Validation and security

A prediction is only useful if you know how wrong it can be

Forgite reports error and uncertainty next to every number, keeps a full record of how each result was made, and protects chemistry your competitors would like to see.

Abstract

Uncertainty
On every prediction
Provenance
End to end
Data
Isolated per tenant
Audit
Planned

01 · Principles

Four rules we hold ourselves to

Rule 1

Show the error bar

No ranking without uncertainty. A confident wrong answer is the most expensive kind.

Rule 2

Hold data out

Accuracy is measured on structures the model never saw, chosen before training.

Rule 3

Write it down

Every result records its inputs, model version, settings and seed.

Rule 4

People decide

Models propose; named people approve anything that costs lab time.

02 · Method

How we validate a potential

  1. 1

    Split before training

    Hold out structures by composition family, not at random, so the test reflects new chemistry.

    Why Random splits flatter models
  2. 2

    Compare with DFT

    Report energy, force and stress errors on the held-out set.

    Output MAE, RMSE, parity plot
  3. 3

    Check calibration

    Confirm predicted uncertainty matches observed error.

    Output Calibration curve
  4. 4

    Compare with experiment

    Where measurements exist, compare predicted properties with them.

    Output Property report

03 · Context

Standards and practices we align with

Forgite is not certified against any of these yet. This table describes intent, not status.
AreaReferenceHow Forgite relates
DataFAIR data principlesResults stored with identifiers, metadata and provenance for reuse
StructuresCIF, POSCAR, OPTIMADERead and write standard formats; OPTIMADE queries Planned
Information securityISO/IEC 27001Controls designed with it in mind; certification Planned
Service controlsSOC 2 Type IIAudit Planned
Export controlApplicable export regulationsCustomer responsibility for controlled chemistry; on-premises option available

04 · Audit trail

From prediction to measurement, linked

provenance record · example
result: cand-0118/sigma
value: 12.4 mS/cm  # ± 2.1, 95% interval
from:
  structure: cand-0118.cif@sha256:9f2c…
  potential: equiva/battery-ft@0.3.1
  md: { T_K: [500, 600, 700], seeds: 3, time_ps: 1000 }
  campaign: sse-li-argyrodite-r1
approved_by: program-lead@customer
measured: pending

05 · Security

Protecting unpublished chemistry

Tenant isolationSeparate storage, models and logs per customer.
EncryptionIn transit and at rest; customer-managed keys in VPC and on-premises.
AccessSingle sign-on, roles, and least privilege for our staff.
Audit logEvery query, run, approval and export recorded.
Deployment choiceCloud, your VPC or on-premises.
DisclosureReport issues to security@forgite.com.

06 · Questions

Questions R&D teams ask

Is Forgite certified?

No. We have not completed any certification or audit. SOC 2 Type II and ISO/IEC 27001 are planned before general availability.

Can predictions be used as evidence in a filing?

Forgite supports the evidence work by recording how each result was produced, but it does not certify materials or properties. Measurements from your lab remain the evidence.

Who at Forgite can see our data?

By default, nobody. Support access is granted per incident, time-limited and logged.

Next step

Ask us how we would validate your chemistry

Send the property you care about and the reference data you trust. We will describe the held-out test before we train anything.