Technical overview

Deterministic AI for regulated decisions, layer by layer

Architecture, request path, guardrails, audit trail, integrations, deployment and controls on one page. Written for CIOs, CISOs, architects and model-risk teams reviewing Palladium before a pilot.

Private deployment · request pathEnforced
  • PII redacted before the model and again on the response
  • Prompt injection screened on every inbound message
  • Outbound screened so the explanation matches the decision
  • No training on your data, by default and by contract
Definition

What is deterministic AI

Deterministic AI is a design where explicit rules make the decision and always return the same output for the same input. A language model is used only to understand the request and explain the result.

In Palladium, rules decide

Eligibility, amounts, limits and escalation are decided by versioned rules in code, written from your policy. The model never decides.

Every step is replayable

Each step is sealed in a record that can be replayed later, so you can show exactly which rule produced which outcome.

Rules in code vs. generative decisioning

Both use a language model. Only one lets it decide.

The difference is what the model is allowed to do.

QuestionRules in code (Palladium)Generative decisioning
Who decidesRules in code, written from your policyThe model, guided by a prompt and retrieved context
Same input, same outputAlwaysNot guaranteed. Outputs can vary run to run.
How you prove a decisionReplay the rule, inputs and outcome step by stepRead the model's rationale and its cited sources
Changing a policyEdit and version the rule. The change is explicit.Change prompts or retrain, then re-test behavior
Changing the modelAffects the wording of explanations onlyCan change decisions and needs re-validation
Model validation scopeNarrow: classify and explainBroad: the model is the decision-maker
Architecture

Every request takes the same path

Screening, the Rulebook, the Concierge and the Ledger, always in that order. The model sits inside the Concierge only.

Platform overview

In
ChannelVoice, chat, SMS, WhatsApp or email. Identity verified.
Channel
01
ScreeningPII masked and prompt injection blocked before any model sees the request.
Screen
02
The RulebookEligibility, amounts, limits and escalation, enforced in code.
Decides
03
The ConciergeThe model explains the decision. Output screened before it is sent.
Explains
04
The LedgerEvery step timestamped and sealed. Nothing written ever changes.
Seals
Request lifecycle

One Reg E dispute, second by second

Record #TR-8841: a customer disputes an $84.50 charge in chat.

TimeWhat happensLayer
9:41:02Request received, identity verified, PII masked. The model classifies it as an unauthorized-transaction dispute.Screening, Concierge
9:41:04Rule reg-e-provisional-credit checks eligibility: within the 10-business-day window, provisional credit eligible.Rulebook
9:41:09The model writes a plain-language response. Output screening checks it matches the decision and leaks no PII.Concierge
9:41:12Provisional credit of $84.50 issued through the payment rail.Specialist agent
9:41:40The full trace is sealed as one record, replayable on request.Ledger

Replay this record step by stepReg E disputes

The Rulebook

Policy as code, reviewed like code

Each policy is a versioned rule with explicit conditions, actions, limits and an escalation path. Natural-language policy enforcement is layered on top, never instead.

  • Hard limits. Dollar and risk thresholds sit outside the model's control.
  • Explicit escalation. Anything outside the rules routes to a person with full context.
  • Proactive triggers. A rule can act before the customer asks, such as a fee waiver when an NSF fee posts.
  • Versioned. Every decision records the rule version that produced it.
# policy/fee-waiver-eligibility.yml
condition: first_nsf_in_window(days: 365)
action: waive_fee(max_amount: 35.00)
requires_human: false
on_breach: escalate(context: full)

Enforced in code. The model never sees this decision.

LLM guardrails

Most guardrails filter what a model says. Ours limit what it may do.

Palladium screens the model's words, and it also limits what the model is allowed to do in the first place.

The model can

  • Classify the customer's request and extract the facts the rules need
  • Explain the Rulebook's decision in plain language
  • Hold a natural conversation across channels
  • Hand off to a person with a summary of the case

The model cannot

  • Approve, deny or change an amount
  • See the Rulebook or override a threshold
  • Send a response that contradicts the decision, because output screening blocks it
  • Receive unmasked PII, or train on your customer data
AI audit trail

A record an examiner can keep

The Ledger is Palladium's AI audit trail. Records are queryable by user, date and event type, and export as a digitally signed PDF with the full trace.

Record
Record ID, request type and channel
Timestamps
Every step, from receipt to seal
Rule
Which rule fired, its version and the inputs it evaluated
Decision
The outcome, and who decided it: the Rulebook
Explanation
What the Concierge said, and its screening result
Handoff
Escalations and the context passed to the person
Seal
Integrity hash. Nothing in the record can change.
Ledger · audit replayReplayable
9:41:02 Received9:41:04 Rule9:41:09 Response9:41:12 Credit9:41:40 Sealed
Record
#TR-8841
Request type
Reg E dispute
Rule
reg-e-provisional-credit
Outcome
Provisional credit
Decided by
Rulebook
Explained by
Concierge
Seal
sha256 · 9f3a2c71e0b44d8a6c1f…e27b
Sealed
Integrations

Channels in, systems of record out

Connectors are pre-built and bi-directional, with granular tool-calling permissions on every action.

Channels

  • Voice
  • Chat and in-app
  • SMS and WhatsApp
  • Email

Systems of record

  • Core banking
  • CRM, including Salesforce
  • Ticketing: Zendesk, ServiceNow
  • Payment rails
  • Identity providers
  • Document systems

Resilience

  • 99.9%+ uptime target
  • Circuit breakers on every external system
  • One shared memory across channels, agents and systems
Deployment

Your VPC or ours, one architecture

Three ways to run Palladium. The rules, agents and Ledger are the same in each.

OptionManaged conciergeCo-managed rolloutPrivate deployment
Where it runsHosted by Sphere, with strict data isolationHosted by SphereYour VPC on AWS, Azure or Google Cloud
Who writes the rulesSphere, from your policiesYour teamYour team
Who operates itSphereSphereYour operating team
ModelsModel-agnosticModel-agnosticPrivate models, or bring your own

Palladium is model-agnostic and not locked to one cloud. Your customer data is never used to train models, by default and by contract. See pricing and engagement options.

Security controls

Built in layers, checked on every message

Conventional controls plus a security layer for the risks that only AI brings.

AI layer

Prompt-injection and output screening

Inbound messages are screened before the model. Outbound responses are screened before the customer sees them.

Data

PII redaction

Runs on both passes, so personal data does not reach the model.

Access

Access and encryption

Role-based access, granular tool-calling permissions and bank-grade encryption.

Residency

Residency and consent

Data residency and consent management set for where you operate.

Supervision

People on every exception

Escalation carries the full context: identity state, request progress and everything already verified.

Audit

SOC 2 Type II

Independently audited controls over security, availability and confidentiality. Palladium is also HIPAA compliant.

Model risk management

Built for examiners, not only for demos

Your model-risk and compliance teams can see which rule decided, what the model said, and what changed and when.

Open the due-diligence kit

  • Decisions come from versioned, testable rules, so examiners can review the exact version that applied.
  • The model's role is narrow and documented: classify and explain.
  • Every step is sealed and exportable as a digitally signed PDF.
  • Vendor documentation for third-party model review and interagency TPRM guidance.
Frameworks

What Palladium maps to

The controls Palladium enforces, mapped to the frameworks your reviewers work from.

DomainRepresentative frameworksWhat Palladium maps to it
US model riskSR 26-2 (April 17, 2026) · OCC Bulletin 2026-13 · FDIC FIL-15-2026Versioned, testable rules; a narrow, documented model role; vendor documentation for third-party model review.
US consumer bankingReg E · ECOA/Reg B · UDAAP · interagency TPRM guidanceRequirements held in the Rulebook as versioned rules, with every decision sealed in the Ledger.
Organizational securitySOC 2 Type IIIndependently audited controls over security, availability and confidentiality.
Data protection and privacyGLBA · GDPR · CCPA/CPRAConsent management, PII redaction before model access, data residency controls.
Banking and tech riskDORA · FFIEC · PRA/FCA · MAS TRM · OSFI · APRAModel risk management, operational resilience, third-party risk documentation.
AI-specific regulationEU AI ActRisk classification, explainability, human-oversight requirements for high-risk use.
Payments and data securityPCI DSSSegmentation and tokenization of payment data anywhere it touches the model layer.
Securities and marketsFINRA · MiFID IIRecordkeeping, communications surveillance, best-execution-adjacent audit trail.

This describes the controls Palladium enforces. Regulatory mapping is not a certification by, or endorsement from, any regulator. SR 26-2 and AI

FAQ

Technical questions

If Palladium uses a language model, how is it deterministic?

The decision is deterministic. The conversation is not. The Rulebook decides eligibility, amounts and escalation in code, so the same facts always produce the same outcome. The model only classifies the request and explains the result, and output screening checks the explanation matches the decision.

Which language models does Palladium use?

Palladium is model-agnostic. You can use private model options, or bring your own model in a private deployment. Because the model never decides, changing models changes the wording of explanations, not the decisions.

What is in Palladium's AI audit trail?

Every prompt, tool call, decision and reasoning step, timestamped and sealed in the Ledger. Records are queryable by user, date and event type, and export as a digitally signed PDF.

Does customer data leave our environment?

In a private deployment, Palladium runs in your VPC on AWS, Azure or Google Cloud with private model options. In every deployment, PII is redacted before the model and your data is never used to train models.

What happens when a connected system is down?

Circuit breakers sit on every external system, and the platform targets 99.9%+ uptime. When a decision cannot be completed, the request goes to a person with its full context instead of a guess.

How does this support SR 26-2 model risk reviews?

Decisions come from versioned, testable rules, and the model's role is limited to classifying and explaining. That keeps the scope of model validation narrow and documented, and the due-diligence kit gives your model-risk team vendor documentation to review.

Review it with an architect

Bring your security questionnaire. A senior architect walks your team through the data flow, the controls and a live sealed record.