Diffusal v1.0.1 — License clarification

This release supersedes v1.0.0 for licensing clarity.

• Diffusal-authored source code is licensed under the MIT License.
• The paper, whitepaper, figures, and other research prose are licensed under CC BY 4.0.
• third_party/ retains its upstream licenses.

No research results or experimental artifacts changed.

Diffusal v1.0.0 — Initial public research release

Initial public research release

This release archives the Diffusal research program, including pre-registered experiment configurations, evaluation scripts, raw result summaries, and the current paper draft.

Current evidence

• INT4 PTQ: matched MDLM diffusion language models show lower relative degradation than their autoregressive controls in the reported likelihood and generative-anchor evaluations.
• Native ternary QAT: the three-seed 7M-scale replication supports the pre-registered no-extra-dLLM-tax criterion; it does not establish superior diffusion robustness.

Scope and limitations

Results are bounded by the evaluated models, hardware, and recipes. They are not a claim about 7B+ models, a 25B ternary dLLM, or calibrated PTQ. See ARXIV-PAPER.md and the experiment result files for methods, raw values, and caveats.

Citation and reuse

Citation metadata is available in CITATION.cff. Repository materials are available under CC BY 4.0.

---

License clarification: v1.0.0 was superseded by v1.0.1. Diffusal-authored code is MIT-licensed; paper and research prose are CC BY 4.0.

Exp 0 pre-registration freeze

Freezes Experiment 0's quantization recipe, metrics, and numeric go/no-go thresholds BEFORE any quantized evaluation. Platform validated: FP16 AR control reproduces BD3-LM Table 5 zero-shot perplexities within 1.6% (exact likelihood); dLLM bound consistent with published values. Eval protocol: bd3lms @ 1c3e8f4, insert_valid_eos=False. GitHub's release timestamp is the server-side pre-registration record; the attached exp0.yaml is the frozen config at this tag.

Pre-registration: Diffusal research program

Timestamps the Diffusal whitepaper — hypotheses (argmax commitment as a quantizer, contraction/expansion crossover), the three-experiment protocol, and the pre-registered kill criteria (§6) — before any experiments were run. The attached PDF is the compiled whitepaper at this commit. GitHub's release creation time serves as the server-side timestamp.