STAGE: Stateful Translation to Agentic Graph Execution with Policy-Scoped Context and Deterministic Control

• Long, branching, multi-stage policies force agents to track both semantic reasoning and procedural state.Source: §1 · p. 1
• Growing context increases risks of omitted steps, misrouting, and propagation of unvalidated judgments.Source: §1 · p. 1
• STAGE represents policies as explicit action and decision nodes connected by valid transitions.Source: §1 · p. 1
• Each node receives only the context, state, and tools required for its local responsibility, while code owns procedural control.Source: §1 · p. 1
• STAGE is evaluated against monolithic full-policy execution across public benchmarks and Smart Dispute, with larger gains on deeper, more conditional workflows.Source: §1 · pp. 1–2

2.1 Experimental Conditions

• The baseline gives one agent the complete policy and control over both semantic reasoning and procedural progression.Source: §2.1 · p. 2
• Graph with full policy uses the reviewed graph for execution while presenting the complete policy at every node.Source: §2.1 · p. 2
• Node-scoped STAGE uses the same graph but gives each node only task-relevant policy context grounded in mapped source passages.Source: §2.1 · p. 2
• Comparing the baseline with STAGE measures bounded node execution, coordinator-owned control, and localized policy context together.Source: §2.1 · p. 2
• Comparing the two graph conditions isolates policy localization while holding control structure fixed.Source: §2.1 · p. 2

2.2 Research Questions

• RQ1 asks whether graph-structured execution improves task success and repeated-run reliability over monolithic full-policy execution.Source: §2.2 · p. 2
• RQ2 tests node-scoped context against complete-policy presentation for success, reliability, and token consumption under fixed control.Source: §2.2 · p. 2
• RQ3 examines how performance advantage and error composition vary with execution-chain length.Source: §2.2 · p. 2

• An LLM-assisted procedure decomposes authoritative policy into bounded tasks, maps source passages, and authors candidate nodes and transitions.Source: §3.1 · p. 2
• The candidate graph is validated, reviewed, and frozen before evaluation.Source: §3.1 · p. 2
• Each node is an executable contract specifying a local judgment, permitted skills and tools, result validation, and failure behavior.Source: §3.2 · p. 2
• At runtime, the coordinator supplies the node contract, case information, shared memory, and selected policy view to the model.Source: §3.2 · p. 2
• The model returns a typed result with its judgment, supporting case-evidence references, and an outcome when required.Source: §3.2 · p. 2
• The coordinator validates results and enforces graph-declared routing, retries, escalation, termination, and attempt limits.Source: §3.3 · p. 2

• The evaluation covers SOP-Bench Referral Abuse, shallower Retail and deeper Telecom workflows from τ^2-bench, and the deepest Smart Dispute workflow.Source: §4.1 · p. 3
• Public workflows are evaluated with GPT-5.6 Luna, DeepSeek V4 Flash, Claude Haiku 4.5, and Sonnet 5.Source: §4.1 · p. 3
• Smart Dispute is evaluated only with Claude Haiku 4.5 and Sonnet 5 because it contains proprietary banking knowledge.Source: §4.1 · p. 3
• One reviewed and frozen graph is constructed for each workflow before evaluation.Source: §4.2 · p. 3
• Within each comparison, the model, task input, environment, and evaluator remain the same.Source: §4.2 · p. 3
• The protocol compares full-policy and node-scoped STAGE conditions, isolates policy-context effects under fixed graphs, and groups Smart Dispute outcomes by chain length.Source: §4.2 · p. 3
• Pass1 measures single-run success, while Pass3 measures cases completed successfully in all three trials.Source: §4.2 · p. 3

• STAGE improves Referral Abuse repeated-run reliability most for weaker models: Luna’s Pass3 rises from 88.0% to 95.5%, and Haiku’s from 79.0% to 88.5%.Source: §5.1 · p. 3
• 32.5 percentage points: DeepSeek’s Retail Pass1 gain is the largest, while Pass3 gains range from 15.0 to 25.0 points across models.Source: §5.1 · p. 3
• Luna’s Telecom Pass1 increases from 47.5% to 90.0%, while Pass3 rises from 32.5% to 87.5%; gains remain visible for all four models.Source: §5.1 · p. 3
• Smart Dispute gains are large: Haiku’s Pass1 rises from 14.3% to 74.3% and Pass3 from 11.4% to 68.6%, while Sonnet’s Pass3 rises from 14.3% to 80.0%.Source: §5.1 · p. 3
• Under fixed graph control, node-scoped context raises Pass3 by 27.5 points for Haiku and 22.5 points for Sonnet, while reducing total token consumption by 37.8% and 56.8%.Source: §5.2 · p. 4
• Smart Dispute gains peak on medium chains at +73.3 points for Haiku and +66.7 for Sonnet; long-chain accuracy declines, with execution/end errors remaining dominant.Source: §5.3 · p. 4
• The current implementation is primarily a reliability and control mechanism rather than a token-reduction method because multiple node-scoped calls can increase output consumption.Source: §5.4 · p. 4
• Luna’s Telecom input consumption decreases by 17.3% while Pass3 increases by 55 points, and Sonnet’s Smart Dispute input consumption decreases by 36.9% alongside a Pass3 increase from 14.3% to 80.0%.Source: §5.4 · pp. 4–5

• Prior systems such as ShieldAgent and AgentLTL verify actions or traces against temporal rules, while other systems gate tool calls with generated, reviewed, or authored constraints.Source: §6 · p. 5
• These approaches share a control boundary in which a persistent agent selects the next task-level behavior and an enforcement layer judges whether it may proceed.Source: §6 · p. 6
• Workflow-oriented systems establish explicit workflow state and controller-mediated execution through reviewed SOP DAGs, executable paged programs, phase orchestration, or WCFGs.Source: §6 · p. 6
• STAGE centers a compliance-carrying execution node that binds policy passages, case evidence, capabilities, outcomes, schemas, and failure or transition behavior.Source: §6 · p. 6
• Within that reviewed interface, the model resolves a bounded semantic question while the coordinator enforces which judgment is active and how the procedure advances.Source: §6 · p. 6

• STAGE substantially improves task success and repeated-run reliability over monolithic full-policy execution across public benchmarks and Smart Dispute.Source: §7 · p. 6
• The Smart Dispute chain-length analysis shows that gains persist on long workflows and remove most errors from skipped or misordered procedural steps.Source: §7 · p. 6
• Remaining failures on the longest chains are concentrated in execution and terminal completion.Source: §7 · p. 6
• Node-scoped policy context improves success and repeated-run reliability while reducing total context processed relative to full-policy exposure.Source: §7 · p. 6
• A conforming trace bounds, reviews, and audits the execution path but does not guarantee a correct semantic judgment.Source: §7 · p. 6

• LLM-assisted construction segments the policy, extracts procedural tasks, and authors nodes, routing, and shared memory.Source: A Policy-to-graph Construction · p. 7
• Candidate fan-in, fan-out, and convergence are reorganized through policy-faithful absorption, serialization, or duplication before rechecking.Source: A Policy-to-graph Construction · p. 7
• The resulting plan is deterministically compiled and validated before human review and freezing.Source: A Policy-to-graph Construction · p. 7
• Generation stops for review when no faithful graph reorganization is possible, rather than emitting an invented rule.Source: A Policy-to-graph Construction · p. 7

SOP-Bench Referral Abuse

• SOP-Bench Referral Abuse contains six sequential stages for risk calculation, violation classification, severity assessment, and enforcement.Source: B Reconstructed Benchmark Graphs · p. 7

Telecom

• Telecom contains 13 nodes and two decisions, adding wider request routing and a deeper technical-support branch.Source: B Reconstructed Benchmark Graphs · p. 7

Smart Dispute

• Smart Dispute contains 26 nodes and nine decisions, producing the longest and most conditional execution structure.Source: B Reconstructed Benchmark Graphs · p. 7, Figure 5 · p. 8

Retail

• Retail provides the shallowest routing structure among the reconstructed benchmark graphs.Source: Figure 5 · p. 8

• The benchmark graph panels include only non-sensitive node titles and outcome labels.Source: Figure 5 · p. 8