Skip to main content
SCIENTIFIC VERIFICATION & DECISION OS · v14.0

Verify what science can support before a model result becomes a decision.

BioAtlas connects claims, scientific tests, model credibility, failures, uncertainty, Bayesian beliefs, change impact, qualification, portfolio risk and reproducible snapshots without inventing a universal readiness score.

Scientific tests3/3No evaluated failures
Claim reproductionverifiedAbsolute error 2.00
Evidence debthigh34 weighted obligations
AbstentionabstainMaterial evidence remains conflicting.
Qualificationblocked-pending-evidencedocumented → reproduced
COMPLETE PRODUCT CONTRACT

All 25 capabilities are represented as governed product primitives.

25/25 implemented
Verify

Can the claim, model or conclusion be trusted in this context?

  • Executable Claim EngineTurn scientific statements into structured, rerunnable claims with population/context, intervention, comparator, outcome, evidence and reproducibility state.
  • Scientific Unit TestsRun golden cases, invariants, calibration checks, failure cases and out-of-domain tests against exact scientific model versions.
  • Model Credibility LedgerTrack multidimensional credibility without collapsing evidence into a universal score.
  • Model Failure AtlasMake documented failure modes, severity, biological context and mitigations first-class model evidence.
  • Scientific Model ArenaCompare compatible models on a user-defined benchmark under standardized versions, inputs, compute and hidden-label evaluation.
  • Benchmark Contamination IntelligenceTrack sequence, scaffold, patient, publication, temporal and corpus-overlap leakage risks for benchmark claims.
  • Adversarial Scientific ReviewConvene role-specific reviewers that are instructed to find weaknesses, disagreements and missing evidence before synthesis.
  • Explicit Abstention EngineAllow analyses to conclude insufficient evidence, conflicting evidence, out-of-domain use, unavailable sources or non-reproducibility.
  • Causal Evidence GraphSeparate causal, mechanistic, predictive and associative edges and attach evidence to every relationship.
Decide

How should evidence change belief, action and the next experiment?

  • Bayesian Scientific Decision TwinRepresent programme beliefs explicitly and update them as evidence changes.
  • Evidence DebtExpose unsupported assumptions, unresolved contradictions, stale evidence, surrogate dependencies and out-of-domain model use.
  • Translational BridgeConnect genetics, mechanism, cell phenotype, animal evidence, PK/PD, biomarkers, patient populations and clinical outcomes in one lineage.
  • Portfolio IntelligenceSurface correlated scientific risk across programmes, shared assumptions, common liabilities and reused out-of-domain models.
  • Portfolio Counterfactual & Reversal EngineRecord what would need to become true to advance, stop, reopen or reverse a programme decision.
  • Scientific EconomicsRank experiments and analyses by expected information gain relative to time, cost, burden and probability of changing a decision.
Continuous Science

What changed, what is affected and what must be replayed?

  • Scientific Change-Impact EngineTrace a changed model, paper, dataset or experiment through dependent claims, analyses, decisions and programmes.
  • Automatic Scientific ReplayRe-run reproducible workflows against changed dependencies and diff scientific outputs rather than only software artifacts.
  • Negative Science / Failure MemoryPreserve failed experiments, non-replications, terminated programmes and negative observations so they remain searchable and decision-relevant.
  • Institutional Scientific MemoryAnswer why a past decision was made from the exact evidence, alternatives, dissent, model versions and reopening conditions available at that time.
Qualify & Govern

Is this scientific asset qualified for its declared context of use?

  • Scientific Qualification LifecycleMove scientific models from discovered to documented, reproduced, benchmarked, validated, context-qualified, monitored and retired using evidence gates.
  • Model Qualification DossierGenerate context-of-use dossiers covering version, provenance, validation, applicability, known failures, uncertainty, change history and review.
Scientific Network

How can evidence move safely across tools, organizations and reproducible snapshots?

  • Federated Scientific OSBring governed computation to protected enterprise data while returning controlled evidence and provenance rather than centralizing raw data.
  • Scientific Connector & MCP EcosystemExpose scientific tools and data systems through typed capability passports with version, license, compute, validation and context-of-use metadata.
  • BioAtlas Evidence Protocol (BEP)Provide an open interoperable schema for claims, evidence, analyses, models, artifacts, uncertainty, contradictions, decisions and reviewers.
  • Citable Scientific SnapshotsFreeze an immutable evidence/decision state with exact versions, hashes, reviewers and timestamp for reproducibility and citation.
VERIFY

Executable evidence instead of citation-only confidence.

Executable Claim Engine

Reproduce a numeric scientific claim

verified

Observed result is inside the declared reproducibility tolerance.

Scientific Unit Tests

Test scientific behavior, not only software uptime

passReference AUC caseexpected 100.0 · observed 102.0 · tolerance 5.00
passNon-negative output invariantexpected 0.00 · observed 0.00 · tolerance 0.00
passCalibration referenceexpected 0.80 · observed 0.76 · tolerance 0.08

Golden cases, invariants, calibration, robustness and out-of-domain tests are version-aware and remain task-specific.

Credibility Ledger

Multidimensional credibility without a fake score

Developer validationdocumented
Independent replicationemerging
Prospective validationunknown
Experimental validationdocumented
External benchmarkdocumented
Calibrationpartial
Training leakage risklow
Out-of-distribution evidencelimited
Biological plausibilitysupported
Reproducibilitysupported
Model driftunder-surveillance
Failure Atlas + Negative Science

Make known failure boundaries impossible to ignore

highOut-of-domain chemistryUnrepresented chemical series · Require domain check and an orthogonal method.
moderateSparse validation evidenceRare population · Flag uncertainty and request prospective evidence.
non-replicationNo effect in the second assay system.Reopen when: Independent rescue experiment succeeds.
MODEL ARENA & CONTAMINATION

Compare models under a frozen scientific contract.

3 Pareto-efficient models
ModelTask accuracyCalibrationFailure rateGPU hoursPareto
Model A · v10.860.820.058.00Yes
Model B · v30.890.710.085.00Yes
Model C · v20.840.880.0311.0Yes

BioAtlas does not generate a universal winner. Rankings are metric-specific; the Pareto frontier identifies models that are not dominated across the declared metrics. Labels remain hidden until scoring, exact versions and parameters are recorded, and contamination assessment is required.

DECIDE

Evidence → belief → uncertainty → next action.

Bayesian Decision Twin

Update a programme belief with new evidence

prior 45.0%posterior 72.4%

Belief change: +27.4%. The probability is meaningful only for an explicitly defined proposition and evidence model.

Value of Information

Choose the experiment most likely to reduce decision uncertainty efficiently

  1. CYP inhibition panel0.05572.0% chance of changing decision · 3 days
  2. Human hepatocyte clearance0.04755.0% chance of changing decision · 4 days
  3. Telemetry safety study0.01845.0% chance of changing decision · 21 days
Evidence Debt

Make unresolved scientific obligations visible

high

34 weighted debt. Evidence debt highlights unresolved scientific obligations; it is not a probability of programme failure.

Abstention Engine

A scientifically valid answer can be: we do not know yet

abstain

Material evidence remains conflicting. Decision uncertainty exceeds the declared acceptance boundary.

ADVERSARIAL REVIEW & CAUSALITY

Structure disagreement rather than hiding it inside one assistant answer.

Mechanistic biologist

Mechanism is plausible, but causal rescue evidence is still missing.

recorded
Statistician

The validation set is small and confidence intervals should remain visible.

recorded
Pharmacologist

No role-specific critique has been recorded yet.

pending
Toxicologist

No role-specific critique has been recorded yet.

pending
Medicinal chemist

No role-specific critique has been recorded yet.

pending
Translational scientist

No role-specific critique has been recorded yet.

pending
Regulatory reviewer

No role-specific critique has been recorded yet.

pending
Skeptical reviewer

A key conclusion depends on one model family and needs an orthogonal check.

recorded
causesinhibitsactivatesrescuesmediatesnecessary_forsufficient_forassociated_withpredicted_byexperimentally_observed
CONTINUOUS SCIENCE

Scientific CI/CD: detect → impact → replay → review.

4 affected objects
model-v2analysisanalysis-17claimclaim-41decisiondecision-9programmeprogramme-amr
highdecision: decision-9planned · bioatlas-scientific-replay/1.0
highprogramme: programme-amrplanned · bioatlas-scientific-replay/1.0
mediumclaim: claim-41planned · bioatlas-scientific-replay/1.0
normalanalysis: analysis-17planned · bioatlas-scientific-replay/1.0

A dependency change can trigger scientific replay, but BioAtlas does not silently rewrite a reviewed scientific decision. Material decision changes require explicit review and sign-off.

Institutional memoryWhy did we make this programme decision?

Current draft — human review required

Reopening conditionResolve the highest-weight evidence debt item.

3 recorded dissent/review notes remain attached.

QUALIFY & TRANSLATE

Context-of-use qualification and preclinical-to-clinical lineage.

DiscoveredDocumentedReproducedIndependently benchmarkedExperimentally validatedContext qualifiedProduction scientific useUnder surveillanceDeprecatedRetired
Target geneticssupported1 connected evidence objects
Mechanismsupported1 connected evidence objects
Cell phenotypesupported0 connected evidence objects
Animal evidenceunresolvedSpecies translation remains uncertain.
Pk pdsupported0 connected evidence objects
Biomarkeremerging0 connected evidence objects
Patient populationunresolved0 connected evidence objects
Clinical endpointunresolved0 connected evidence objects
Trial outcomeunresolved0 connected evidence objects
Qualification dossier preview
{
  "schema": "bioatlas-model-qualification-dossier/1.0",
  "intendedUse": "Scientific decision support inside the declared project context",
  "contextOfUse": "Demonstration context",
  "model": {
    "id": "model-a",
    "version": "v1"
  },
  "trainingDataProvenance": "unknown",
  "validationDatasets": [
    "project-benchmark-v1"
  ],
  "applicabilityDomain": [
    "declared project data regime"
  ],
  "knownFailures": [
    "Out-of-domain chemistry",
    "Sparse validation evidence"
  ],
  "uncertainty": {},
  "softwareEnvironment": {},
  "changeHistory": [],
  "reviews": [
    {
      "role": "Mechanistic biologist",
      "critique": "Mechanism is plausible, but causal rescue evidence is still missing.",
      "state": "recorded"
    },
    {
      "role": "Statistician",
      "critique": "The validation set is small and confidence intervals should remain visible.",
      "state": "recorded"
    },
    {
      "role": "Skeptical reviewer",
      "critique": "A key conclusion depends on one model family and needs an orthogonal check.",
      "state": "recorded"
    }
  ],
  "reproducibilityEvidence": [
    "golden-auc",
    "invariant-positive",
    "calibration-check"
  ],
  "deviations": [],
  "regulatoryBoundary": "This dossier supports model qualification and governance; it does not itself confer regulatory qualification or acceptance."
}
PORTFOLIO INTELLIGENCE

Find correlated scientific risk before it becomes correlated portfolio failure.

Shared assumptions
biomarker translates2 programmes · P1, P3
target biology valid2 programmes · P1, P2
Shared liabilities
CYP3A42 programmes · P2, P3
hERG2 programmes · P1, P2
Out-of-domain model reuse
Model X2 programmes · P1, P3
Counterfactual & reversal engine

Pre-commit the conditions that would change the decision

advanceexposureMargin >= 3condition metAdvance only with predefined exposure margin.
stopsafetySignal >= 2not metStop when material safety liability crosses the programme boundary.
Scientific Economics

Information gained per cost, time and burden

  1. CYP inhibition panel0.054scientific value 0.233 · cost 2 · 3 days
  2. Human hepatocyte clearance0.045scientific value 0.243 · cost 3 · 4 days
  3. Telemetry safety study0.014scientific value 0.347 · cost 12 · 21 days
SCIENTIFIC NETWORK

Federation, typed connectors, an open evidence protocol and citable snapshots.

BEP 1.0
Federated Scientific OS

Bring computation to protected data.

Raw enterprise data stays inside its declared boundary. BioAtlas can exchange approved aggregate results, validation summaries and provenance manifests.

rawDataLeavesBoundary = false
BioAtlas Evidence Protocol

BioAtlas Evidence Protocol

Interoperable objects: claim · evidence · dataset · analysis · model · artifact · uncertainty · contradiction · decision · review · snapshot.

Open BEP specification →
Citable Scientific Snapshot

BA-4b1207bf32f5

Immutable canonical state with project, decision, evidence, model/data versions, artifact hashes, reviewers and timestamp.

4b1207bf32f5c4b0
model+dataAlphaFold / AlphaFold DBStructure prediction and reference structuresdirect-or-managed
modelBoltzComplex structure, affinity and generative-design workflowsdirect-or-managed
computeSchrödingerDocking, FEP, molecular dynamics and physics-based modellingdirect-or-managed
computeOpenMMMolecular simulationdirect-or-managed
dataPubChemChemical structures, identifiers and bioactivity datadirect-or-managed
dataChEMBLBioactivity and drug-discovery evidencedirect-or-managed
dataRCSB PDBExperimentally determined macromolecular structuresdirect-or-managed
dataOpen TargetsTarget-disease evidence and tractabilitydirect-or-managed
dataClinicalTrials.govClinical-trial registry evidencedirect-or-managed
enterpriseBenchlingELN/LIMS and registry evidence ingestionbring-compute-to-data
enterpriseSnowflake / warehouseFederated enterprise evidence queriesbring-compute-to-data
enterpriseDatabricks / lakehouseFederated scientific data and model executionbring-compute-to-data
computeBayesPKPK/PD, PBPK and population pharmacometrics evidencedirect-or-managed
computeBayesPharma ToxicologyMechanistic and evidence-linked toxicology intelligencedirect-or-managed