Use the right mechanism
A query or calculation may already satisfy the task without another model call.
Model Routing & Deterministic-First Execution
Use what you know.
Resolve uncertainty.
Route with a reason.
Use known facts, exact queries and rules first. Call a model when interpretation adds value, then select among routes that satisfy the task’s quality, privacy, region and output requirements. Keep the reason, usage and outcome connected.
When work remains unresolved, identify the gap. More evidence, verification, a policy check or an authorized reviewer may be the right next capability. Keep every route within the application’s contracts, authority and limits.
FROZIONExecution FabricGoverned from request to outcome
Establish exact factsINV-1042 · 2.0% PO varianceGoverned query + Feature Store · versions retained
Does the result satisfy the task?Evidence · validation · uncertainty · consequence
Contract satisfiedAccept this resultNo general reasoning model required.
Interpretation unresolvedEligible specialist modelSame facts, scope and output contract.
Required evidence absentRetrieve the evidenceRecheck sufficiency before reasoning.
Bounded attempts · Eligible fallback · Explicit stop
Choose the capability the work needsDeterministic firstEligibility before optimizationEscalation by causeRouting evidence retained
Model and tool routing across your adoption path
Govern existing agents
Route work your existing agent sends through Frozion; its own model calls stay with that agent.
Grow tools into workflows
Choose eligible model or tool capabilities for unresolved workflow tasks.
Build native agents
Native agents use configured routing rules, business context and output contracts.
Build systems of engagement
Application model steps follow the same eligibility and data-handling rules.
The problem this solves
A model call should have a reason—and a business owner.
Routine checks can become inference work by default, and model availability can become an accidental routing policy. Make task requirements and effort limits explicit.
Respect the data boundary
A cheaper or faster model remains ineligible if it fails the workload requirements.
Explain the spend
Link recorded model usage and attempts to the agent, initiating identity and workflow.
01 / Start with the work, not a model name
One workflow.
Different kinds of problems.
An exact value, a known calculation and an ambiguous document need different capabilities.
A task contract defines the required inputs, typed output, validation, acceptable uncertainty and consequence. Routing chooses a capability able—and permitted—to satisfy that contract.
Retrieve current state.
Get an authoritative invoice value, supplier identity or case status under the active scope.
Compute a known result.
Apply a declared formula and preserve the input versions and units.
Apply explicit logic.
Evaluate thresholds, required fields and business conditions as defined.
Check the contract.
Verify types, invariants, schema requirements and consistency before use.
Reuse a calculated fact.
Consume a governed feature with its definition, time and freshness semantics.
Continue from known state.
Use durable state and validated transitions for waits, approval and continuation.
Let models interpret
what the exact stages cannot settle.
A specialist extractor, classifier or reasoning model can help with unstructured language and ambiguity. Its output still needs the task’s validation and sufficiency checks.
Deterministic-first reserves probabilistic reasoning for work that benefits from it. It does not make all model behavior deterministic.
Carry established facts
through every escalation.
Invoice amounts, PO variance, exposure, scope and policy limits remain typed inputs. A model should reason around those facts without recreating arithmetic from prose or silently replacing a source value.
Explore reusable decision factsTool Call Resolver/Router
Select the capability. Resolve the inputs. Govern the call.
Give the model a connected set of capabilities: internal tools, Frozion workflows and external MCP endpoints. A workflow can itself be a native MCP tool, from a protected resource call to a complete durable process.
Internal and external tools together.
The router presents available internal tools, workflows and connected external MCP tools for the LLM to choose from.
Populate inputs across model turns.
Resolve the selected tool’s parameters over multiple turns before invoking it. Choosing a tool is separate from having the inputs required to call it.
Apply policy around the external call.
Invoke external MCP tools with policy before and after execution. The model’s selection does not expand its authority.
Selection and parameter population remain subject to the task’s contracts, authority and policy.
Explore governed connections02 / Continue the invoice review
What is unresolved?
Route to the capability that helps.
INV-1042 has versioned records and calculated facts. Change the unresolved condition to see which capability the workflow needs next.
For an interpretation task, compare routing preferences within the eligible set. Then see what happens when private capacity disappears or the escalation budget is exhausted.
Compare capability profiles for this task. Relative cost and latency are example values.
Tenant A · INV-1042 · Purpose: invoice reviewModel processing constraint: private deployment only
Established facts stay attachedInvoice r5: $12,240 · PO r2: $12,000
Typed inputsFS-1042-A · PO variance 2.0% · exposure $148,500 · 0 prior matches. Sources and freshness are checked for this review.
Return the invoice’s PO variance.
Required output: a current, versioned percentage with source references.
Exact facts satisfy this bounded task.
The Feature Store already supplies the calculation from the matching invoice and purchase-order revisions.
Versioned Feature Store lookup
Consume the governed fact and validate its definition, input revisions and freshness.
Model eligibility before preference
Which interpretation routes may compete?
All eligible candidates must meet the same task-specific quality floor and output contract. Actual cost and latency depend on the deployment.
| Candidate route | Mandatory conditions | Preference profile | Route decision |
|---|---|---|---|
| Private Specialist v2Tenant private deployment · region A | Eligible · evaluation EP-note-S2 covers this task and output schema. | Lower cost · longer latency | Selected |
| Private Reasoner v5Tenant private deployment · region A | Eligible · evaluation EP-note-R5 covers this task and output schema. | Higher cost · shorter latency | Eligible alternative |
| Shared Cloud v7External shared endpoint | Ineligible · fails the private-deployment requirement. | Excluded before preference comparison | Not eligible |
ACCEPT
Use the exact result. No general reasoning model needed.
The result is 2.0% from FS-1042-A. Acceptance applies to this calculation task, not to invoice approval or payment.
Next boundaryContinue the review under AP policy v4 and its human-approval requirement.
Model attempts in this example0 / 2
The exact-result path consumes no model attempt. All later work remains subject to the declared escalation limits.
- Task & scope
- RR-1042-A · PO-variance contract v1 · Tenant A · invoice review · RP-1042-v2.
- Eligibility & objective
- The governed fact satisfies the task. No model candidate comparison is needed.
- Selected next step
- Feature Store lookup · FS-1042-A · original source revisions and calculation definition retained.
- Sufficiency & stopping
- ACCEPT the bounded calculation; retain the separate approval boundary.
- Invocation receipt
- An invocation receipt connects the selected capability to its validation outcome.
Confidence is a signal.
It does not grant permission.
A classifier probability, self-reported certainty, verifier result and Context Sufficiency outcome answer different questions. High confidence cannot fill an evidence gap, resolve a policy question or authorize a business effect.
Stop when the contract is satisfied.
Stop when the limits are reached.
Bound attempts, elapsed time, inference cost, tokens, retrieval and human escalation. When no eligible route can satisfy the contract, wait, open an exception, request appropriate judgment or abstain.
03 / Establish eligibility before choosing a winner
Hard requirements first.
Trade-offs within the allowed set.
Price, quality or speed cannot compensate for a route that is not permitted.
Build the Model Eligibility Set from the active task, typed contract, identity, tenant, purpose, data class, provider policy, deployment constraints, evaluation coverage and operational state.
Can this route participate?
- Task, input and output contract fit
- Identity, purpose and data permissions
- Classification, retention and residency
- Approved evaluation coverage and available capacity
Which eligible route fits best?
- Evaluated quality for the application task
- Latency, cost and resource intensity
- Capacity, availability and resilience
- Expected verification, retry and review burden
The application declares which dimensions are mandatory. For example, latency can be a preference in one task and a hard deadline in another.
Bind model choices
to relevant evaluation evidence.
An Evaluation Profile connects a capability to the task, input class, population, language, modality, version, deployment and consequence it has been evaluated to handle.
General reputation and benchmark scores do not establish suitability for a particular business decision.
- Capability sufficiency
- Check evidence, freshness, schema, validation, unresolved conflicts and required verification—not simply whether the endpoint returned successfully.
- Deployment as part of the route
- Keep provider, model version, adapter, runtime target, region, context limits and retention posture attached to the selection.
- Fallback under the same contract
- Re-evaluate each alternative against current eligibility. An outage does not authorize a new region, provider or data-retention posture.
Optimize the complete path. A low-cost call followed by retries, a larger model and manual correction may cost more than an exact query and a suitable specialist. Connect inference, retrieval, verification, waiting and human effort to the governed outcome.
Usage you can attribute. Effort you can bound.
Connect model usage to the agent, initiating user and workflow. Set applicable token, cost, time and attempt limits alongside the route’s contract, then inspect the receipt and attempts when a case consumes more work.
Eligibility comes before cost preferences. Historical aggregated spend is a decision input; it is not automatically an atomic account-wide spending cap.
04 / Make the routing policy inspectable
See the possible paths.
Explain the one that ran.
Routing belongs in the visible application model: task contracts, deterministic alternatives, eligibility filters, sufficiency checks, escalation, fallback, human review and stop conditions.
Engineering, security, operations and finance can inspect the same behavior before publication and connect it to execution evidence afterward.
Why this route was chosen.
Inspect the task's decision contract, candidate capabilities, eligible and rejected routes, selection criteria, evaluation evidence and reason for selecting or stopping a route.
Selection evidence · application contract · permitted alternativesWhat actually executed.
See the runtime target, capability or model version, provider, attempt, validation and finish state, with available usage, latency and cost evidence.
Invocation evidence · actual execution · resulting validationA selected model gets
a reasoning role, not broader authority.
A model may interpret a case without being permitted to invoke every downstream capability. Tool calls, action scope, human approval and effect-boundary checks still follow the initiating authority and application policy.
Explore Visual Policy BuilderKeep the model infrastructure
that serves the application.
Existing gateways, private model servers, provider services and evaluation pipelines can retain their roles. Frozion connects them to capability contracts, task eligibility, deterministic alternatives and the surrounding business process.
Explore Private & Hybrid AIRoute-to-outcome lineage
Trace the route to the business outcome.
Follow the routing decision and established facts through capability attempts, added evidence, accepted results, policy and human decisions, and the business operation's reported effect.
Compare quality, cost and escalation using the complete outcome—not just the model name.05 / Questions, answered
Choose deliberately.
Escalate specifically.
Model routing is part of the application’s execution design, alongside exact capabilities, evidence and human judgment.
No. Use a sufficiently exact capability when it satisfies the task, then use an eligible specialist or reasoning model where interpretation adds value. Calculations, current records and known constraints remain typed inputs when the route escalates.
Model capacity is only one possible limitation. Missing evidence calls for retrieval; conflicting authoritative facts need verification; unresolved policy needs evaluation; an authority gap needs an eligible reviewer; and an uncertain effect needs reconciliation. Route to the capability that addresses the actual unresolved condition.
No. Hard requirements establish the eligible set first. A Routing Objective Profile optimizes among the remaining routes. The application decides which dimensions are mandatory and which permit trade-offs; a weighted preference must not erase a mandatory control.
Evaluate fallback as another routing decision. The alternative must satisfy current scope, data, provider, residency, evaluation and operational requirements. If none qualifies, use a declared wait, eligible alternative capability, human route, exception or abstention path.
No. Confidence signals have different meanings and may need calibration for the actual task. Capability Sufficiency also considers context, freshness, schema, conflicts, validation, consequence and verification obligations. Confidence does not establish permission to act.
Existing infrastructure can continue to manage endpoints, provider adapters, traffic, credentials, quotas and inference. Frozion connects model choices to task contracts, deterministic options, eligibility, human escalation and application outcomes. Supported integrations and enforcement behavior must be established for the selected infrastructure.
Yes. An explicit abstention, wait or exception is valid when no eligible capability can satisfy the task within its limits. Preserve the unresolved condition and stopping reason instead of turning insufficient evidence into a consequential answer.
Yes. Define applicable attempt, token, elapsed-time and inference-cost limits in the workflow and route. Inspect recorded usage by agent, user and workflow. The configured process determines the next step when the contract cannot be satisfied within its limits.
Start with one task and its unresolved conditions
Choose the right capability.
Keep the whole path visible.
Bring your models, exact services, policies and review requirements.
We’ll map where each belongs—and where the application should stop.