Zentari

Main programme · Future objective

Five product rails. One evidence-aware foundation model.

Zentari’s main scientific goal is to build a reusable multimodal model from the governed records produced by Scaffold Intelligence, Nano, Quantum Energy, Quantum Materials and Cell.

The model is intended to connect designs, simulations, observations and decisions while retaining source evidence, uncertainty and the option to abstain. A cross-product foundation model has not yet been trained or scientifically admitted.

See what exists todayRead the admission standard
Current stateFive product-specific engines and governed evidence rails
Foundation modelFuture objective · not yet trained or admitted
Admission gateLeakage-safe transfer tests and prospective validation

What foundation means here

A reusable scientific representation—not a replacement for solvers or laboratories.

The programme is not a chatbot wrapper or a claim that one network can replace domain physics. It is a governed model system that must retrieve evidence, use native tools when fidelity matters and refuse work outside its admitted domain.

Target architecture

Five scientific environments. One governed learning system.

Each product remains useful on its own while contributing typed representations, specialist tools and evidence-bound evaluations to a shared model programme.

  1. 01Scaffold

    Geometry · flow · mechanics

  2. 02Nano

    Carriers · fields · dynamics

  3. 03Energy

    Workloads · forecasts · decisions

  4. 04Materials

    Candidates · assays · selection

  5. 05Cell

    State · environment · lifecycle

Shared evidence language

Identity · provenance · units · geometry · fields · uncertainty · applicability · authority

North-star systemZentari Foundation Model

A multimodal, tool-using scientific model system designed to learn across governed records, call specialist engines and keep every proposal attached to its evidence and limits.

  1. RepresentGeometry, fields, traces, assays and programmes
  2. RetrieveSource-bound evidence and applicable prior cases
  3. OrchestrateSurrogates, optimisers, solvers and compute
  4. Decide carefullyCalibrated proposals, uncertainty and abstention
Programme status · target architecture. No unified cross-domain model is trained or admitted today. The five products currently create the governed evidence, tools and evaluations required to build one.

Evidence available today

The learning substrate starts with records that can be challenged.

These figures are current product evidence—not foundation-model training or performance results.

OpenFOAM pressure-drop chart across coarse, medium and fine gyroid meshes
FIG 02 OpenFOAM mesh-refinement observation564,291 → 1,960,469 cells · ΔP 10.661 → 11.399 Pa Fine-grid GCI 19.945%; under-resolved, standalone and not a measurement.
Open evidence →
Cell-F1 seven-day simulated population fate with absolute-unit oxygen, glucose, lactate and pH traces
FIG 08 Cell-F1 deterministic axial referencecelltrace-7d87ae1e… · 174 snapshots · 0–168 h Reduced-order, literature-order structural priors; no biological or laboratory validation.
Open evidence →

Five governed inputs

Every product contributes a different observation space.

Shared representation must preserve each domain’s units, identity, uncertainty and authority rather than flattening them into unqualified tokens.

Scaffold Intelligence

Geometry, OpenFOAM flow and transport fields, FEniCSx mechanics, design identity and numerical uncertainty.

Current public CFD is under-resolved and no scaffold outcome is laboratory validated.

Zentari Nano

Carrier identity, analytic and stochastic dynamics, optimisation histories and future rig trajectories.

Current public evidence is analytic or internal in-silico; there are no hardware labels.

Quantum Energy

Workload constraints, forecast context, execution cost, decisions and bounded telemetry for deciding where and when to spend compute.

This is the resource-control rail, not scientific ground truth; current evidence is counterfactual replay.

Quantum Materials

Typed candidates, molecular or material descriptors, assays, selections and active-learning records.

The current public loop is emulated and contains no measured material result.

Zentari Cell

Lifecycle programme identity, population and environment traces, and future observation operators linking simulation to microscopy, omics and assays.

Cell-F1 is an uncalibrated reduced-order prior; the protein programme is context-only today.

Honest state split

An evidence spine exists. The foundation model does not—yet.

No current figure, trace or benchmark is foundation-model performance evidence.

Available now

Governed product rails

Deterministic engines, source-pinned records, evidence classes, native-solver handoffs, abstention controls and bounded telemetry.

Still required

Cross-domain model evidence

A permissioned training corpus, trained multimodal system, demonstrated transfer, calibrated out-of-distribution behaviour and prospective external validation.

Intended model behaviour

A six-stage scientific reasoning loop.

Tool use and abstention are part of the model contract, not fallback copy added after prediction.

  1. 01Retrieve

    Find source-linked records within the admitted scope.

  2. 02Represent

    Align modalities without erasing units or identity.

  3. 03Propose

    Generate candidates, experiments or compute plans.

  4. 04Quantify

    Report uncertainty, applicability and alternatives.

  5. 05Route

    Escalate decisive cases to native solvers or laboratories.

  6. 06Update

    Retrain only through reviewed, versioned evidence.

Admission programme

The model earns authority task by task.

Scale, fluency or one benchmark cannot grant scientific authority across the suite.

  1. Corpus integrity

    Verify rights, provenance, units, coordinate frames, duplicates, evidence class and withdrawal state before training.

  2. Leakage-safe evaluation

    Hold out grouped candidates, time periods, sites and complete domains so memorisation cannot masquerade as transfer.

  3. Strong baselines

    Compare with each product’s analytical, statistical, surrogate and solver-backed methods under equal evidence and compute budgets.

  4. Scientific reliability

    Measure calibration, domain shift, out-of-distribution abstention, invariants or conservation where applicable, and inference cost.

  5. Prospective challenge

    Predeclare tests on unseen solver cases and independently retained laboratory or operational observations.

  6. Governed release

    Retain immutable model and data receipts, evaluation reports, telemetry caps, rollback and no silent online learning.

Foundation status is earned by reusable transfer—not inherited from scale.

Help build the evidence the model cannot generate for itself.