# Task-specific model cards

Common task: predict a 2,000-HVG HepG2 perturbation delta for genes held out by identity. The five outer seeds are 20260712–20260716. Their test sets overlap and are interpreted as retrospective stability analyses. Raw cosine relative to train mean is the primary model-direction endpoint; RMSE is secondary; common-effect-adjusted residual cosine is diagnostic.

## Zero-delta and train-mean controls

- Role: mandatory no-effect/common-effect references, not trained models.
- Coverage: all 540 eligible targets in each outer split.
- Output: either a zero vector or the mean training-target delta.
- Interpretation: a model that does not exceed train-mean raw direction cannot claim perturbation-specific biological direction, even if RMSE or residual cosine improves.

## Ridge v1 and v2

- Implementation: `scripts/run_repeated_heldout_gene_ridge_benchmark.py` using standardized inputs and ridge penalties selected from 0.1, 1, 10, 100 and 1000 by three-fold training-only cross-validation.
- Data: 2,160 targets with at least 20 cells; 25% held out after cell-count-quintile stratification.
- Coverage: 540/540 per split.
- v1 feature emphasis: dataset summaries and DepMap functional-dependency features. v2 adds disease/context features.
- Result: v1 RMSE improves 8.5–10.1% versus zero and 4.5–5.2% versus train mean. Raw cosine gain versus train mean ranges from −0.0041 to +0.0007 and does not pass the primary direction gate. Residual cosine is 0.242–0.267.
- Leakage boundary: DepMap 26Q1 post-dates the benchmark work and contains CRISPR dependency outcomes. A leave-HepG2-out rebuild preserves broad RMSE gains but is an outcome-proxy sensitivity, not temporal validation.
- Permitted claim: narrow same-dataset internal RMSE/residual generalization.
- Prohibited claim: leak-free evaluation, raw direction superiority, cross-cell-type generalization or DMD prediction.

## Published linear-PCA reimplementation

- Implementation: `scripts/run_nature_methods_linear_baseline.py`.
- Hyperparameters: PCA dimension 10, ridge penalty 0.1, random state 20260712.
- Native support rule: perturbation target must occur in the response-gene space.
- Coverage: 48–70 of 540 targets per split (8.9–13.0%).
- Fair comparison: ridge v1 was rescored on the identical covered genes.
- Result: neither linear-PCA nor ridge v1 consistently wins the strict same-coverage composite gate.
- Boundary: low native coverage is part of the result; missing targets are not silently removed from the task denominator.

## GEARS

- Implementation: cell-gears 0.1.2, PyTorch 2.5.1+cu121, PyG 2.8.0; five 20-epoch CUDA runs, hidden size 64, training batch 32 and test batch 128.
- Compatibility patches: SHA-256 `804fc556...3805`, `14bbf0c3...361` and `65bd5f9a...cba7`; the full hashes are in `gears_full_benchmark_preflight.json`.
- Runtime exclusions: 11 symbol-case unsupported conditions were conservatively removed.
- Native coverage: 526–535/540 targets. Common comparison: ridge v1 on the same targets.
- Checkpoints: per-seed `model.pt` files retained; SHA-256 values are `5e84a38a...be5`, `2a7398ac...b11`, `0bc33db3...20d`, `f3089579...e25` and `fb42cf3f...a63`.
- Result: GEARS RMSE 0.1304–0.1386 is worse than both no-effect baselines; raw direction also fails. Ridge v1 is better on identical coverage for RMSE, raw cosine and residual cosine in every split after Holm correction.
- Boundary: this is a negative result for this task, runtime and support policy, not a general verdict on GEARS.

## scGPT

- Implementation: scGPT 0.2.5, source commit `cebd6fae655b9c585a4807daa3ac31bb764f06b4`; cell-gears 0.0.2; PyTorch 2.5.1+cu121; five 15-epoch CUDA runs; early stop 10; batch 64; maximum sequence length 1,536; learning rate 1e-4.
- Whole-human checkpoint hashes: `args.json` `c18e075e...d60`; `vocab.json` `acca93d1...80f`; `best_model.pt` `6cb5d451...922`.
- Native support rule: targets must occur in both response-gene space and checkpoint vocabulary.
- Coverage: 46–67/540 targets per split. Common comparison: ridge v1 on identical targets.
- Result: scGPT fails RMSE and raw-direction support on all splits and has higher RMSE than ridge v1 on each common set; residual-only support is diagnostic.
- Boundary: this is a checkpoint/interface-limited negative task result, not a general verdict on foundation models.

## STATE

- Implementation provenance: Arc Institute STATE, observed source commit `cc22e69`, arc-state 0.11.1; server environment recorded in `reports/STATE_SERVER_READINESS_2026-07-06.md`.
- Task variants: one-hot, feature, seen-perturbation and supervised-delta experiments; all retained in the GPU closure manifest.
- Result: current STATE configurations fail the no-effect gates. The environment and inference route work, but the tested training objectives do not support a positive held-out-gene claim.
- Boundary: failed STATE runs are first-class negative evidence; they do not imply that all STATE configurations or datasets fail.

## Seen-perturbation measured-delta positive control

- Implementation: `scripts/build_seen_pseudobulk_delta_benchmark.py`.
- Task: cell-heldout evaluation for perturbation identities already observed in training; 96 targets.
- Test result: RMSE 0.1381 versus zero 0.1526 and train mean 0.1525; cosine 0.3623; declared gate passes.
- Role: shows that the scoring stack can register recoverable signal.
- Boundary: it is a reference ceiling for seen-perturbation semantics, not a deployable unseen-identity predictor.

## Reproducibility status shared by all cards

Scores, gates, coverage manifests, predictions and principal checkpoint/source hashes are retained. Deterministic local resource regeneration passed in an isolated output tree, but a DOI-backed public snapshot and portable independent GPU/full-model rerun remain pending. The cards therefore document the local/remote release candidate rather than a final archival release.