Explain this case
Follow the decisions and evidence that support the reported outcome.
Behavior Intelligence, Causality & Replay
Inspect the path.
Explain the outcome.
Test the change.
Follow why a case took its recorded path, compare it with other executions and test a proposed change against retained evidence inside an explicit simulation boundary.
Connect what the published application permitted to what one execution actually did. Compare runs, reconstruct the historical record and test a proposed change inside an explicit replay boundary.
FROZIONExecution FabricGoverned from request to outcome
Published GraphAP-Review v7Evidence gates · review · revalidation · effect controls
Realized execution · EX-1042-A↓
- Context suppliedInvoice r5 · 2.0% variance · retained inputsPresent
- Policy required reviewAP policy v4 · exact proposalEnforced
- Human decision boundReviewer · assurance · proposal r5Approved
- State revalidatedAuthority · versions · current conditionsPassed
- External effect confirmedERP payment PAY-4821 · effect receiptRecorded
Application meaning across the whole runOutcome-first inspectionTyped causal bindingsCross-execution analysisExplicit replay modes
Behavior Intelligence across your adoption path
Govern existing agents
Inspect the decisions, attempts and reported results of actions governed through Frozion.
Grow tools into workflows
Compare the defined workflow with the recorded path and its exceptions.
Build native agents
Review native-agent participation alongside the rules, tools and evidence it used.
Build systems of engagement
Use application execution records and simulation to evaluate the next change.
The problem this solves
A trace needs the business decision behind it.
When an outcome is questioned, teams need to connect the result to the proposal, facts, rule and approval—and understand how a change would affect recorded work.
Compare realized paths
Locate the boundary where executions diverged and inspect its inputs.
Test before live action
Keep simulations separate from current authority and side effects.
01 / Begin with the business outcome
What happened?
What shaped the path?
A published graph defines permitted behavior. A realized execution records the path one case took.
Connect each runtime boundary to its graph node, version, context, decision, authority and evidence. Follow the business relationships across traces, APIs, model calls and human tasks.
Explore Visual OrchestrationPublished Graph AP-Review v7Tenant A · invoice settlementInspect-only · no external effects
Recorded outcome
Payment confirmed
Evidence coverage8 of 8 required
01 · InputsInvoice r5 + contextSources and feature snapshot
02 · JudgmentReview requiredDecision + AP policy v4
03 · Human reviewApproved r5Bound review package
04 · RevalidationBinding matchedAuthority, state and versions
05 · Operation / effectPayment confirmedOne attempt + effect receipt
What did the external system confirm?
Application-level relationship
Keep the identities connected.
An Evidence Correlation Spine joins the Published Graph and Execution Instance to context, proposal, decisions, approval, Business Operation Identity and effect. Trace IDs, request IDs and attempt IDs retain their distinct roles.
Follow judgments as they change.
See earlier proposals, added evidence, policy requirements and human conditions alongside the decisions that superseded them. An approval for r5 remains part of the history when r6 requires new review.
Causal bindings describe application-visible relationships. “This policy required this review” or “this receipt confirmed this operation” connects evidence to behavior. These records do not expose a model’s private chain-of-thought or establish that every input caused its output.
02 / Move from a case to a pattern
Same application.
Different realized paths.
Compare the boundary where executions diverged, then inspect the supporting record.
Segment by application version, policy, route, capability, reviewer, tenant or outcome where evidence and permissions allow. Keep missing evidence visible in the analysis.
| Execution | Recorded outcome | Relevant boundary | What the evidence supports | Inspect |
|---|---|---|---|---|
| EX-1042-AAP-Review v7 | Payment confirmed | External effect | One approved proposal, one attempt, a confirmed ERP receipt. | |
| EX-1042-BAP-Review v7 | Effect unresolved | Attempt timed out | No terminal receipt yet; reconciliation is required before a retry decision. | |
| EX-1042-CAP-Review v7 | Blocked before invocation | Proposal revalidation | Approval covers r5; the requested r6 proposal does not match. |
Compare three alternative outcomes for the same invoice.
Where does work wait?
Compare policy obligations, evidence gaps, review assignments and capability-sufficiency failures.
What needs reconciliation?
Distinguish invocation failure, unknown effects, confirmed effects and separate compensating operations.
What differs across versions?
Compare paths, routing, validation, abstention and review burden with the relevant context and scope.
03 / Declare the replay mode
Reconstruct what happened.
Test what could change.
Historical inspection, simulation and a new live action have different permissions and effect boundaries.
Choose the mode explicitly. Keep historical authority in the record, and re-establish current controls before any present-day effect.
Read the record.
Inspect preserved evidence and the realized path.
External effects: none.Rebuild the history.
Resolve the historical graph, state, context, decisions, waits and attempts.
Effects represented from evidence; never repeated by reconstruction.Test a declared change.
Evaluate historical inputs under an alternate policy, route, model or graph.
Effects suppressed, stubbed, sandboxed or simulated.Authorize current work.
Re-establish current identity, policy, state, proposal, approval and operation identity.
A live effect is possible only through the required safety gates.Historical case EX-1042-AInvoice r5 · $12,240 · CP-1042-AAll external effects stubbed
Fixed historical basis
AP-Review v7 · AP policy v4 · one bound approval · source age 2 minutes against a 5-minute limit.
Replay boundary
No production credentials, endpoints, queues or payment calls. Recorded model output is retained; no model is rerun.
Declared fidelity
Compare proposed policy rules against retained inputs. External actions are disabled.
Payment confirmed.
- Policy
- AP v4: eligible approval required.
- Recorded review
- One eligible reviewer approved the r5 proposal.
- Recorded effect
- ER-1042-A confirms ERP payment PAY-4821.
Counterfactual resultNot compared
Ready to compare.
Apply a declared policy change to the historical inputs.
- Changed rule
- Two distinct eligible approvers for invoices of $10,000 or more.
- Resulting path
- No comparison performed yet.
- Effect posture
- External effects are stubbed; no new payment.
Replay Comparison Record
A comparison will identify the historical baseline, declared change, fidelity and resulting path.
It creates no approval task, retrieves no current data and performs no real replay. In an application, the Replay Plan and safety controls determine which evidence and capabilities may participate.
Test a change across recorded cases.
Batch simulation can evaluate a declared workflow, policy, route or model change against a set of recorded cases. Compare changed decisions and paths with the historical baseline, preserving evidence availability and replay fidelity.
A simulation does not approve a case or authorize a live tool call. Controlled re-execution remains a separate mode with current authority, policy and effect checks.
04 / Make reconstruction limits visible
Know what the replay preserves—and
what differs.
Historical reconstruction should use the application semantics that governed the case.
A Historical Execution Capsule can retain snapshots, protected references, digests, resolved dependencies, state, decisions and receipts. Replay Fidelity identifies what is preserved, reconstructed, substituted, simulated or unavailable.
Graph and dependency versions
Resolve the recorded graph, policy, routing, prompts, mappings, indexes and feature definitions. Do not silently substitute current configuration.
Historical decision inputs
Use the Context Assembly Record, Model Input Manifest and consumed feature snapshot. A fresh query against today’s data is a different input.
Probabilistic model behavior
A provider or model version may be unavailable, and rerunning a model may produce different tokens. Recorded-output reconstruction and a new invocation have different fidelity.
Unavailable or protected evidence
See which evidence is present, pending, available only as a protected reference, or missing. A reference alone may not contain the information required for a replay.
Separate investigation from live effects.
The Replay Boundary isolates production credentials, live endpoints, queues and consequential tools. Side-effecting nodes can be denied, stubbed, substituted, sandboxed or simulated according to the declared plan.
Controlled re-execution requires a Replay Safety Gate for current authority, policy, resource state, exact proposal, approval and effect eligibility.
Keep the viewer inside the access boundary.
Operations, Engineering, Security, Risk, Audit and business owners can inspect the same execution through role- and purpose-appropriate views.
Selective snapshots, protected references, redaction and retention preserve useful evidence without granting universal access to the underlying data.
ReplayUnderstand or compare behavior.
RetryAnother attempt at the intended operation.
ReconciliationEstablish whether an uncertain effect occurred.
CompensationA new governed action addressing a committed effect.
Event redeliveryDeliver the event again under its processing contract.
Controlled re-executionPerform current work after current controls are established.
Build on existing evidence systemsKeep observability.
Keep observability.
Connect application meaning.
Logs, metrics, distributed traces, SIEM, data lineage and model telemetry remain useful. Frozion connects their records to the proposal, context, authority, decisions, approval, operation and effect. The Data Plane preserves or references evidence; Behavior Intelligence provides the investigation and comparison experience.
05 / Questions, answered
Inspect with evidence.
Replay with boundaries.
Make the difference between recorded history, a proposed change and a current action explicit.
No. Observability platforms can correlate rich telemetry and application attributes. Frozion keeps business identities and typed governance relationships in the application model, connecting technical records to the proposal, policy, human approval, operation and outcome.
Not necessarily. Model randomness, unavailable versions, changed providers and incomplete historical inputs can limit reproduction. Replay Fidelity must distinguish preserved outputs, resolved historical dependencies, new invocations, substitutions and unavailable evidence.
Historical reconstruction represents external effects from evidence. Its Replay Boundary excludes production effects. A current live action belongs to controlled re-execution, with current authority, policy, state, proposal, approval and operation controls established explicitly.
No. Historical approval explains the past decision and its exact proposal. It does not automatically authorize a present-day effect. Current identity, authority, policy, resource state, proposal and approval must be evaluated at the live action boundary.
No. It compares behavior under a declared alternate configuration using the inputs and fidelity available to that run. A changed policy can yield a different obligation without invalidating a decision made under the historical policy.
The evidence architecture preserves or references artifacts and their relationships. Behavior Intelligence uses that record to inspect a run, compare executions, reconstruct history and evaluate alternate configurations under explicit replay controls.
Start with one outcome you need to explain
Understand the path.
Improve the application.
Bring a workflow, its evidence and one proposed change.
We’ll map what can be inspected, compared and reconstructed.