Self-edit evidence, backend compatibility, and larger model lanes

Release: InvarLock 0.12.0 - Self-edit workflow evidence, backend compatibility, and larger-model public basis

Highlights

• Knowledge-edit and self-edit workflows now have a documented evidence shape: external editor, subject checkpoint or declared runtime artifact, baseline-vs-subject evaluation, verification, and optional evidence pack.
• BYOE edit evidence is broader, with LoRA merge, full fine-tune, and magnitude-prune lanes plus training evidence campaign artifacts, checkpoint references, runtime manifests, verification summaries, and hash inventories.
• Public evidence is organized around clearer categories for runtime backend compatibility, attention-backend behavior, lifecycle stress/resume evidence, larger-model validation findings, and published-basis model lanes.

0.12.0 is about widening what can be reviewed without blurring the claim boundary. InvarLock still verifies a configured baseline-vs-subject comparison. The new Knowledge & self-edit workflows guide explains how to carry context for knowledge edits, self-adaptation loops, LoRA merges, fine-tunes, pruning, quantization, dynamic runtime adapters, and privacy-sensitive deltas without turning those labels into hidden strict gates.

That guide introduces optional metadata for evaluation realism, edit provenance, edit-impact scenarios, topology, and delta privacy. The important constraint is explicit: fields such as edit_family, target_set_digest, metric_is_generation_realistic, proxy_metric_warning, artifact_kind, and delta_privacy help readers understand the subject artifact, while the verifier verdict remains tied to the selected report and runtime-provenance contract.

The public evidence surface also grows. The Public Evidence Walkthrough now includes separate BYOE examples for magnitude pruning, LoRA merge, and fine-tune subjects. Evidence Packs and Evidence Pack Internals describe the validation-subject taxonomy, tuned clean edits, stress edits, deployable edit lanes, stricter release-review packaging, queue reliability controls, and package-native verification behavior.

Model and backend coverage get a clearer map as well. The synced Model Family Catalog and Model Adapters include updated published-basis distinctions for Gemma, Qwen, GPT-OSS, Phi, Falcon, OpenLLaMA, and related lanes. The upstream release also records CUDA 12.8 runtime backend compatibility evidence for bitsandbytes, GPTQModel, HQQ, Quanto, TorchAO, and compressed-tensors, plus attention-backend evidence for FA2-unavailable eager execution.

For the immutable release record, read the tagged CHANGELOG.md for v0.12.0.