacher/gpt-oss-20b-heretic-v2-GGUF:MXFP4_MOE, which uses the same native MXFP4 mixture-of-experts quantization and 20B architecture as stock gpt-oss:20b.

Decision

  1. A two-variant sovereign model catalogue (src/vibey/domain/sovereign_model.py). SovereignModels (declared by SovereignModelsInterface) defines the two sovereign builds of SOVEREIGN_FAMILY = "gpt-oss:20b":
  2. heretic (default): hf.co/mradermacher/gpt-oss-20b-heretic-v2-GGUF:MXFP4_MOE, derived from p-e-w/gpt-oss-20b-heretic-v2 (Apache-2.0 weights; ~12.1 GB download).

  3. stock: gpt-oss:20b from the Ollama library (Apache-2.0 weights; ~14 GB download).

  4. Why v2 over v1. Heretic v2 reduces refusals further (49/100 vs 58/100) while keeping KL divergence at 0.1248 (nearly an order of magnitude tighter than v1's 0.96), preserving the non-refusal behaviour and coding capability of the base model.

  5. License boundary. The Heretic tool itself is licensed under AGPL-3.0 and is not a dependency or runtime component of vibey. vibey consumes only the published Apache-2.0 model weights through Ollama.

  6. One resolution precedence everywhere (12.c). SovereignModels.from_environment() resolves the active model tag in order:

  7. Explicit --ollama-model CLI flag;
  8. VIBEY_OLLAMA_MODEL environment variable;
  9. VIBEY_SOVEREIGN_MODEL environment variable (heretic or stock, case-insensitive; an unknown value raises ValueError naming the valid keys);
  10. DEFAULT_SOVEREIGN_TAG (heretic).

Both --ollama-model and VIBEY_OLLAMA_MODEL also accept the variant keys heretic and stock and resolve them to their full tags, while passing any other Ollama tag through unchanged. Empty or whitespace-only values count as unset.

  1. Single source across conductor and runner. DEFAULT_SOVEREIGN_TAG is wired into OllamaChatClient, OllamaResidency, domain/local_stack.py, vibey doctor --sovereign-fit, and qwenloop's DEFAULT_ENDPOINT_MODEL for gptossloop. Because the qwenloop workspace tenant does not import vibey, tests/meta/test_sovereign_model_agreement.py asserts that the tenant's constant equals DEFAULT_SOVEREIGN_TAG.

Evidence and verification status (sub-doctrine 10.f)

  • Verified:
  • The model card metrics for p-e-w/gpt-oss-20b-heretic-v2 (49/100 refusals vs 98/100 original, KL divergence 0.1248, Apache-2.0 license, Heretic v1.1.0).
  • The Hugging Face Ollama manifest at https://huggingface.co/v2/mradermacher/gpt-oss-20b-heretic-v2-GGUF/manifests/MXFP4_MOE resolves cleanly to a ~12.1 GB MXFP4_MOE GGUF model layer plus a template layer (Q4_K_M is ~15.8 GB).
  • Unit, CLI golden, and cross-tenant agreement tests pass at 100% branch coverage.
  • Not yet verified:
  • hf.co/mradermacher/gpt-oss-20b-heretic-v2-GGUF:MXFP4_MOE has not yet been pulled with ollama pull or run through vibey doctor --sovereign-fit or an end-to-end gptossloop tool-calling benchmark on the reference host. Existing measured latency, VRAM, and throughput tables in docs/reference/system-requirements.md and docs/guides/local-models-ollama.md remain explicitly attributed to stock gpt-oss:20b until re-measured against the Heretic v2 tag.

Consequences

  • Default local installs and vibey loops --json now report hf.co/mradermacher/gpt-oss-20b-heretic-v2-GGUF:MXFP4_MOE as gptossloop's model and 12.1 GB as its download size.
  • Operators who want the unmodified OpenAI release can select it with VIBEY_SOVEREIGN_MODEL=stock, VIBEY_OLLAMA_MODEL=stock, or --ollama-model stock.
  • Rule status. Choosing which build of the era's default model sits at the top of the catalogue is a mechanism and catalogue update under sub-doctrine 8.d, not a new rule of conduct (ADR-0020).