Skip to content

Release Notes

v1.0.0 — 2026-05-16

First stable release. This release closes a correctness-hardening cycle: an audit found and fixed 10+ serious bugs in code that had previously been marked "done." The public API is stable and the Stable-tier discovery path is validated. Experimental- and research-tier backends remain explicitly marked as such.

Highlights

Stable discovery path validated

  • EAP and EAP-IG are validated across GPT-2, Llama-3.2-1B/3B, Gemma-2-2B, Gemma-3-4B, and Qwen2.5-1.5B
  • EAP-IG-Activations and EAP-Clean-Corrupted remain Research tier — validated only on GPT-2 IOI
  • 13 discovery algorithms across 4 backends (EAP, ACDC, IBCircuit, CD-T) — but only 2 are validated at production scale (eap, eap-ig); acdc and ibcircuit are Experimental (GPT-2 scale, ibcircuit OOMs above ~3B) and the other 9 are Research (GPT-2 IOI only)
  • 14 registered selectors; 16 built-in tasks
  • 6-pillar faithfulness framework with Pillar 6 marked preliminary

bf16 / GQA correctness

Discovery backends (PEAP, CD-T, ACDC) and all applications (knowledge editing, CircuitWeightSteering, circuit-localized fine-tuning) now run correctly on bf16 models and grouped-query-attention architectures.

Chat-template handling

Per-task chat_template_mode (auto/on/off), BOS consistency across all components, auto chat-model detection, and discovery freezing the resolved policy into artifact metadata.

Checkpoint export protocol

Intervened (pruned/quantized) checkpoints are exported as standard reloadable HuggingFace checkpoints and are vLLM-evaluable via lm-eval. 8 checkpoint export/eval correctness bugs fixed.

API surfaces

  • Flat typed API (ck.discover, ck.prune, ck.quantize, ck.faithfulness, ck.export_checkpoint, ck.benchmark, ck.load_scores, ck.selective_finetune, ck.visualize_circuit)
  • Pipeline class with method chaining and from_artifact / from_scores constructors
  • Unified scope / protect_layers component selection across all applications

Removed

  • lrp pruning selector removed (orphan — replaced by taylor and wanda)
  • Old Sphinx documentation replaced by MkDocs

Known Issues

  • Pillar 6 (Generalization) is implemented but not validated at scale — treat scores as preliminary
  • IBCircuit OOM above ~3B parameters on single-GPU
  • ROME knowledge editing: MLP-only target selection (attention target selection is not yet implemented)

For the full bug-fix history, see CHANGELOG.md in the repository root.