Glossary

Architecture

Action. A structured genomic edit passed to the predictor as a first-class input.

Action encoder. The trainable module that maps an edit into a vector embedding.

Autoregressive rollout. Repeatedly applying the predictor over a sequence of edits, feeding each predicted state into the next step.

Carbon. The HuggingFaceBio DNA foundation model used as GenoLeWM's frozen state encoder in the released path.

Carbon pure-DNA tokenizer. The self-contained tokenizer used by the corrected source path for GenoLeWM's restricted <dna>...</dna> input. It is constructed from local Carbon configuration and does not execute the upstream wrapper's unpinned Qwen/Qwen3-4B-Base lookup. This runtime repair is not model-quality evidence.

Cross-attention predictor. The trainable model that combines a latent state and action embedding to predict the post-edit latent state.

JEPA. Joint-Embedding Predictive Architecture: a model family that predicts in representation space instead of reconstructing raw input.

Predictor. The trainable GenoLeWM module that maps (state, action) to a predicted next state.

State. A latent vector embedding of a DNA window under an explicit state contract.

State encoder. The frozen DNA encoder that maps a sequence window to a state. The released path uses Carbon-500M.

State contract. The transformation, scale, and pooling coordinate shared by source, target, candidate, and predicted states. legacy_raw_v1 preserves historical raw pooled Carbon states; l2_normalized_v2 applies L2 normalization after coordinate-matched pooling and is the intended contract for new training.

DNA content start. The hidden-state index of the first six-mer after the leading <dna> control token. Correct centered pooling uses dna_content_start + edit_locus // 6. Historical code omitted dna_content_start, shifting every intended center one hidden token left and sometimes centering on the control token.

Cache schema 3. Raw post-pooling, pre-normalization Carbon states keyed by encoder runtime identity, layer, pooling mode/radius, center_token, and logical compute dtype. Pooling emits canonical FP32 before consumption/storage; normalization also rounds its final view to canonical FP32. New Parquet shards separately declare fixed-size FP32 physical storage and use SHA-256-derived ASCII components for encoder ID and contig. Strict shard/index validation, serialized no-clobber publication, and schema/encoding provenance prevent silent collisions or ambiguous resume. Cache v2 remains readable only under an explicit labeled replay policy; corrected training and rollout evidence require v3. Cache v1 omitted center_token and is intentionally invalidated. No full cache build or corrected model run is claimed by this schema foundation. Race-resistant cache I/O is supported on Linux and macOS and fails closed on Windows or when secure directory-descriptor/no-follow primitives are unavailable; no path-only mutation fallback is used.

Finite cache build. A request-scoped geno-lewm-cache-windows operation driven by immutable JSONL rows containing a DNA window and its edit_locus. The committed plan preserves distinct center_token keys for the same window, binds batch, hardware/device, and resolved-config identity, and is rederived exactly from immutable inputs on recovery. The resumable state SHA-256-binds each evidence-owned shard before any new encoder batch runs. Existing global logical keys are inspected and reused; only misses are encoded. Its report binds request-scoped immutable shard mappings, not mutable shared-index bytes, and proves only the exact supplied request artifact; it is not evidence of 10% corpus coverage or the 24-hour hardware target.

Surprise. A candidate residual between predicted and encoded post-edit states under the same validated state contract. Published legacy_raw_v1 values mixed state scales and are not valid scientific surprise scores.

Genomics

bp. Base pairs.

ClinVar. NCBI's public database of human variants with clinical significance labels.

Coding / non-coding. DNA regions that do or do not translate to protein.

Edit / variant. A change to a reference DNA sequence.

EditSpec. GenoLeWM's chromosome-position-reference-alternate edit object.

RelEdit. Window-relative edit object used after placing an edit inside a sequence window.

gnomAD. Genome Aggregation Database, used by GenoLeWM data and calibration tooling.

Haplotype. A coordinated set of variants on the same chromosome copy.

Indel. Insertion or deletion of one or more bases.

MNV. Multi-nucleotide variant.

SNV. Single-nucleotide variant.

SV. Structural variant. Large SV support is not established by the public release.

TraitGym. A benchmark of trait-associated variants for variant-effect prediction.

VCF. Variant Call Format.

Window. A contiguous DNA sequence region passed to the state encoder.

Scores And Artifacts

Calibration table. Table that maps raw residuals to contextual calibrated values. The published legacy table was fitted to invalid mixed-contract residuals and is retained only for artifact replay.

Checksum receipt. JSON document binding model identity, input commitment, output commitment, and runtime metadata. It is not a model quality or runtime-assurance certificate.

Content addressing. Identifying an artifact by the cryptographic hash of its canonical bytes. It binds only included files and dependencies; hashing historical Carbon tokenizer.py did not bind its unpinned transitive Qwen tokenizer load.

Input commitment. Hash of an inference input payload.

Manifest. JSON document describing a model package's artifact identity and provenance.

Model id. The manifest hash for a specific released checkpoint artifact set.

Output commitment. Hash of a score output payload.

sigma_raw. Uncalibrated latent residual. Published legacy_raw_v1 values are invalid as edit-effect or surprise scores.

sigma_calibrated. Calibration-table mapping of sigma_raw. Published legacy values are historical implementation outputs, not validated scores.

Tuple builder. Data-pipeline component that produces (reference window, edit, edited window) training examples.