Source-linked AI summary
DNative-Twin: Decision Graphs and Digital Twins for Reconstructable Agentic Decisions
Junjie Pang, Zhenzhen Xie, Haoke Han, Ying He, Jing Wang, Gang Liu
TL;DR
AI agent outputs do not by themselves reveal the evidence, tool state, rules, or authority path behind a committed decision. DNative-Twin records that mechanism as a typed trajectory graph and replays it under declared conditions; experiments show that replay-contract state and verification results recover unresolved divergences that graph structure alone misses.
Problem
A final agentic decision output does not reveal the observed state, execution path, authorization, or commitment dependencies needed to reconstruct the decision mechanism.
Method
DNative-Twin records a committed agentic decision as a graph-native typed trajectory, synchronizes relevant state, replays the mechanism in isolation, and compares it under controlled changes.
Results
Unresolved-divergence recall increased from 0 to 0.667 with replay-contract state and to 1.0 when injected verification results were available.
Takeaways & Limitations
Graph structure, replay context, and verification evidence play separate roles in reviewing agentic decision mechanisms.
Takeaways & Limitations
The evidence is bounded by schema and extraction choices, fixed replay configuration, provider-specific probes, and injected adjudication cases with no held-out critical instance.
Abstract
from arXiv · showhide
AI agents increasingly gather evidence, invoke tools, apply constraints, and produce decisions that people or software may commit to action. A final output alone cannot show which evidence, tool state, rule, authorization, or action path produced it. We present DNative-Twin, a graph-native digital twin that records a committed agentic decision as a typed trajectory and re-executes its decision mechanism under declared conditions. The graph links the state observed by the agent, the path it followed, and the authority behind the resulting action. The twin synchronizes this information, replays the mechanism in isolation, and compares it under controlled changes. We instantiate the framework in enterprise decision processes using three public process logs and controlled replay suites. The experiments identify a specific failure: graph structure localizes represented changes but cannot determine the consequence of an unobserved tool state. In a three-condition controlled experiment with 300 injected instances, unresolved-divergence recall increased from 0 to 0.667 when replay-contract state was added and to 1.0 when verification results were also available; the held-out set contained no critical-class instance. Across 500--5,000 BPI 2020 cases, median end-to-end time increased from 0.794 to 8.889 seconds on the reported platform. These results separate the roles of graph structure, replay context, and verification evidence in reviewing a decision mechanism.
I. INTRODUCTION
DNative-Twin addresses the need to reconstruct agentic decisions by linking observed state, execution, authorization, and commitment in a decision trajectory graph. It replays the recorded mechanism under controlled conditions so aligned trajectories can be compared and reviewed.
- Motivation: A final APPROVE label can conceal different evidence, tool states, policies, and authorization paths, making decision reconstructability necessary for explanation and safe revision.The paper defines decision reconstructability as recovering the path that produced the decision.
- Motivation: Existing operational and provenance records often disconnect what the agent observed, how it acted, and what authorized the resulting action.Reconstruction requires linking these dependencies over time to the observed outcome.
- Decision representation: The decision trajectory graph is a dynamic heterogeneous subgraph bounded to one decision path, supporting reconstructability checks, constrained-path queries, and aligned comparison.GraphDiff records differences between aligned trajectories, while synchronized state and replay conditions provide interpretive context.
- DNative-Twin: DNative-Twin copies relevant state into an isolated twin, replays the recorded mechanism under fixed conditions, changes selected conditions, and compares the resulting trajectory with baseline.Rule-based review assesses whether observed differences change the decision under declared constraints.
- Decision representation: The framework connects decision objects, versioned knowledge, policies, tools, actors, agents, and outcome signals when changes could alter what an authorized participant should know or do.The state model includes tool states, authority grants, bounded skills, and outcome signals as decision-relevant elements.
- Decision representation: Its temporal attributed heterogeneous multigraph uses typed nodes and edges with environment, timing, validity, version, provenance, and confidence metadata.The multigraph preserves repeated interactions and multiple relations among the same entities.
B. Decision Trajectory and Reconstructability
The section defines a decision trajectory as a constrained, typed path from recommendation to commitment and outcome, then specifies how reconstructability and cross-environment comparison are assessed. It distinguishes benign structural variation from governance-relevant divergence and classifies unresolved cases when evidence is insufficient.
- Cross-environment comparison: The same mechanism can be replayed against different synchronized states, while aligned terminal trajectories support GraphDiff comparison of nodes, edges, attributes, and paths.Environment labels keep perturbed data separate from real-world evidence and preserve divergences for review.
- Governance interpretation: Traceability, validity, and contextual applicability are distinct properties: a valid artifact need not apply to the current decision state.State-sensitive scope and preconditions are recorded explicitly, with unresolved natural-language conditions routed to governed review.
- Trajectory definition: A decision mechanism separates recommendation from commitment, linking the latter to an authorized actor and applicable authority path.Rejections are recorded as ReviewEvents, while human or authorized machine actors may create CommitmentEvents.
- Reconstructability: A reconstructable trajectory must preserve required antecedents, admissible relations, temporal and version constraints, and outcome coverage.The minimum configuration includes decision objects, evidence, knowledge or policy, agents or tools, authorized actors, and authority grants.
- Cross-environment comparison: The synchronization contract maps observed state into a twin while declaring freshness, transformations, failed updates, and any governed substitutions.Identifiers may be tokenized, sensitive evidence represented by references, and unavailable tools replaced by consequence-preserving stubs.
- Governance interpretation: Governance equivalence treats structural differences as benign when required evidence, policy, authority, invariants, commitment class, and outcome constraints remain equivalent.Changes affecting governance-relevant projections are material rather than benign.
- Governance interpretation: Adjudication returns unresolved when available evidence cannot support classification, otherwise prioritizing critical, material, and benign labels in that order.Mandatory authority, isolation, evidence, or review violations are critical; frequency does not override mandatory constraints.
III. DNATIVE-TWIN ARCHITECTURE
DNative-Twin organizes decision reconstruction into graph projection and synchronization, isolated replay and perturbation, and semantic adjudication with revision.
- DNative-Twin uses three stages: graph projection and synchronization, isolated replay and perturbation, and semantic adjudication and revision.
A. Graph Projection and Synchronization
Graph projection converts operational and agent records into a synchronized decision trajectory that joins otherwise disconnected evidence, actions, authority, and outcomes.
- A purchase request is copied into an isolated twin after its supplier alert and state effects are recorded.The twin first replays the recorded decision, then tests a declared change such as a budget-tool timeout.
- Figure 1 presents the workflow as synchronized graph projection, replay under baseline and perturbation conditions, and GraphDiff adjudication with revision feedback.
- Source connectors normalize operational and agent records into typed graph updates while preserving origin and time.This connects decisions to the information and authority on which they depended and to later outcomes.
- DNative-Twin joins common record types around one decision trajectory, making that trajectory the unit of replay and comparison.No single source record normally contains the full path.
B. Isolated Replay and Perturbation
Isolated replay re-executes a recorded mechanism against synchronized state without external commitment, then applies declared perturbations to attribute trajectory differences to specific dependencies.
- The twin reuses synchronized alerts, policies, and budget state while replacing real-affecting calls with read-only or simulated calls.The twin has no authority to commit an external action.
- Replay comparisons are valid only when real and twin runs share the same decision identity, mechanism version, cutoff, and controlled execution context.Different internal traces may remain governance-equivalent when they satisfy the same declared constraints.
- The replay contract records decision identity and cutoff, governed state, execution configuration, and environment, including unbound or substituted fields as possible divergence sources.
- Tool substitution supports recorded-response reuse, read-only execution, or result-preserving stubs, each answering a different replay question.Response reuse isolates downstream mechanism behavior, whereas execution also tests synchronized tool state.
- Perturbation replaces one declared baseline dependency, such as a successful budget check with a timeout, and records the resulting divergent trajectory.Each intervention remains tied to the original decision and replay state for attribution.
C. Semantic Adjudication and Revision
DNative-Twin adjudicates replay differences against declared constraints, links candidate revisions to divergences and verification, and gates release on invariant, scenario, equivalence, freshness, and unresolved-case criteria.
- C. Semantic Adjudication and Revision: Semantic adjudication determines whether a replay difference affects a declared constraint and whether the mechanism requires adjustment.A candidate revision can add a fallback for a missing escalation path and remains linked to the divergence and verification record.
- C. Semantic Adjudication and Revision: Release readiness tests whether candidate mechanism M′ remains within its declared constraints.
- C. Semantic Adjudication and Revision: Cinv = 1, Cscn = 1, rmat ≤ǫmat, req ≥τeq, ftwin ≥τfresh, Ncrit,unres = 0.
- C. Semantic Adjudication and Revision: Passing the criteria makes a mechanism eligible for authorization, but release also requires an independent GOVAPPROVES decision.Drift, policy or tool changes, incidents, or new critical scenarios can reopen the revision cycle when they invalidate a release condition.
- C. Semantic Adjudication and Revision: Algorithm 1 organizes synchronization, replay, perturbation, graph difference, and adjudication over a declared observation window.The loop compares aligned real, twin, and perturbed trajectories using source records, a mechanism, constraints, and declared scenario changes.
A. From Architecture to Executable Specification
The framework turns architectural claims into executable graph checks, reproducible artifacts, and evidence-labeled evaluation procedures for replay and revision.
- A. From Architecture to Executable Specification: DNative-Twin expresses declared decision constraints as graph properties linked to concrete checks and failures.These checks test whether implementations preserve the distinctions needed for reconstruction and replay.
- A. From Architecture to Executable Specification: Six graph operations verify dependencies, authority and review paths, temporal applicability, prohibited shortcuts, and differences between aligned trajectories.A property graph with a rule or query engine is sufficient to implement them.
- A. From Architecture to Executable Specification: A minimal reproducible artifact tests whether the specification can execute outside the deployed system.
- A. From Architecture to Executable Specification: Real and twin examples must share a decision object, replay lineage, mechanism version, and observation window before graph differences are computed.The checker reports pass/fail results with witness paths or missing-type diagnostics for I1–I7, while the difference tool reports node, edge, attribute, and path changes.
- A. From Architecture to Executable Specification: Enterprise processes supply persistent events, identifiable decision objects, approval paths, and observable outcomes for instantiating and testing the model.
- A. From Architecture to Executable Specification: The evaluation separates public-log graph construction, controlled replay and release suites, and external-model probes, labeling each field by evidence source.
- A. From Architecture to Executable Specification: The public-log layer uses OCEL 2.0 Procure-to-Pay, BPI Challenge 2020 PermitLog, and BPI Challenge 2019, while prompts, replay modes, authority grants, and applicability rules come from overlays or injections.
- A. From Architecture to Executable Specification: An independent rerun reproduced all 18 declared checks and recovered the reported traces, events, sample, graph counts, and byte-identical sampled case-ID files.
B. Illustrative Procurement Case Study
The procurement case connects trajectory graphs and synchronized replay to concrete dependency analysis, then evaluates graph construction, replay, perturbation, adjudication, revision, and runtime behavior.
- Illustrative case: The procurement example contrasts two APPROVE outputs whose differing evidence, tool state, policy, and authorization paths determine reconstructability.The baseline and perturbed trajectories connect the formal model to an end-to-end decision path.
- Illustrative case: The trajectory graph retains dependencies that call-level traces omit, enabling explanation of divergent governed paths.GraphDiff compares aligned trajectories, while synchronized state and replay conditions provide interpretation context.
- Graph construction: All 3,371 evaluable commitments passed recommendation–commitment separation and reconstructability reachability, while 1,598 purchase-order objects contained Create–Approve–Pay trajectories.Checks requiring absent fields were evaluated only after declared mapping supplied them.
- Limitations: Graph-only detection classified all 40 unresolved tool-timeout instances as benign, while replay-contract state and verification results recovered the unresolved cases.The evaluation therefore separates structural localization from consequence determination for unrepresented tool states.
- Controlled evaluation: Replay-contract state recovered 60 of 90 injected unresolved instances, while verification-result fields recovered all 90.The three-condition experiment measures how additional replay and verification information supports adjudication.
- Controlled evaluation: Under budget or supplier drift, executable replay exposed current state changes, whereas response injection retained historical observations and stubbing preserved consequences without production writes.The comparison establishes that tool replay mode changes the meaning of trajectory comparison.
- Revision and runtime: Revision improved critical-invariant pass rate from 0.7143 to 1.0 and reduced unresolved critical cases from six to zero; median runtime reached 8.889 seconds at 5,000 cases.M0 failed the evaluated thresholds, while M1 passed; the runtime series covered 500, 1,000, 2,500, and 5,000 cases.
VI. RELATED WORK
Related work spans agent safety, provenance, tracing, business process management, digital organizational twins, and knowledge graphs. DNative-Twin combines these perspectives around reconstructing and replaying governed decision trajectories.
- Agent safety: Agent-safety systems test action safety, whereas decision reconstruction asks which observed state and authorization condition produced an action and how its path changes under replay.DNative-Twin retains the governed trajectory and its cross-environment difference.
- Provenance and tracing: Provenance and distributed tracing record lineage and operational execution, but DNative-Twin extends the record through the authorized actor and authority grant.The resulting record supports synchronized replay and comparison rather than historical tracing alone.
- Process management: Process mining derives behavioral models from event data, while agentic BPM addresses agents that observe process state and act through organizational tools.These foundations motivate the enterprise setting for governed agentic decisions.
- Digital twins: Organizational digital-twin work connects process-oriented representations to operational data and action-oriented analysis; DNative-Twin requires synchronized state that affected the governed path.Its isolated execution compares the same decision mechanism across synchronized states.
- Knowledge graphs: Knowledge graphs provide typed, temporal representations, while DNative-Twin turns temporal relations into inspectable checks for dependencies, authority, and cross-environment differences.The framework links process models, guardrails, traces, and provenance through a committed decision executable under a replay contract.
VII. DISCUSSION AND LIMITATIONS
DNative-Twin is scoped to consequential processes with identifiable, recordable decision trajectories, and its evidence is bounded by what the representation and evaluation settings expose. The experiments show that replay context and verification evidence complement graph structure when relevant tool state lies outside the recorded trajectory.
- Scope: DNative-Twin applies when an agentic process has an identifiable decision object and a bounded trajectory whose dependencies, constraints, authority, and outcome can be recorded.Generation without a committed action falls outside this scope.
- Evaluation setting: Enterprise processes are used because persistent events, approval paths, and outcomes make the framework’s required conditions observable.Evaluation in another domain requires evidence, authority rules, constraints, and outcomes defined for that domain.
- Layered evidence: Graph structure can localize represented changes but cannot determine the consequence of an unobserved tool state.Replay contracts record tool state and expected behavior, while verification results record whether the relevant condition held.
- Evidence boundaries: The evidence is bounded by schema design, event-to-graph extraction, derived or overlay fields, fixed replay configuration, and recorded external-probe conditions.The reported measurements and adjudication evidence therefore describe specific samples and injected cases rather than unrestricted deployment behavior.
- Deployment limits: The release demonstration evaluates four of eight declared conditions and leaves freshness, critical-scenario coverage, iteration limits, and replay budget for deployment evaluation.Deployment also requires access control, retention limits, and privacy-preserving references because the graph records sensitive operational relationships.
- Reported results: In the three-condition injected experiment, unresolved recall increased from 0 to 0.667 with replay-contract state and to 1.0 with injected verification results.The held-out set contained no critical instance; across 5,000 BPI 2020 cases, median processing time was 8.889 seconds on the reported platform.