Skip to content

10 — Glossary

  • Status: Authoritative for v0.1
  • Companion file: docs/glossary.md (synonymous; kept for the existing docs convention).

This is the canonical term registry across the spec corpus, RFCs, and implementation. Any term that appears in more than one document belongs here. When the spec and the glossary disagree, the spec wins; please open a PR fixing the glossary.

ML / architecture

Action. A structured genetic edit (SNV, INS, DEL, MNV, INDEL, SV) passed to the predictor as a first-class input. Canonical type: EditSpec (RFC-0003).

Action embedding. Fixed-size vector representation of an action, produced by the action encoder. Dimension d_action = 512 in v1.

Action encoder. Trainable module mapping EditSpec (or RelEdit) to an action embedding. See RFC-0003.

Autoregressive predictor (ARPredictor). Wrapper around the predictor that unrolls it step-by-step over a sequence of actions, with KV caching, for multi-edit haplotype rollout and planning. See RFC-0004 §3.3.

Carbon. DNA foundation model family released by HuggingFaceBio (500M / 3B / 8B). Used as GenoLeWM's state encoder. See RFC-0002.

Cross-attention predictor. Default predictor architecture: 4 cross-attention blocks (state ↔ action) plus 2 self-attention blocks on the fused sequence. See RFC-0004 §3.1.

JEPA. Joint-Embedding Predictive Architecture. A class of models that predict in representation space rather than input space.

LeJEPA regularizer. Isotropic-Gaussian regularizer on encoder output distributions that prevents collapse. Live in Phase 2 only. See RFC-0005 §3.2.

LeWorldModel (LeWM). Reference architecture and training recipe (Maes et al., 2026) for stable end-to-end JEPA world models. Basis for GenoLeWM's training recipe.

Predictor. Trainable module mapping (state, action) to predicted next-state in latent space. See RFC-0004.

State. Vector embedding of a contiguous DNA window produced by the state encoder (Carbon). Dimension d_state = 1024 for Carbon-500M.

State encoder. Carbon (frozen by default in v1). See RFC-0002.

Surprise. Predictor's residual: distance between predicted post-edit latent and the encoder's actual post-edit latent. Calibrated to a per-context percentile to produce a pathogenicity score. See RFC-0009.

Genomics

bp. Base pairs. Unit of DNA length.

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

Coding / non-coding. Regions of DNA that do (coding) or do not (non-coding) translate to protein.

Edit / variant. Change to a reference DNA sequence. Used interchangeably; "edit" is the action-conditioned framing, "variant" is the genomics framing.

EditSpec. Canonical edit type: (chrom, pos, ref, alt, edit_type). See RFC-0003.

EditType. Enumeration: {SNV, INS, DEL, MNV, INDEL, SV}.

FASTA. Standard text format for nucleotide sequences. Indexed via .fai files (samtools / pysam convention).

gnomAD. Genome Aggregation Database. Reference catalog of human genetic variation, used both for biological-prior edit sampling during training and for the surprise calibration distribution.

Haplotype. Coordinated set of variants on the same chromosome copy. In GenoLeWM, a multi-edit input to the predictor.

Indel. Insertion or deletion of one or more bases.

MNV. Multi-nucleotide variant: substitution of equal-length ref and alt, both > 1 bp.

Pathogenic / Likely pathogenic (P/LP). ClinVar classes for variants asserted to cause disease (P) or strongly suspected to (LP).

RelEdit. Window-relative form of an EditSpec, with rel_pos as the offset within the window in bp. See RFC-0003 §3.3.

SNV. Single-nucleotide variant: substitution of one base for another.

SV. Structural variant: large-scale rearrangement (inversion, duplication, large indel, translocation). Not supported in v1; deferred to v2.

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

VCF. Variant Call Format. Standard file format for representing genetic variants.

Window. Contiguous DNA region passed to the state encoder. Default length 12,288 bp (2,048 Carbon 6-mer tokens).

Operations

Calibration table. Per-context empirical CDFs of σ_raw over gnomAD common variants. The v0.1 model package ships calibration.parquet, but population-stratified calibration validity is not established by the first release. See RFC-0009 §3.4.

Cache (window cache). On-disk Parquet store of pre-computed reference-window embeddings. Content-addressed by (window_hash, encoder_hash, state_layer, pool_type, pool_radius, dtype). See RFC-0002 §3.6.

Encoder hash. SHA-256 of the encoder weights file. Part of every cache key and checksum receipt.

Holdout. Set of data excluded from training and reserved for evaluation. Three holdouts: holdout-chr (chr21), holdout-clinvar, holdout-haplotypes. See RFC-0006 §3.8.

Manifest. Canonical JSON document describing a GenoLeWM checkpoint's identity, configuration, and provenance. Its hash is the model_id. See RFC-0011 §3.7.

Receipt. JSON document that release scoring paths can emit to bind model identity, input commitment, and output commitment. See RFC-0011.

Tuple builder. Data-pipeline component that produces (w_ref, a, w_alt) training tuples. See RFC-0006 §3.5.

Artifact provenance

Checksum receipt. Receipt mode currently supported by GenoLeWM. It binds the model manifest identity, input commitment, output commitment, and runtime metadata; it is not a model-quality or runtime-assurance guarantee.

Content addressing. Identifying data (weights, inputs, outputs) by the cryptographic hash of their canonical serialization, not by name or location.

Input commitment. SHA-256 hash of the canonical serialization of an inference's inputs.

model_id. SHA-256 of a GenoLeWM checkpoint's manifest. Globally identifies a specific released checkpoint artifact set.

Project & process

Current public status: alpha implementation with a completed v0.1 paper/demo release. Phase names below describe the roadmap, not broader model-quality evidence.

Phase 0 (Design, complete). Spec and RFC bootstrap.

Phase 1 (MVP). Carbon-500M frozen, SNVs only, ClinVar coding/non- coding eval. See ROADMAP.

Phase 2 (Full edits). Adds indels, MNVs, LoRA-adapted encoder, LeJEPA regularizer, planning, calibrated surprise. Full eval suite.

Phase 3 (On-device). Export pipeline, quantization, desktop app skeleton.

RFC. Request for Comments. A numbered design document under rfcs/. The source of truth for individual design decisions.

Tracking issue. A GitHub issue tagged type:tracking that lists the implementation issues for a subsystem with a checkbox per child.

Conventions

  • RFC numbers are zero-padded to four digits (RFC-0001).
  • File names are kebab-case after the number (0001-project-scope-and-goals.md).
  • Mathematical notation: lowercase italic for scalars (s, a), lowercase plain for vectors (s_t, a_emb), uppercase for sets / matrices.
  • → denotes function mapping; ≜ denotes definition.
  • Hashes are SHA-256 unless otherwise noted; rendered in lowercase hex.