GPT-OSS benchmark readiness

gptoss is registered as the text-only mlx-community/gpt-oss-20b-MXFP4-Q8 model served by mlx-lm. The model is downloaded and has completed core, executable, realistic-RAG, and practical full runs. Use gptoss-final for structured requests; it uses the same cached weights with a final-channel Harmony template and explicit JSON-schema injection.

Capability behavior

Workload	Harness behavior
Vision, camera, mixed JSON/text/image	Capability-skipped because `multimodal: false`; not scored as failure
Executable coding and no-change safety	Runs through disposable fixture repositories
Cybersecurity, reasoning, instruction following, RAG, structured output	Runs through the core benchmark suite
Sports, hallucination, REST text workflows, repository reasoning	Runs through the practical full profile
Realistic 2K/8K/16K RAG	Runs through the natural-corpus benchmark
Responses API	Recorded as unavailable (`responses_api: false`), not as model-quality failure

The chat-completions benchmark path is independent of /v1/responses. A missing Responses endpoint affects direct Codex-provider compatibility only; it must not be interpreted as an intelligence result.

Completed validation

On 2026-08-19, gptoss-final completed all 21 core requests with 90.1% mean deterministic quality. The executable-v2 run applied 5/5 decisions, passed 5/5 targeted and full suites, met all five minimality thresholds, produced no regressions, and correctly avoided the adversarial no-change edit. The three supported hallucination-v2 text cases all abstained correctly, producing a 10.0/10 category score with the image case capability-skipped. Realistic RAG scored 100.0% at 2K, 86.5% at 8K, and 95.8% at 16K. The practical full profile scored 7.3/10 overall; its strongest quality categories were agentic (10.0), repository (9.67), and sports (8.8), while hallucination resistance was the primary weakness (2.5).

See docs/results/gptoss-evaluation-2026-08-19.md for the detailed snapshot and result provenance.

Next round

# Run from the local-mlx-lab repository root.
make model MODEL=gptoss-final
make health
# Repeat representative executable and hallucination cases three times.
make stop

The corrected hallucination rerun reported zero failures, zero unsupported claims, and one capability-aware image skip. Do not compare v2 scores directly with unversioned rows without noting the prompt change. Photographic, camera, and vision workloads remain capability-skipped for this text-only model.