Security & Assurance — governed, auditable autonomy

Caskent is designed for safety-sensitive production, recipe protection, and regulated operations with human checkpoints and immutable decision records.

Trust by design.

Controls, isolation, and evidence are part of the product architecture, not a separate reporting layer.

Bounded action

Autonomy never writes outside approved safety, quality, or process limits.

Grounded reasoning

Recommendations cite recipes, SOPs, safety procedures, and relevant run context.

Operator authority

Human review modes define when the system can suggest, assist, or act.

Evidence first

Every action preserves the state, rationale, approval, and resulting signal change.

Grounding and citations for plant decisions.

Autonomy uses retrieval over approved plant knowledge so a recommendation can be traced to procedures and safety limits.

  • RAG over recipes, SOPs, standards, and batch records.
  • Clear citations for operator review and QA records.
  • Fallback paths when evidence is incomplete or conflicting.
CITATION FLOW
SOPRECIPERAGACTION+ CITES
Autonomy modes

Human in the loop, then graduated autonomy.

  1. Step 1

    Shadow

    Caskent observes plant decisions, predicts alternatives, and builds evidence without control write-back.

    Observe
  2. Step 2

    Assist

    The system recommends setpoints with citations and requires operator approval before action.

    Approve
  3. Step 3

    Graduated autonomy

    Approved policies write inside signed envelopes with pause, rollback, and review controls.

    Bounded

Immutable assurance-grade audit log.

Each recommendation and write-back stores the evidence bundle needed by operations, QA, food safety, security, and excise stakeholders.

  • Model version, retrieved citations, operator, time, and plant state.
  • Before and after signals tied to the production batch.
  • Exportable records for reviews and customer audits.
AUDIT LOG
TIMEACTIONEVIDENCE

Data isolation and recipe protection.

Per-tenant scoping

Recipes, sensory notes, batch records, and production data are isolated by tenant and site permissions.

Recipe and IP protection

Access boundaries are designed around formulas, cut preferences, maturation plans, and brand-specific profiles.

On-prem option

Selected deployments can keep runtime and sensitive data inside the customer environment.

Least-privilege access

Roles separate operations, QC, IT/OT, food safety, and executive reporting needs.

Compliance and controls.

Status reflects current design and roadmap posture for design-partner discussions.

ControlWhat it coversStatus
SOC 2 Type ISecurity controls and operating designIn progress
SOC 2 Type IIControl operating effectiveness over timePlanned
SSO / RBACIdentity, roles, approval modes, separation of dutiesPlanned for enterprise
Encryption in transit and at restCustomer data, telemetry, records, and backupsDesigned baseline
Methanol and food-safety limitsSafety guardrails and quality release checksWorkflow specific
Excise traceabilityBatch, lot, and production evidenceWorkflow specific
On-premLocal runtime and data boundariesAvailable by design

Safety guardrails.

Hard limits

Methanol, temperature, pressure, reflux, ABV, and line constraints can be enforced as non-negotiable bounds.

Fallback states

Ambiguous evidence, sensor drift, or missing data route the system back to operator review.

Change control

Policy updates require environment, recipe, and approval context before deployment.

Traceable releases

Every model, prompt, retrieval source, and control policy is versioned.

Continuous evaluation before release.

Models and policies are tested against golden datasets, judged for citation quality, and gated in CI before reaching the plant.

  • Golden datasets cover cut calls, congener interpretations, and safety cases.
  • LLM-as-judge checks completeness, citation support, and refusal behavior.
  • CI gating blocks policy release when regression thresholds fail.
EVAL GATES
DATAJUDGECIEDGE

Operational evidence by stakeholder.

TeamEvidence they reviewFrequency
Master distillerRecipe fit, sensory context, cut changesPer run
QualityCongener, methanol, release gatesPer batch
IT / OTAccess, integrations, write-back logsContinuous
SecurityTenant boundaries, audit exports, SSOReview cycle
FAQ

Security questions.

Can Caskent run without sending sensitive recipes to the cloud?

Deployment architecture is selected with each customer. On-prem and strict tenant boundaries are supported for sensitive workflows.

Who approves autonomous actions?

The customer defines approval roles and modes. Shadow and assist require review, while graduated autonomy acts only inside signed envelopes.

What happens when sensors disagree?

The system can refuse action, request operator review, or fall back to a safe state when evidence is incomplete or conflicting.

Is SOC 2 complete?

SOC 2 Type I is in progress and Type II is planned. Current design-partner conversations include the control roadmap and evidence plan.

Review governed autonomy for your site.

Bring security, QA, IT/OT, and operations into an early design-partner review.

We reply within two business days.