# Pipeline Claims and Provenance Every pipeline stage in this repository is classified under `reproduction` — implementing an EvoFlows or Edit Flows decision as empirical evidence regarding the published method. Stages and classifications are tracked in `editjumps/core/provenance.py` and validated by repo invariant tests. Regenerate the table below with `make provenance`. ### `reproduction` — 14 stages Implements an EvoFlows/Edit Flows decision. Evidence about the paper. | stage | why | |---|---| | `download_oas` | §3.1/§4.2 sequence database — they use UniRef30 + a ColabFold environmental DB; antibody-only OAS is ours **(deviation)** | | `seed_homologs` | built only to put numbers in their Table 2's columns | | `split_corpus` | §4.2 train / inference / holdout split; their ratios are not stated | | `fetch_base_checkpoint` | §3.4 'the encoder trunk of a pre-trained ESM-2 model' | | `pretrain_esm` | §3.4 states NO domain-adaptive pretraining; our OAS-adapted trunk is arm A of the 2x2, an addition rather than a reproduction step **(deviation)** | | `build_homolog_pairs` | §4.2 pair construction + eq 7-8; MMseqs clustering replaces their iterative profile search, and pairs are capped per family **(deviation)** | | `train_edit_flows` | eq 6 Bregman rate-matching loss, eq 13-16 architecture | | `build_deterministic_pairs` | §4.1's synthetic dataset — the rules verbatim, applied to natural sequences | | `train_deterministic_editor` | §4.1 trains on those pairs; evaluating an editor that never saw the rules measures zero-shot transfer, not the paper's experiment | | `deterministic_benchmark` | §4.1 synthetic ground truth — the only evaluation with a known answer | | `evotune_esm` | §2.2 Evotuning (Alley et al., 2019) — the same MLM objective on one homolog family; the family and its train/holdout split are ours | | `evotune_baseline` | §4.3's 'Evotuned PLM' baseline — eq 12 entropy profile picks positions, the evotuned MLM infills them | | `evotune_baseline_forced` | §4.3's 'Evotuned PLM with forced substitutions' — same, with the original amino acid blocked at every masked position | | `evodiff_msa_baseline` | §4.2's 'EvoDiff-MSA (Alamdari et al., 2024)' baseline — their released MSA model and weights; the mutation-budget masking that makes it §4.2-matched is ours **(deviation)** | Specific experimental decisions and deviations are detailed in the reproduction ledger write-up. Model cards citing the write-up note numbers as `LEDGER-SOURCED`. Empirical measurements reproducible from repository artifacts are recorded in [`docs/findings.md`](findings.md). ## Target Property Scope EvoFlows generation is unconditional; the editor does not incorporate target property conditioning, and all reported numbers reflect unconditional generation. To prevent inadvertent misconfiguration: - `property` in `params.yaml` / `TargetConfig`: only `null` is supported (`editjumps/core/run_config.py`). - `homolog_pairs.direction` in `params.yaml`: property-based orientation is rejected (`editjumps/pipeline/preprocess/pretrain/homolog_pairs.py`). - Pair orientation checks in the training pipeline are skipped in unconditional mode (`editjumps/pipeline/train/train_edit_flows.py`).