EVAR: Evidence-Validated Hypothesis Admission for Budget-Aware Narrative Reasoning¶
Conference: EMNLP 2026
Paper: ACL Anthology
Area: LLM Reasoning
Keywords: Narrative Reasoning, Evidence Validation, Hypothesis Admission, Budget-Aware Inference, Large Language Models
TL;DR¶
EVAR addresses the issues of intermediate hypothesis contamination and uniform revision costs in non-interactive long-form narrative reasoning by compiling an immutable source-linked atomic evidence store, dynamically routing budgets via instance complexity, gating hypothesis admission through multi-faceted validation challenges, and triggering sufficiency-based early stopping.
Background & Motivation¶
Real-world complex decision-making frequently hinges on long-form narrative comprehension, such as legal case analysis, criminal investigation attribution, and longitudinal medical history diagnosis. In non-interactive long-form narrative reasoning, large language models are expected to operate within a closed-world setting—assembling dispersed factual cues across thousands of words of static context without the luxury of follow-up human queries or external web search. Despite dramatic advances in standard multi-hop benchmarks and localized question answering, language models degrade sharply when faced with extended narratives. This breakdown stems primarily from two intertwined failure modes: premature commitment, where models form early conjectures based on partial premises and subsequently rationalize those biases into coherent but unfaithful explanations; and uniform reasoning depth, where iterative self-refinement and multi-step prompting frameworks indiscriminately apply the same fixed computational depth across all instances, wasting massive inference budgets on easy cases while potentially corrupting initially correct outputs without reliable external feedback.
Analyzing the root cause of these failures reveals that conventional reasoning strategies—ranging from Chain-of-Thought (CoT) and Self-Refine to trajectory-based search algorithms such as Tree-of-Thoughts (ToT) and Graph-of-Thoughts (GoT)—fundamentally attempt to resolve ambiguity by expanding search spaces and multiplying reasoning paths. In long-form narrative environments, however, an unconstrained exploration space becomes fertile ground for hallucination and error propagation. Once an unsupported or speculative intermediate inference enters the active context, the model treats it on equal footing with factual narrative premises, amplifying logical drift across multi-hop inference steps. Existing evidence-aware refinement methods such as RARR rely heavily on external search engines and post-hoc editing, which are inapplicable to closed-world narrative comprehension where external knowledge cannot fill internal story gaps. Conversely, full graph-based explorations like GoT incur unsustainable token expenses while failing to prevent invalid assertions from spreading across states. Developing a rigorous verification firewall grounded strictly in narrative facts—one that prevents speculative hypotheses from polluting the core trajectory while adaptively allocating computation according to instance difficulty—remains an urgent and unsolved challenge.
The crucial perspective of this work is to reframe long-form narrative reasoning away from unconstrained generation and indiscriminate self-critique into a strict hypothesis admission control problem: models are free to propose candidate explanations for unresolved narrative gaps, but no candidate is admitted into the answer-supporting state until it is explicitly validated against verified textual evidence. Core idea: compile the narrative into an immutable, provenance-preserving atomic evidence store, dynamically estimate instance difficulty from unresolved gaps and local uncertainty tags to assign an adaptive refinement budget, and gate iterative updates through hypothesis-conditioned validation challenges that admit supported hypotheses, quarantine unverifiable ones, and discard contradictions.
Method¶
Overall Architecture¶
The overall workflow of EVAR enforces an evidence-first principle and strict one-way evidential flow across four sequential stages: immutable evidence store compilation, budget-aware routing, verifier-gated hypothesis admission, and sufficiency-based early stopping and terminal synthesis. Given an input narrative and reasoning goal, the system first atomizes the text into source-linked evidence units with local consistency annotations, establishing an unalterable factual bedrock. Next, a lightweight probe detects unresolved gaps to compute an instance complexity score, cleanly bifurcating instances into a single-pass fast synthesis route or an iterative refinement loop bounded by an explicit instance-specific budget. Within iterative refinement, the model generates hypotheses for unresolved gaps, constructs targeted three-dimensional validation challenges, and passes them to a verifier against the locked store, admitting only fully substantiated hypotheses into the answer-supporting state. Finally, once evidence sufficiency is satisfied or the budget cap is met, early stopping halts exploration and synthesizes the final verdict exclusively from pure, validated evidence.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Narrative Context & Reasoning Goal"] --> B["Immutable Evidence Store Compilation"]
B --> C["Budget-Aware Routing"]
C -->|K=0 Fast Route| E["Sufficiency-Based Early Stopping & Synthesis"]
C -->|K>0 Iterative Route| D["Challenge-Gated Hypothesis Admission"]
D --> E
E -->|Insufficient & Budget Remaining| D
E -->|Sufficient or Budget Exhausted| F["Final Verdict & Narrative Attribution"]
Key Designs¶
1. Immutable Evidence Store Compilation: Preventing Unverified Intermediate Drift -> Extracting Source-Linked Atomic Claims with Consistency Tags -> Establishing an Unalterable Factual Reference Base
When processing long and convoluted narratives, models frequently conflate intermediate conjectures with explicit textual evidence, injecting unverifiable claims into early reasoning steps. To construct an unbreachable firewall against factual drift, EVAR avoids operating directly on raw continuous narrative text. Instead, an atomic extraction operator \(M^\pi_{\text{ATOM}}(X)\) decomposes the narrative \(X\) into explicitly stated atomic propositions \(\{(c_j, \rho_j)\}_{j=1}^m\), where \(c_j\) denotes an atomic proposition and \(\rho_j \subseteq \{1, \dots, n\}\) records its exact sentence or span identifiers. Inferred motives, speculative causal bridges, and subjective assumptions are strictly barred at this extraction stage. The framework then binds source-observable metadata \(\eta_j = (\text{entity}_j, \text{time}_j, \text{polarity}_j)\) without merging claims into synthetic macro-events, forming provenance-preserving evidence units \(u_j = (c_j, \rho_j, \eta_j)\). Subsequently, a local consistency tagger \(M^\pi_{\text{TAG}}(u_j, U \setminus \{u_j\})\) evaluates pairwise tensions and assigns consistency metadata \(\kappa_j = (\text{status}_j, \text{sev}_j, \text{note}_j)\), where \(\text{status}_j \in \{\text{OK}, \text{Uncertain}, \text{Conflict}\}\) and \(\text{sev}_j \in \{0, 1, 2, 3\}\) indicates issue severity. The resulting immutable base evidence store:
remains strictly locked throughout all subsequent reasoning stages. Because downstream operators are prohibited from mutating \(\mathcal{B}\), future hypotheses cannot retroactively rewrite the narrative's underlying factual ground truth.
2. Budget-Aware Routing: Eliminating Uniform Computation Waste -> Probing Unresolved Gaps and Uncertainty Signals to Allocate Adaptive Caps -> Partitioning Instances Between Fast and Budgeted Iterative Paths
Narrative cases exhibit vast disparities in evidence dispersion, suspect count, and inferential subtlety. Subjecting straightforward cases to extensive multi-turn deliberation squanders inference budget and exposes correct priors to overthinking noise, whereas overly complex cases fail under shallow reasoning depths. Instead of allowing the model to freely iterate without computational boundaries, EVAR executes a lightweight initial gap probe \(Z_0 \leftarrow M^\pi_{\text{GAP}}(\mathcal{B}, G)\) to identify the set of underspecified premises \(Z_0\) that directly prevent a conclusive answer. Combining these unresolved gaps with the uncertainty and conflict annotations stored in \(\mathcal{B}\), EVAR evaluates an explicit complexity score \(\Gamma\):
This metric unifies the breadth of missing causal links, the quantity of uncertain facts, and their severity into an objective difficulty indicator. EVAR then converts \(\Gamma\) into an instance-specific refinement iteration cap \(K\) under a global maximum budget \(B_{\max}\):
Instances with \(K = 0\) bypass iterative refinement entirely and follow the FAST route, synthesizing answers directly from \(\mathcal{B}\). Instances with \(K > 0\) enter the ITER route, where \(K\) imposes an immutable ceiling on refinement loops. This gap-driven routing guarantees that computational effort scales dynamically with actual narrative intricacy while enforcing strict predictability on worst-case execution costs.
3. Challenge-Gated Hypothesis Admission: Intercepting Unsupported Hypotheses from State Updates -> Structuring Hypothesis-Conditioned Multi-Faceted Challenges and Three-Way Partitioning -> Preventing Hallucinated Speculation from Contaminating the Core Trajectory
In multi-turn iterative reasoning, the most insidious failure mode occurs when a plausible yet ungrounded speculation enters the working memory and subsequently reinforces itself in later turns as an established fact. In EVAR, at each iteration \(t\), the system identifies active blocking gaps \(Z_t \leftarrow M^\pi_{\text{GAP}}(S_t, G)\) and generates targeted candidate hypotheses \(H_{z,t} \leftarrow M^\pi_{\text{HYP}}(z, S_t)\). Critically, these candidates remain external possibilities and are barred from immediate state inclusion. For every candidate \(h \in H_t\), EVAR constructs three tailored validation challenges \(V_{h,t} = \{v_{h,t}^{\text{sup}}, v_{h,t}^{\text{ctr}}, v_{h,t}^{\text{req}}\}\), demanding direct source justification (support challenge), probing for conflicting facts (counterevidence challenge), and verifying mandatory prerequisite conditions (prerequisite challenge). A verification operator \(M^\pi_{\text{VER}}(h, V_{h,t}, \mathcal{B})\) then evaluates \(h\) against the locked store \(\mathcal{B}\) and assigns a discrete outcome label \(\ell(h) \in \{\text{Support}, \text{Unknown}, \text{Contradict}\}\) along with supporting units \(E_h \subseteq \mathcal{B}\). EVAR enforces strict tripartite admission logic:
Only hypotheses receiving the Support verdict enter the active answer-supporting state \(\mathcal{H}^+_{\text{HIST}}\); candidates labeled Unknown are isolated in a quarantine buffer \(\mathcal{H}^?\) (preventing unproven assumptions from propagating into downstream generation), while Contradict candidates are discarded into \(\mathcal{H}^-\). This gate ensures that only fully evidenced deductions ever influence the reasoning trajectory.
4. Sufficiency-Based Early Stopping & Synthesis: Pruning Redundant Loops and Ensuring Pure Evidence Answers -> Monitoring Evidence Sufficiency and Constructing a Decontaminated View -> Achieving Cost Efficiency and Hallucination-Free Conclusion Synthesis
Exhausting the entire allotted iteration budget mechanically increases inference latency and runs the risk of self-correction degradation. EVAR incorporates a dual termination criterion after updating the working state \(S_{t+1} \leftarrow (\mathcal{B}, V_{\text{HIST},t+1}, \mathcal{H}^+_{\text{HIST},t+1})\). Beyond verifying whether blocking gaps are fully cleared (\(Z_t = \emptyset\)), the system computes an evidence sufficiency score \(\sigma_{t+1} \leftarrow M^\pi_{\text{SUF}}(S_{t+1}, G)\) to determine whether the admitted evidence network is already adequate to resolve the target goal \(G\). If \(\sigma_{t+1} \ge \tau_{\text{SUF}}\) or no blocking gaps remain, iteration immediately terminates regardless of remaining iterations in budget \(K\). To prevent audit-level validation challenges, exploratory query text, or quarantined hypotheses from leaking into final generation, EVAR builds a clean, decontaminated terminal state:
The answer synthesizer \(M^\pi_{\text{ANS}}(S^{\text{ans}}_{t^\star}, G)\) operates exclusively over this decontaminated view to produce the textual response \(\hat{y}\) and, for narrative crime benchmarks, a normalized probability distribution \(p\) across suspects. This maintains a strict one-way evidential flow, pairing cost-efficient early exit with guaranteed groundedness.
A Worked Example¶
Consider a complex case from NarraCrime involving a 1,400-word story with principal suspect Charlie, associate Alice, and bystander Bob. 1. Evidence Store Compilation: The narrative is compiled into 22 atomic claims with source citations. Unit \(u_5\) ("Alice was recorded by bank security cameras at the time of the incident", source [span 8], status OK, severity 0) and unit \(u_{12}\) ("A promissory note bearing Charlie's signature went missing from the victim's desk", source [span 17], status OK, severity 0) are locked. 2. Budget-Aware Routing: The gap probe flags 2 blocking gaps (\(z_1\): Charlie's financial motive and weapon access; \(z_2\): Bob's potential accomplice role in poison procurement). Coupled with 1 temporal uncertainty tag, the complexity score is computed as \(\Gamma = 4.8\), mapping to budget cap \(K = 2\) and routing the case to ITER. 3. Refinement Iteration 1: - For gap \(z_1\), candidate hypothesis \(h_1\) ("Charlie committed the crime to destroy the promissory note and avoid imminent bankruptcy") is generated. The verification challenges match \(u_{12}\) and bank records, yielding \(\ell(h_1) = \text{Support}\); \(h_1\) is admitted to \(\mathcal{H}^+_1\). - A candidate hypothesis \(h_2\) ("Alice physically entered the study") is tested against counterevidence challenge \(v^{\text{ctr}}\), matching \(u_5\) (bank alibi), yielding \(\ell(h_2) = \text{Contradict}\); \(h_2\) is discarded to \(\mathcal{H}^-_1\). - For gap \(z_2\), candidate hypothesis \(h_3\) ("Bob procured the toxic compound") is tested. The locked store contains no record of Bob obtaining the chemical and no alibi. The verifier returns \(\ell(h_3) = \text{Unknown}\); \(h_3\) is shunted into quarantine \(\mathcal{H}^?_1\), blocking it from being treated as true. 4. Early Stopping & Synthesis: At the close of iteration 1, sufficiency assessment confirms that Charlie is unequivocally identified as the sole principal culprit with established motive, means, and opportunity, scoring \(\sigma_2 \ge \tau_{\text{SUF}}\). Refinement stops early at iteration 1. The challenge metadata and quarantined \(h_3\) are stripped out. Grounded purely in \(\mathcal{B} \cup \{(h_1, E_{h_1})\}\), the synthesizer assigns \(0.90\) guilt probability to Charlie, clears Alice, and avoids making ungrounded accusations against Bob.
Key Experimental Results¶
Main Results¶
The authors evaluated EVAR on the non-interactive long-form benchmark NarraCrime (split into Easy, Medium, and Complex subsets based on story length, evidence density, and suspect count) and three widely adopted public reasoning benchmarks: HotpotQA (HQA), StrategyQA (SQA), and Big-Bench Hard (BBH). All systems utilized DeepSeek-V3.2 as the primary backbone under identical decoding constraints. Evaluated metrics include Role-Aware Verdict Score (RVS), Intent Recall (IR), Action Schema Recall (ASR), Evidence Coverage (EC), answer exact match/F1, accuracy, and average LLM call count (\(T\)).
| Method | Split / Benchmark | NarraCrime RVS | NarraCrime IR | NarraCrime ASR | NarraCrime EC | HQA Ans / SF | SQA Acc | BBH Acc | Avg Calls T |
|---|---|---|---|---|---|---|---|---|---|
| Direct | Easy / Global | 65.6 ± 0.54 | 64.8 ± 0.38 | 60.7 ± 0.61 | 70.8 ± 0.47 | 62.1 / 52.6 | 82.7 ± 0.41 | 79.0 ± 0.46 | 1.0 |
| Direct | Complex | 40.1 ± 0.69 | 60.0 ± 0.63 | 58.2 ± 0.65 | 59.4 ± 0.67 | - | - | - | 1.0 |
| CoT | Easy / Global | 67.7 ± 0.43 | 77.9 ± 0.31 | 64.6 ± 0.49 | 76.2 ± 0.36 | 68.9 / 60.0 | 86.9 ± 0.33 | 86.6 ± 0.39 | 2.4 |
| CoT | Complex | 46.0 ± 0.62 | 60.3 ± 0.51 | 60.4 ± 0.63 | 61.6 ± 0.56 | - | - | - | 2.4 |
| Self-Refine | Complex | 54.8 ± 0.63 | 62.7 ± 0.57 | 60.6 ± 0.64 | 66.1 ± 0.61 | 67.9 / 59.7 | 87.6 ± 0.35 | 88.0 ± 0.37 | 6.2 |
| SC (k=5) | Complex | 62.4 ± 0.55 | 62.3 ± 0.60 | 61.9 ± 0.58 | 71.4 ± 0.57 | 72.7 / 61.6 | 90.8 ± 0.29 | 88.9 ± 0.32 | 12.1 |
| CRITIC | Complex | 65.6 ± 0.47 | 65.2 ± 0.52 | 63.8 ± 0.57 | 72.7 ± 0.50 | 74.6 / 68.0 | 91.1 ± 0.25 | 87.3 ± 0.33 | 29.2 |
| S2R-style | Complex | 69.0 ± 0.44 | 62.9 ± 0.59 | 68.0 ± 0.47 | 77.2 ± 0.51 | 72.9 / 68.9 | 92.4 ± 0.21 | 89.9 ± 0.27 | 18.5 |
| SELF-DISC. | Complex | 68.0 ± 0.40 | 65.1 ± 0.48 | 64.6 ± 0.53 | 76.3 ± 0.44 | 76.0 / 67.9 | 92.9 ± 0.19 | 90.7 ± 0.23 | 4.8 |
| GoT | Complex | 69.6 ± 0.30 | 70.1 ± 0.33 | 66.7 ± 0.39 | 77.6 ± 0.31 | 76.6 / 72.7 | 92.2 ± 0.18 | 90.0 ± 0.22 | 35.5 |
| EVAR (Ours) | Easy | 85.9 ± 0.15 | 82.7 ± 0.18 | 85.6 ± 0.16 | 93.5 ± 0.14 | - | - | - | - |
| EVAR (Ours) | Medium | 81.6 ± 0.20 | 80.8 ± 0.19 | 77.6 ± 0.21 | 90.6 ± 0.18 | - | - | - | - |
| EVAR (Ours) | Complex / Global | 78.6 ± 0.24 | 72.6 ± 0.25 | 72.2 ± 0.29 | 83.3 ± 0.23 | 78.2 / 74.1 | 94.1 ± 0.13 | 93.6 ± 0.17 | 15.8 |
Ablation and Faithfulness Diagnostics¶
To isolate the exact contribution of each architectural pillar, controlled ablations were conducted on NarraCrime-Complex and StrategyQA. In parallel, output faithfulness was audited at the atomic claim level by a GPT-5.5 evaluation judge, reporting Unsupported Claim Rate (UCR) and Contradiction Rate (CR).
| Variant / Baseline | NarraCrime RVS | NarraCrime EC | NarraCrime UCR (↓) | NarraCrime CR (↓) | StrategyQA Acc | Description / Diagnostic Impact |
|---|---|---|---|---|---|---|
| EVAR (Full) | 78.6 | 83.3 | 8.6 | 3.8 | 94.1 | Complete pipeline with locked store, admission gating, and routing |
| w/o Admission | 71.5 | 74.4 | 17.8 | - | 90.5 | Hypotheses admitted without verification; UCR more than doubles |
| w/o Evidence Store | 70.3 | 72.5 | 19.4 | - | 83.2 | Removes locked atomic basis; unanchored reasoning degrades |
| w/o Budget Routing | 62.9 | 64.7 | 21.7 | - | 77.5 | Removes complexity-guided budget assignment; sharpest performance drop |
| Direct Baseline | 40.1 | 59.4 | 31.6 | 14.2 | 82.7 | Single-pass generation suffering from over 31% unsupported content |
| CoT Baseline | 46.0 | 61.6 | 28.4 | 12.7 | 86.9 | Multi-step prompting vulnerable to premature bias rationalization |
| Self-Refine Baseline | 54.8 | 66.1 | 25.1 | 10.5 | 87.6 | Blind self-critique without external anchor; UCR remains high at 25.1% |
| GoT Baseline | 69.6 | 77.6 | 15.9 | 7.4 | 92.2 | Large graph structure requiring 35.5 calls yet sustaining 15.9% UCR |
Key Findings¶
- Hypothesis admission is the primary firewall against hallucinations: Disabling admission gating (w/o Admission) causes NarraCrime-Complex RVS to drop from 78.6 to 71.5 and UCR to surge from 8.6 to 17.8. This demonstrates that EVAR's superior quality is not an artifact of additional refinement computation, but directly stems from filtering ungrounded and conflicting intermediate claims.
- Diminishing marginal returns govern reasoning budget caps: Sweeping \(B_{\max}\) from 0 to 1 yields dramatic gains on NarraCrime-Complex (+7.8 RVS, +9.6 EC, and +12.3 StrategyQA accuracy). However, escalating \(B_{\max}\) from 3 to 4 produces marginal improvements of only +1.3, +0.6, and +0.5, respectively. This confirms that a modest, well-targeted refinement budget resolves the vast majority of critical gaps, making \(B_{\max} = 4\) the optimal cost-utility operating threshold.
- Sufficiency-based early stopping prevents overthinking: In the early stopping study, disabling early stopping (w/o EarlyStop) noticeably inflated normalized iteration and token call costs without providing accuracy gains, while fixed iteration scheduling (\(K = 2\)) proved inferior to EVAR's adaptive stopping triggered by \(\sigma \ge \tau_{\text{SUF}}\).
- Budget routing correlates strongly with structural difficulty: On HotpotQA, only 33% of cases with 2 supporting facts triggered iterative refinement, compared to 68% for cases with >2 facts. On StrategyQA, routing rates scaled monotonically across question-length terciles from 26% to 45% and 71%, proving that the complexity score reliably differentiates simple cases from compositional challenges.
- Remarkable robustness across diverse backbone architectures: Experiments across Qwen3-235B, GPT-5.5, Gemini-2.5, and Claude-4.5 showed consistent advantages, boosting RVS by 8.1–9.8 points and EC by 5.2–6.5 points over GoT while slashing LLM invocations by ~55%.
Highlights & Insights¶
- Reframing generative revision as admission control: While traditional iterative prompting blindly trusts the LLM to critique and rewrite its own thoughts, EVAR treats all generated statements as provisional hypotheses requiring explicit empirical validation against an unalterable evidence base before admission, fundamentally eliminating the confirmation bias trap.
- Provenance-preserving atomic evidence grounding: By disaggregating long narrative text into fine-grained atomic claims mapped to explicit sentence spans and tagging local inconsistencies, the framework constructs a verifiable, auditable ground truth anchor that prevents speculative drift.
- Generalizable budget-aware inference paradigm: The architectural philosophy—detecting premise gaps, computing instance complexity, routing between fast and budgeted iterative paths, and gating state updates—readily generalizes beyond narrative analysis to complex domain applications like legal contract compliance, medical diagnosis reasoning, and root-cause software debugging.
Limitations & Future Work¶
- Reliance on upstream atomic decomposition: EVAR presupposes that the atomic extraction operator accurately captures all critical narrative details; if subtle subtext, implicit metaphors, or emotional cues are dropped during initial claim compilation, the locked store will inherit irreversible omission blind spots.
- Evaluation reliance on commercial LLM judges: Output claim decomposition, semantic matching, and UCR/CR scoring are automated via a GPT-5.5 judge. While prompts and decoding settings are held constant across all baselines, intrinsic model preferences may introduce slight measurement variances.
- Constrained to closed-world, English narrative text: The current framework has been validated primarily on closed-world English narratives. Extending hypothesis-conditioned admission to multi-lingual contexts, open-domain web retrieval, and multimodal egocentric reasoning represents an important future frontier.
Related Work & Insights¶
- vs GoT (Graph-of-Thoughts): GoT expands and merges diverse thought branches over complex graphs, but in non-interactive long narratives this leads to massive call overhead (35.5 calls) and propagates unchecked false premises; EVAR enforces strict one-way evidential flow and tripartite gating, outperforming GoT by 9.0 RVS points on complex narratives while using only 15.8 calls.
- vs SABA (Proactive Cognitive Awareness): SABA proactively evaluates premise sufficiency prior to final synthesis; EVAR extends this intuition into a full test-time hypothesis admission framework with dynamic instance-level budget routing, three-dimensional verification challenges, and quarantine buffers.
- vs RARR (Researching and Revising): RARR relies on external web search APIs to fetch references and post-edit text; EVAR addresses a closed-world setting where external retrieval is unviable, anchoring all verification internally to an immutable narrative evidence base.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ Formulates narrative reasoning as hypothesis admission with tripartite gatekeeping and dynamic complexity budgeting.
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ Rigorously evaluated across NarraCrime and 3 public benchmarks, with ablations, faithfulness diagnostics, and 4 additional backbones.
- Writing Quality: ⭐⭐⭐⭐⭐ Clear problem formulation, modular architectural descriptions, rigorous mathematical notation, and comprehensive analysis.
- Value: ⭐⭐⭐⭐⭐ Delivers an actionable, cost-controlled, and hallucination-resistant engineering template for complex long-context reasoning.