Repository Issue Activity (beta)

#1769 [Bug Report] stop_at_layer silently no-ops on RavenArchitectureAdapter (Huginn depth-recurrent decoder)

#1791 [Bug Report] _BLOCK_LIST_ATTRS hardcoded name list silently drops Raven's blocks from composition-score / head-label analysis

#1505 [Proposal] Jacobian Lens (J-lens): load published lenses, fit natively, read out and intervene

#1742 [Proposal] Attribution Patching + EAP/EAP-IG: linearized activation patching and faithful edge-level circuit discovery

#1787 [Bug Report] attribution_patch reports non-zero causal effects for identical inputs in train mode

#1784 [Proposal] Validate, generalize, and stress-test Projection-Kernel Head Affinity (follow-up to #1721)

#1755 [Proposal] Relevance Lens (R-lens): a RelP/ LRP-based transport-matrix estimator for Jacobian Lens (J-lens) fit, readout, and intervention

#1776 [Proposal] Causal coordinate-swap benchmark for coordinate_patch_hooks

#1779 [Bug Report] SVDInterpreter returns plausible non-zero MLP directions from fabricated zero weights on MoE models

#1777 [Proposal] Backward Lens: generalize Backward Lens beyond GPT-2 (dense MLP families)

#1772 [Bug Report] IOIDataset double-prepends BOS and ignores prepend_bos=False

#1767 [Proposal] SVD Circuits: singular-vector decomposition of a head's QK/ OV into causally-validated subfunctions

#1746 `JacobianLens.swap_hooks` re-reads live coordinates, causing odd/even cancellation across layer bands

#1728 [Proposal] Add leakage-safe k-sparse probing over activation tensors

#1765 [Bug] JacobianLens reuses a stale full-vocabulary dictionary after unembedding weights change

#1743 [Bug Report] maybe_cast_floating_params skips dtype normalization for an entire quantized model, not just quantizer-owned tensors

#1686 [Proposal] Backward Lens: vocabulary readout of multilayer perceptron gradient factors

#1751 [Bug Report] TransformerBridge: NeoX/Pythia unembed still resolves to embed_out instead of lm_head (transformers >= 5.13)

#1748 [Proposal] Dynamic hooks for J-space coordinate patching

#1760 [Bug Report] hf_generate right-pads batched strings and drops the attention mask

#1762 [Bug] JacobianLens.fit silently mixes dropout realizations in train mode

#1756 [Bug Report] generate_stream(return_type="str") decodes only the first sequence of a batch

#1726 [Bug Report] Cohere compatibility mode double-applies logit_scale in forward

#1739 [Proposal] Add J-space coordinate patching over sparse decompositions

#1665 [Proposal] Add a bridge-vs-HF numerical parity test for BERT