family_id	title	core_mechanism	best_fit_question	input_contract	output_contract	strengths	failure_modes	minimum_data	permanent_baselines	evaluation_requirements	reference_methods	local_state	local_runs	next_gate
MF-01	Linear & non-parametric baselines	Predict no change, a matched mean, a regularized linear response or a nearest observed neighbor.	Is there learnable signal beyond systematic assay and context structure?	matched feature space; frozen split; context labels for matching where permitted	aggregate response vector or neighbor-based reference	transparent | low variance | fast leakage diagnostic | permanent comparator	cannot represent complex interactions | matching can leak held-out context | good mean error may miss perturbation identity	one task-compatible training response set plus held-out truth	zero change | train/control mean | perturbed or matching mean when legal | ridge | nearest neighbor when legal	mean fidelity | perturbation-specific delta | split integrity	Systema | ridge | matching mean	LOCALLY_EXECUTED_LIMITED	MRUN-RIDGE-SAFE-2.3-G0-REPEATED-FOLD	Retain ridge as a permanent same-assay comparator; do not extend it to DMD transfer.
MF-02	Factorized latent generative models	Disentangle basal cell state, perturbation and covariates in a latent representation, then compose an unseen condition.	Can known factors be recombined across dose, time or cell context?	cell-level expression with perturbation, covariate, dose/time and control labels	counterfactual cells or an expected post-perturbation distribution	compositional representation | covariate conditioning | counterfactual generation	disentanglement is not guaranteed | performance declines as unseen covariates accumulate | batch may be encoded as biology	factorial overlap across perturbation and covariate combinations	matching mean | context-conditioned ridge | nearest observed condition	mean fidelity | delta recovery | distribution fidelity | OOD covariate stress test	CPA | scGen | biolord	SOURCE_VERIFIED_NOT_RUN		Run only after a factorial task exposes which covariate combinations are genuinely unseen.
MF-03	Graph-informed perturbation models	Propagate gene and perturbation information over co-expression, ontology or learned graphs.	Can structured gene relationships improve unseen-gene or combination response prediction?	perturbation responses plus gene identities and a versioned graph or embedding source	post-perturbation mean expression or effect vector	uses gene relationships | supports structured inductive bias | can represent combinations	graph mismatch | single-perturbation data may not identify combinations | relation priors can dominate sparse truth	single and preferably combinatorial perturbations with stable gene coverage	additive single-perturbation baseline | ridge | matching mean	held-out genes | held-out combinations | sign and interaction recovery | seed stability	GEARS | TxPert	LOCALLY_EXECUTED_FAILED_OR_PENDING	MRUN-GEARS-0.1.2-FIVE-SEED-20260713 | MRUN-TXPERT-CONFIG-GAT-SEED-20260712	GEARS remains failed on five frozen splits; TxPert remains calibration-pending. Neither unlocks DMD prediction.
MF-04	Foundation transformer & embedding models	Pretrain token or rank-based cell representations at scale, then adapt embeddings or decoders to a downstream task.	Does broad pretraining improve a precisely frozen perturbation or cell-state task?	gene-aligned expression plus the exact tokenizer, vocabulary, checkpoint and adaptation recipe	cell embeddings, labels or decoded response vectors depending on the adapter	broad representation prior | transferable embeddings | large reference context	embedding quality is not perturbation accuracy | vocabulary/context mismatch | scale can obscure task leakage	checkpoint-compatible genes plus task-specific adaptation and held-out truth	PCA or linear embedding | ridge | train mean | task-specific shallow model	task-level baseline comparison | OOD split | ablation of pretraining | reproduction receipt	scGPT | Geneformer | scFoundation | CellFM	LOCALLY_EXECUTED_FAILED	MRUN-SCGPT-0.2.5-FIVE-SEED-20260714	The local scGPT adapter lost to same-coverage baselines; a new checkpoint is a new frozen ModelRun, not an inherited upgrade.
MF-05	Set-to-set & context-prompted models	Represent a cell population as a set and condition one set of cells on another context or prompt.	Can a model predict population transitions while using context examples at inference time?	cell sets, context labels, perturbation identity and a task-compatible feature universe	predicted cell set, state embedding or transition distribution	population-native | context prompting | heterogeneity-aware representation	prompt leakage | set composition confounding | source-reported scale may not transfer to disease context	multiple comparable cell populations per context and held-out population truth	matching population | stratified mean | optimal-transport baseline	distribution fidelity | composition recovery | prompt ablation | unseen-context holdout	Arc State | Arc State/Stack	COMPATIBILITY_ROUTE_EXECUTED_NO_ADVANCEMENT	NMD-B1-COMPAT-HEPG2-R1-20260816	STATE: Do not rerun by reputation; require a corrected preregistration, untouched truth or materially different task-compatible checkpoint. Stack: Identify a task-compatible pretrained genetic checkpoint and preregister an untouched benchmark before any model-performance claim.
MF-06	Optimal-transport population maps	Learn a transport map from an unpaired control population to a treated population.	How does a distribution of control cells move under treatment when cells are not paired?	unpaired control and treated cell populations with shared features and sufficient state coverage	transported cells or a treatment-conditioned population distribution	distributional output | unpaired design | population geometry	rare states are unstable | transport assumptions may not identify biology | composition shifts can mimic state transitions	adequate cells across represented states in both conditions	identity map | mean shift | nearest-neighbor transport	MMD or energy distance | Wasserstein distance | state composition | rare-state stratification	CellOT	SOURCE_VERIFIED_NOT_RUN		Use only for a population-output task with explicit rare-state and composition stress tests.
MF-07	Probabilistic sparse-effect models	Estimate interpretable perturbation effects with a probabilistic prior and calibrated uncertainty.	Which gene-level effects are supported, and where should the model abstain?	replicated perturbation responses with biological units, covariates and stable features	effect posterior, uncertainty interval and sparse active set	uncertainty-aware | interpretable effects | appropriate for sparse signals	prior sensitivity | poor scaling | cell-level pseudo-replication produces false confidence	independent biological replication, not cells treated as replicates	regularized linear model | empirical Bayes shrinkage | no-effect model	interval coverage | calibration | effect-sign recovery | donor-level resampling	GPerturb	SOURCE_VERIFIED_NOT_RUN		Prioritize when replicated DMD perturbation outcomes exist and uncertainty is decision-critical.
MF-08	Distributional flow & diffusion generators	Learn a conditional generative process that samples heterogeneous post-perturbation cells.	Can the full conditional response distribution be generated rather than only its mean?	large cell-level perturbation datasets with context, dose/time and robust controls	sampled post-perturbation cell population	multimodal distributions | heterogeneity | sample-level counterfactuals	plausible-looking hallucinated states | mode collapse | weak calibration | high compute burden	large, balanced and context-rich perturbation populations	matching population | CellOT or transport baseline | conditional Gaussian baseline	distribution metrics | mode coverage | calibration | biological state validity | prospective function	conditional flow matching | diffusion perturbation models	WATCHLIST_NOT_RUN		Do not adopt until the population task, compute budget and prospective validation route are frozen.