Global model & competitor watch · 2025-08-16 → 2026-08-16

Stop counting models. Start grading evidence.

The newest model is a source lead, not a local capability. NMD-VCell now binds every external update to a TaskContract, leakage audit, permanent baselines and the highest claim it could legally support.

Inspect TaskContract 1.0 Open full registry

What is known?Evidence AtlasMeasured, contextual and missing evidence.What can be estimated?Task-bound benchmarksOnly within the highest passed claim level.What still needs experiment?Missing target-context outcomeDonor × perturbation × time × function.

v1.2.1-model-intelligence · frozen 2026-08-17

The first compatibility release is frozen—and it does not upgrade DMD claims.

GSE264667/Nadig TRADE HepG2 CRISPRi · five stratified 25% perturbation-gene holdouts. Highest permitted claim: E1 IN DOMAIN HEPG2 HELDOUT PERTURBATION ONLY NO DMD PREDICTION.

Patch manifest Execution JSON Download report

ReleaseNMD-B1-COMPAT-HEPG2-R1-20260816

Independent seeds5

Eligible perturbations2,160

Response coordinates2,000

ridge_v1EXECUTED FIVE SEED

5/5RMSE gate

5/5Perturbation-specific gate

0/5Raw composite gate

2160 eligible perturbations

ridge_v2EXECUTED FIVE SEED

5/5RMSE gate

5/5Perturbation-specific gate

0/5Raw composite gate

2160 eligible perturbations

nature_methods_linear_pca10EXECUTED FIVE SEED SAME COVERAGE

0/5RMSE gate

0/5Perturbation-specific gate

0/5Raw composite gate

test coverage 0.0963–0.1167

ridge_v1_same_coverageEXECUTED FIVE SEED SAME COVERAGE

0/5RMSE gate

0/5Perturbation-specific gate

0/5Raw composite gate

test coverage 0.0963–0.1167

Adapter decisions

TxPertE0 EXECUTION ONLY

HISTORICAL LOCAL GATE FAIL NO REPEAT

STOP UNTIL NEW PREREGISTRATION OR UNTOUCHED TRUTH

STATEE0 EXECUTION ONLY

HISTORICAL RECEIPTS MIGRATED

NO ADVANCEMENT · 8 receipts

StackE0 EXECUTION ONLY

LOCAL NVME CORE BASE FORWARD PASS NO CHECKPOINT

NO NMD B1 BENCHMARK NO COMPATIBLE PRETRAINED GENETIC CHECKPOINT · 2 receipts

Tahoe-x1E0 SOURCE ONLY

SOURCE PINNED INPUT CONTRACT FAIL

DO NOT RUN ON LOG NORMALIZED INPUT; ACQUIRE RAW COUNTS FIRST · 2 receipts

SLIME0 SOURCE ONLY

PRIMARY SOURCE PENDING

NOT ELIGIBLE FOR EXECUTION

Canonical model states Stack NVMe receipt Tahoe input audit Baseline results Adapter decisions STATE migrations Source identities

TaskContract 1.0 · DRAFT_NOT_REGISTERED_OUTCOME_MISSING

A model enters only after the biological task is frozen.

Can a frozen model predict a candidate-conditioned response in a previously unseen DMD biological donor and disease-relevant myogenic state?

Template JSON Schema

01disease

DMD

02donor policy

Hold out complete biological donors; cells are never independent donor replicates.

03cell state

Disease-relevant human myogenic state; exact maturation state must be frozen before registration.

04perturbation

Candidate intervention identity and modality must be frozen; no healthy-myoblast or HepG2 substitution.

05dose

Prespecified per intervention; missing until the experimental protocol is registered.

06time

Prespecified baseline and post-intervention time points; missing until the experimental protocol is registered.

07readout

transcriptomic response · cell-state distribution · at least one disease-relevant functional endpoint

08target quantity

perturbation-specific delta · response direction · population shift · functional response

09split policy

Donor-disjoint first; perturbation, cell-state, batch and disease overlap reported separately.

10metric bundle

RMSE/MAE · signed delta correlation · DE-gene recovery · pathway recovery · distribution distance · calibration/coverage · functional endpoint

11prohibited claims

clinical efficacy · treatment recommendation · patient digital twin · DMD prediction before matched truth exists

Current truth: MATCHED DMD CANDIDATE RESPONSE OUTCOME MISSING

E0–E6 claim ladder

Every stronger sentence requires a stronger holdout.

Passing one level never silently upgrades the next.

Download ladder →

E0Pipeline execution

random cell or engineering smoke split

Can prove: The adapter, data path and metric code execute.
Cannot prove: Biological generalization.

Frozen task, complete receipts and no fatal schema error.

E1Held-out perturbation

unseen perturbation in a matched context

Can prove: Bounded perturbation interpolation or extrapolation in that context.
Cannot prove: New-donor, new-cell-state or disease transfer.

Beat every legal permanent baseline across seeds on perturbation-specific metrics.

E2Held-out donor

complete biological donor

Can prove: Donor transfer within the frozen disease, state and assay.
Cannot prove: New cell state or disease transfer.

Donor-disjoint split, donor-level uncertainty and no cell pseudo-replication.

E3Held-out cell state

unseen cell state or cell type

Can prove: Cross-state transfer for the frozen output contract.
Cannot prove: Cross-disease or prospective validity.

State-disjoint evaluation and state-compatible baselines.

E4Held-out disease or context

unseen disease or biological context

Can prove: Retrospective cross-context transfer on a named dataset.
Cannot prove: Future experimental performance.

Disease/context-disjoint truth with no pretraining exposure inheritance.

E5External retrospective dataset

independent external dataset

Can prove: External retrospective validity within the named dataset and endpoint.
Cannot prove: Prospective or clinical validity.

Independent source, frozen preprocessing, external units and complete failure reporting.

E6Prospective preregistered experiment

future independent donors or experiments

Can prove: Prospective task-level evidence when the preregistered endpoint and baseline gate pass.
Cannot prove: Clinical efficacy, treatment selection or patient-level validity.

Immutable Prediction before Outcome, independent experiment, calibration and all negative branches returned.

Permanent baseline suite

Eight reality checks cannot be removed from the leaderboard.

Deep models advance only on identical data, split, entities, genes, seeds and metric code.

BASE-00identity / no changeTests whether any modeled shift adds information.

BASE-01held-out or train meanTests whether the model exceeds systematic average response.

BASE-02perturbed or matching meanControls systematic perturbation effects when legal for the split.

BASE-03ridge / linearPermanent transparent learned comparator.

BASE-04nearest neighbourTests whether a nearby observed context already explains the result.

BASE-05PCA + regressionTests compact representation without foundation-model complexity.

BASE-06random forestTask-compatible simple nonlinear comparator.

BASE-07context onlyTests whether perturbation identity contributes beyond context.

Priority model watch

Five research priorities, five different evidence states.

TxPert is already a local failed gate; SLIM remains blocked on primary-source verification.

Download model watch →

WATCH-TXPERT-2026P0 ALREADY EXECUTED NO ADVANCEMENT

TxPert

Unseen single and double perturbations plus cross-cell-line transcriptomic response.

Source state: PEER REVIEWED PRIMARY AND OFFICIAL CODE
Local state: EXECUTED LOCAL GATE FAIL
Local result: The released NMD-VCell seed did not beat train mean on its frozen HepG2 gate; confirmatory seeds remain locked.
NMD evidence: NOT ESTABLISHED

WATCH-STATE-2025P1 EXECUTED NO ADVANCEMENT

STATE

Population-level perturbation response across experimental contexts.

Source state: OFFICIAL SOURCE AND CODE VERIFIED
Local state: EXECUTED 8 LEGACY CONFIGS ALL GATE FAIL
Local result: Eight CPU/GPU and held-out/seen-perturbation receipts were migrated; none beat both zero and train mean on every split.
NMD evidence: NOT ESTABLISHED

WATCH-STACK-2026P1 BLOCKED TASK COMPATIBLE CHECKPOINT

Stack

In-context representation and condition transfer; official source reports pretraining on 149 million cells.

Source state: OFFICIAL SOURCE PREPRINT AND CODE VERIFIED
Local state: LOCAL NVME CORE SMOKE PASS NO COMPATIBLE CHECKPOINT
Local result: After one 180-second dependency-path timeout, a bounded local-NVMe retry passed the official StateICLModelBase forward path on a V100 with finite outputs. No pretrained checkpoint or task benchmark was run.
NMD evidence: NOT ESTABLISHED

WATCH-TAHOE-X1-2025P1 BLOCKED RAW COUNT INPUT

Tahoe-x1

70M, approximately 1B and 3B checkpoints linked to Tahoe-100M.

Source state: OFFICIAL MODEL CARD AND WEIGHTS VERIFIED
Local state: SOURCE PINNED INPUT CONTRACT FAIL
Local result: The frozen 145,473-cell H5AD is log-normalized, has no raw layer and therefore fails Tahoe-x1's raw-count input contract; no model run was claimed.
NMD evidence: NOT ESTABLISHED CANCER TO MUSCLE SHIFT

WATCH-SLIM-2026P1 HOLD SOURCE VERIFICATION

SLIM

The supplied research report describes a STRING-informed linear perturbation comparator.

Source state: RESEARCH REPORT LEAD PRIMARY SOURCE PENDING
Local state: NOT ELIGIBLE FOR EXECUTION
Local result: No source-verified artifact and no local run.
NMD evidence: NOT ESTABLISHED

Leakage Auditor 1.0

Unknown overlap is a blocker, not a pass.

Unknown overlap never passes a claim gate; it remains an explicit migration blocker.

Audit schema Download audit migration

ModelRun	Task	Audit state	Donor	Batch	Claim ceiling	Decision
`MRUN-RIDGE-SAFE-2.3-G0-REPEATED-FOLD`	G0	INCOMPLETE LEGACY RECEIPT MIGRATION REQUIRED	NOT ASSESSED NO DONOR AXIS IN RELEASED TASK	NOT RECORDED IN NORMALIZED FIELD	E1 LIMITED SAME CONTEXT	LIMITED PASS SAME CONTEXT ONLY
`MRUN-TRANSFER-DIAGNOSTIC-1.0-G1`	G1	INCOMPLETE LEGACY RECEIPT MIGRATION REQUIRED	NOT ASSESSED NO DONOR AXIS IN RELEASED TASK	NOT RECORDED IN NORMALIZED FIELD	E0 EXECUTION ONLY	FAIL
`MRUN-GEARS-0.1.2-FIVE-SEED-20260713`	G0	INCOMPLETE LEGACY RECEIPT MIGRATION REQUIRED	NOT ASSESSED NO DONOR AXIS IN RELEASED TASK	NOT RECORDED IN NORMALIZED FIELD	E0 EXECUTION ONLY	FAIL
`MRUN-SCGPT-0.2.5-FIVE-SEED-20260714`	G0	INCOMPLETE LEGACY RECEIPT MIGRATION REQUIRED	NOT ASSESSED NO DONOR AXIS IN RELEASED TASK	NOT RECORDED IN NORMALIZED FIELD	E0 EXECUTION ONLY	FAIL
`MRUN-TXPERT-CONFIG-GAT-SEED-20260712`	G0	INCOMPLETE LEGACY RECEIPT MIGRATION REQUIRED	NOT ASSESSED NO DONOR AXIS IN RELEASED TASK	NOT RECORDED IN NORMALIZED FIELD	E0 EXECUTION ONLY	FAIL
`MRUN-MORPH-DEPMAP25Q3-VALIDATION-20260722`	G0-VALIDATION-PILOT	INCOMPLETE LEGACY RECEIPT MIGRATION REQUIRED	NOT ASSESSED NO DONOR AXIS IN RELEASED TASK	NOT RECORDED IN NORMALIZED FIELD	E0 EXECUTION ONLY	NO GO

These are migration receipts over historical runs. They do not rewrite the frozen result and do not invent overlap counts that were not recorded.

Competitor absorption matrix

Copy the operating principle, not the marketing claim.

Every source is paired with an absorb decision, a non-adoption boundary and a local response.

Download matrix →

COMP-ARCOFFICIAL SOURCE VERIFIED

Arc Virtual Cell Initiative

General virtual-cell data, models and evaluation

Absorb: Model/data/benchmark separation, cell-eval discipline and lab validation framing.
Do not copy: Generic cell-count scale or predicted atlas outputs as local DMD truth.
Local response: Use external models as replaceable adapters; protect disease outcome and outcome history.

COMP-CZIOFFICIAL SOURCE VERIFIED

CZI Virtual Cells + CELLxGENE

Open data, model and benchmark product ecosystem

Absorb: Entity discovery, task-based benchmarks and navigation across data and models.
Do not copy: A broad model marketplace before disease tasks and adapters are stable.
Local response: Keep typed Disease/Gene/Cell/Dataset/Model/Study search and frozen task objects.

COMP-TAHOEOFFICIAL MODEL HUB VERIFIED

Tahoe Therapeutics

Large chemical perturbation corpus and cellular foundation models

Absorb: Dataset-to-model card, public weights and explicit licence surface.
Do not copy: Cancer-cell performance as muscle or DMD transfer.
Local response: Run domain-shift probes; do not compete on generic corpus size.

COMP-RELATIONOFFICIAL ANNOUNCEMENT NOT INDEPENDENTLY VALIDATED

Relation MORGAN

Human multi-omic perturbation data plus experimental feedback

Absorb: Disease-relevant perturbation outcome and lab-in-the-loop operating model.
Do not copy: Commercial claims or model capability without independently auditable task evidence.
Local response: Build a smaller DMD donor × perturbation × time × function outcome matrix.

COMP-OPEN-TARGETSOFFICIAL API VERIFIED

Open Targets

Target–disease evidence product and release discipline

Absorb: Source-by-source drilldown, API, release identity and downloadable evidence.
Do not copy: Opaque association scores as causal or therapeutic predictions.
Local response: Keep source, context, direction, inferential unit and claim ceiling visible.

COMP-TREAT-NMDOFFICIAL NETWORK SOURCE VERIFIED

TREAT-NMD

Neuromuscular registry network and harmonized core datasets

Absorb: Disease terminology, longitudinal fields, outcome-measure discipline and registry interoperability.
Do not copy: Patient registry operation, clinical authority or access rights not held by NMD-VCell.
Local response: Map research objects to public core-dataset concepts and pursue collaboration rather than duplication.

Disease-level moat

The scarce asset is intervention outcome evidence, not another Transformer.

DMD perturbation outcome comes before spatial expansion, generic model scale or patient-digital-twin language.

Open DMD outcome pilot draft →

01MISSING

Matched DMD candidate perturbation response panel

disease-relevant human myogenic cells · independent donors · candidate perturbation · dose · time · molecular response · functional endpoint · negative interventions

First proprietary disease-task truth and model eligibility.Cannot substituteHealthy myoblast, HepG2 or unperturbed DMD context.

02MISSING

Donor-resolved longitudinal intervention response

same donor baseline · multiple post-perturbation times · donor-level replication · complete outcome return

Real donor transfer and temporal evaluation.Cannot substituteRandom-cell splits or cells treated as biological replicates.

03MISSING

Transcriptome plus functional rescue paired truth

molecular response · prespecified disease-relevant function · toxicity · null outcomes

A bridge from state similarity to useful biological outcome.Cannot substituteTranscriptomic resemblance alone.

Execution program

Infrastructure now; disease outcome next.

30 DAYSIMPLEMENTED THIS CHANGESET

TaskContract 1.0, Leakage Audit 1.0, E0–E6 claim ladder, permanent baselines and source-gated model watch.

Machine-readable schemas, public route and tests pass; legacy overlap fields remain explicitly incomplete.

90 DAYSBATCH 01 EXECUTED PARTIAL ACCEPTANCE

Benchmark Release NMD-B1-COMPAT-HEPG2-R1 executed with five seeds, permanent-baseline results and source/adapter decisions for TxPert, STATE, Stack, Tahoe-x1 and SLIM.

Baseline and STATE receipts are frozen; Stack and Tahoe-x1 remain explicit hard blocks; no matched DMD task was available.

180 DAYSAWAITING EXPERIMENT AND AUTHORITY

Registered DMD truth-generation Study created before execution; Prediction objects remain reserved for a separate eligible validation track.

Donor, intervention, dose, time, endpoint, QC, abstention and decision rule are frozen before experiment.

12 MONTHSBLOCKED DISEASE TRUTH

First prospective DMD perturbation benchmark with returned positive, null, toxic and failed outcomes.

Previously unseen donors/experiments, prespecified endpoint, stable advantage over all permanent baselines and calibrated uncertainty.

Identity-changing milestone

A page or model name cannot turn this into a DMD predictor.

Upgrade only after an immutable preregistered model predicts previously unseen DMD biological donors or experiments, exceeds every permanent baseline on prespecified molecular and functional endpoints across independent experiments, and returns calibration, failures and abstentions.

External releases are monitored references. Source verification is not local execution; local execution is not DMD validation; retrospective performance is not prospective or clinical evidence.