v0.2.0

0.2.0 - 2026-07-23

Release notes

• SAE steering with portable feature artifacts, decoder-direction additions, and latent clamp/ablate operators.
• Sequence-level conditional gates and plan-composition diagnostics for controlled multi-intervention execution.
• Qwen2.5-VL and InternVL adapters with semantic vision/projector sites, image-token and patch selectors, and processor-driven generation.
• Multimodal provenance, compatibility checks, evaluation schemas, and SAE/VLM guides for reproducible steering workflows.

Added

• A backend-independent SAE protocol, lazy SAELens provider, custom/functional
  adapters, common SAELens hook-name-to-site inference, and portable
  SAEFeatureArtifact bundles.
• Supervised SAE feature ranking and causal-effect ranking, including a
  high-level contrastive capture path and supervised shortlisting before a
  user-supplied causal evaluator.
• SAE decoder-direction steering through the ordinary Add operator,
  direct-latent Clamp/Ablate, and residual-preserving
  SAEClamp/SAEAblate.
• ProbeGate, SAEActivationGate, cosine/callable gates, and logical gate
  composition. Explicit language-site gates are evaluated in prefill and
  cached per sequence for decode.
• Plan composition diagnostics for stable execution order, direction norms,
  same-target pairwise cosine, effective rank, condition number,
  orthogonalized bases, additive-fusion candidates, and non-commutative pairs.
• Qwen2.5-VL and InternVL architecture adapters with language, vision-residual,
  projector-input, and projector-output semantic sites.
• An adapter-produced ModalityMap plus serializable ImageTokens,
  ImagePatches, and normalized-XYXY ObjectPatches selectors. Qwen mapping
  accounts for dynamic static-image grids, spatial merge, and window order;
  InternVL mapping accounts for CLS offsets, pixel shuffle, and supported tile
  grouping.
• Static-image processor-driven generation, including native placeholder
  insertion for the unambiguous one-prompt/one-image case and fail-closed
  cardinality checks for batched or multi-image input.
• Processor and modality provenance in artifacts, exact preprocessing
  fingerprint compatibility, and a finite-only multimodal evaluation schema
  split into representation, causal-behavior, and capability/quality metrics.
• SAE/VLM guides and a compensatory-steering example that combines a
  vision-side prefill control and language-side decode control in one
  SteeringPlan.

Changed

• Hugging Face loading now selects an image-to-text model and AutoProcessor
  for recognized VLM configs; generation preserves its validated modality map
  across language decode.
• The compiler resolves and validates gate read dependencies, rejects
  non-language explicit sequence-gate sites, preserves priority/declaration
  order, and exposes composition data through diagnostics() and
  explain().
• A statically zero Constant or NormRelative schedule is compiled as a
  strict no-op: no target/gate hooks run and random-number-generator state is
  unchanged.

Fixed

• Sequence-gate caches are reset at every new generation and cannot leak into a
  first decode call backed by pre-populated KV state.
• Compiled plans cannot be rebound after their source model wrapper has been
  collected, even if Python later reuses the same object identity.
• Composition JSON no longer emits non-finite numbers or compares directions
  from different concrete targets.
• Exact artifact/processor compatibility includes preprocessing
  configuration, not only a declared processor id and revision.
• Public dtype= loading is translated to the Transformers 4.x-compatible
  torch_dtype keyword across the declared dependency range.
• InternVL ObjectPatches fails closed for multi-tile source images when the
  processor does not expose tile-to-original-image crop transforms;
  ImagePatches remains explicitly tile-local.