DACRI: Decision-Aware Causal Intervention Ranking for Critical Supply Chains

DACRI: Decision-Aware Causal Intervention Ranking for Critical Supply Chains Shiqi Huang* , Jiani He* , Dingyan Shang* , Yihua Xu, Jize Li, Yan Lyu, Lashimi Muraleedharan Nair Independent Researcher {juliahuangsq01, dingyanshang, yxu442, lmuraleedharanair}@gmail.com, jianihe@alum.mit.edu, lyu.yan@northeastern.edu, jizel@bu.edu

arXiv:2608.11154v1 [cs.LG] 11 Aug 2026

Abstract—Detecting or attributing a supply-chain disruption is not the same as selecting the intervention that maximizes recoverable net value. We present CriticalSCM-Bench v1, a controlled synthetic benchmark with causal ground truth, paired factual/counterfactual rollouts, and an explicit net-value objective. Relative to a full-information train-selected static benchmark, LambdaMART improves median normalized net value by 5.7– 16.2%, with paired statistical support on the semiconductor and critical-material archetypes but not on digital infrastructure. On digital infrastructure, a domain-informed constant-buffer policy remains stronger, showing that greater model complexity is not uniformly justified. Across partial and delayed settings, LambdaMART retains 33–75% of full-clamp value. Stress tests further show that intervention fidelity, timing, cost, and held-out disruptions can alter policy ordering. Critical materials show the weakest out-of-distribution retention. Separately, a guarded explanation study over 540 generations preserves every fixed intervention decision after deterministic validation and template fallback, although exact wording remains unstable. Within this controlled setting, the results identify regimes in which adaptive ranking adds value and those in which simpler structural policies remain preferable. Index Terms—Supply chain resilience, causal discovery, anomaly detection, intervention ranking, learning-to-rank, decision support, critical infrastructure, benchmark, large language models, grounded explanation.

2. Decision-aware evaluation: a net-value objective and a baseline hierarchy spanning information-light heuristics, attribution-as-action, a full-information static benchmark, an oracle-like structural rule, learned ranking, and a perfect-value upper bound. 3. Assumption stress tests: partial/delayed interventions, four feasibility specifications, cost and detection timing, learning curves, held-out disruptions, and an oracle ceiling. A secondary guarded-explanation study evaluates whether a fixed DACRI decision can be rendered into operator-facing text without allowing the language model to own or alter the recommendation. II. R ELATED W ORK

Supply-chain and inventory models. Ripple-effect work traces disruption propagation [4], while multi-tier visibility remains incomplete [5], [6]. METRIC [7] optimizes buffers for known topology and demand/lead-time distributions; DACRI instead ranks an episode-specific action after an anomaly. Constant-buffer tests the added value of learning beyond this downstream-buffer insight. Related data-driven work I. I NTRODUCTION we are aware of addresses forecasting and planning under Small upstream signals can cascade into shortages before demand uncertainty—mixed-frequency production planning [8], downstream risk becomes visible [1]–[3]. Anomaly detection newsvendor ordering under shifting demand [9], and throughput answers “what is abnormal?” Root-cause analysis asks “where forecasting [10]—rather than selecting where to act once a did it begin?” The technical gap is not detection alone. It is disruption is already propagating. deciding where an operator should act now to recover the most Anomaly detection. Isolation Forest [11], long short-term value. memory (LSTM) autoencoders, USAD [12], and MTAD-GAT DACRI addresses this gap by framing intervention selection [13] capture multivariate patterns. Microsoft’s deployed service as a learning-to-rank problem. The objective combines counter- reports lead time as the customer-visible quality measure [14]. factual loss avoided with the cost of acting. Our contributions Alert triage and prioritization. Severity- or anomalyfollow. magnitude ranking is a common information-light default 1. CriticalSCM-Bench v1: a controlled benchmark with across operational domains, from security operations [15] to intervention costs and paired counterfactual labels for multilevel anomaly detection [16]. A recurring finding is that three critical-supply-chain archetypes. severity alone is insufficient and that prioritization should reflect downstream or business impact rather than deviation size [17]– * Shiqi Huang, Jiani He, and Dingyan Shang contributed equally to this [19]. Supply chains lack a corresponding benchmark with work. causal ground truth and explicit intervention cost; DACRI This is the authors’ accepted manuscript for presentation at the International Conference on Electrical, Computer, Communications and Mechatronics Engineering (ICECCME evaluates impact as counterfactual net value. 2026), 15–17 October 2026, Bali, Indonesia. © 2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current Temporal causal discovery. Granger causality [20] is or future media, including reprinting or republishing this material for advertising or a linear baseline. PCMCI/PCMCI+ [21] estimates lagged promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. relationships under autocorrelation. DYNOTEARS [22] frames

structure learning as continuous optimization. DACRI treats the graph as an intermediate consumed by the ranker. Treatment-effect estimation. Meta-learners estimate heterogeneous effects from observational data [23]. DACRI instead uses paired simulator counterfactuals to rank cost-adjusted feasible actions; it does not estimate a calibrated treatmenteffect surface. Guarded language-model explanations. Retrieval-grounded generation provides evidence-conditioned context [24], while unsupported claims remain a concern in hallucination detection and factuality evaluation [25], [26]. DACRI fixes the intervention upstream and measures decision preservation and packet-level grounding, not general faithfulness. III. P ROBLEM S ETTING A. Archetypes and KPI Schema Key performance indicators (KPIs) define the observable dimensions of each archetype. Three archetypes support full evaluation: digital infrastructure hardware lifecycle (DI), calibrated using drive-reliability and electronics lifecycle data [27], [28]; semiconductors (SE) (UCI SECOM [29]), and critical materials (CM) (USGS [30]). Table I summarizes the aligned KPI schema across the three supply-chain archetypes. Directionality is normalized so larger values indicate greater risk. Public datasets provide statistical calibration. Causal graphs and counterfactual labels come exclusively from the controlled benchmark (§V). B. Problem Formulation For archetype a, represent the supply chain as a temporal directed acyclic graph (DAG) G(a) = (V (a) , E (a) , Λ(a) ), where Λ(a) stores edge-lag assignments. Detector Dψ emits pertimestep scores. The first threshold crossing is the detection time. It is calibrated to a false-alarm rate (FAR) ≤ 1/month. For anomaly e = (k, t, sk,t ), where k is the flagged KPI, t its detection time, and sk,t its score, candidates are upstream ancestors of k in the candidate graph Gc . In v1 the detection target is not an arbitrary KPI: an episode enters the ranking study only when the flagged KPI is the archetype’s service-risk node (Table I), so k always denotes that node. Candidates are then

Yei,e = Bi,e ϕi,e − ci,e ,

Yi,e = max{0, Yei,e }.

(2)

where: (0) (i) • Bi,e = Le − Le is paired loss avoided; • ϕi,e = clip(1 − γ|∆i,e |/σi , ϕmin , 1), where ∆i,e = xi (t)−µi , σi is the v1 calibration-scale standard deviation, and the main setting is γ = 0.75 and ϕmin = 0.1; (0) • ci,e = ca Le /dG (i, k), where ca is the archetype cost (0) scale, Le is no-action loss, and dG is directed shortestpath length in benchmark G; unreachable nodes use the implemented surrogate dG = |V |. Yei,e is forced-action net value. The ranking label Yi,e truncates negative raw values to create a zero-valued no-action reference; clipping does not implement test-time abstention. Score-based abstention is evaluated separately in §VI-F. The LambdaMART score orders candidates; it is not a calibrated estimate of Yi,e . In the no-graph regime, causal strength and lag use fixed uninformative defaults 0.1 and 1 step for every candidate. The inverse-distance term is a stylized proxy for intervention type rather than coordination distance. Candidates closer to the service-risk node typically represent downstream buffering, emergency capacity, or allocation actions requiring sustained resource commitment, whereas upstream process corrections are modeled as lower-cost one-time actions. Accordingly, the benchmark assigns higher cost to candidates closer to the service-risk node. We do not interpret graph distance as an empirical monetary-cost model. The pre-specified scale is ca = 0.3 for DI/SE and ca = 0.1 for CM; sensitivity is reported in §VI. From detection time onward, do(stabilize i) hard-clamps KPI i to baseline mean µi . This is the primary benchmark label specification. Factual and counterfactual rollouts share a seed, so Yi,e is paired. Because this idealization is strong, we treat partial, delayed, and stochastic responses as assumption stress tests. IV. DACRI F RAMEWORK A. Anomaly Detection

We compare exponentially weighted moving average plus median absolute deviation (EWMA+MAD), Isolation Forest, a C(e) = {i ∈ V : i ⇝Gc k}. (1) long short-term memory autoencoder (LSTM-AE), UnSupervised Anomaly Detection (USAD), and multivariate time-series Here Gc is benchmark G in the headline regime, recovered Ĝ anomaly detection with graph attention networks (MTAD-GAT). in the graph-swap regime, and an explicit all-non-k construction Validation thresholds target FAR ≤ 1/month and prioritize lead without graph features in the no-graph regime. Stored reported time. runs have nonempty benchmark and recovered ancestor sets. For completeness, an empty set falls back to all non-k nodes. C(e) is therefore the service-risk node’s ancestor set: 11/9/8 B. Lagged Causal-Structure Estimation candidates under benchmark G for DI/SE/CM. Temporal We compare Granger, PCMCI/PCMCI+ (primary), and a precedence and lag residual are features, not filters. nonlinear check [20]–[22]. Intrinsic recovery and downstream Each candidate receives an ordinal score from a learned policy utility are reported separately. The reported analysis ranker Rθ . The ranker is trained to order candidates using net configuration pre-specifies αPCMCI = 0.01 before validation intervention value as relevance: and test evaluation.

TABLE I KPI SCHEMA FOR THE THREE CRITICAL SUPPLY- CHAIN ARCHETYPES . Dimension

Digital Infra.

Semiconductors

Critical Materials

Demand Supply Inventory Lifecycle Quality Service-risk Criticality

repair demand spare availability spare buffer recovery yield fault rate downtime risk asset priority

production demand component capacity component inventory rework yield production yield production delay product priority

refined-material demand material availability stockpile buffer recycler yield purity rate allocation shortfall end-use criticality

C. Decision-Aware Intervention Ranking

A. Benchmark Archetype Specification

Digital Infrastructure (DI). 12 nodes, two-stage topolThe ranker scores each candidate using: anomaly severity, causal-strength statistic and lag, temporal precedence, lag- ogy. Drive reliability is calibrated from Backblaze [27]; alignment residual, downstream criticality, lifecycle-recovery lifecycle and bill-of-material context comes from electronics feasibility, and inventory-buffer depletion. Relevance labels data [28]. Disruptions: recycler yield drop (15–40%), faultcome from paired simulations. The headline ranking experiment rate spike (1.5–3×), lead-time extension (+5–21d). Lags: For the uses benchmark G to isolate ranking; a separate graph-swap 7–14d (yield→buffer), 1–3d P (buffer→downtime). + + 30-day horizon, Le = = retrains under recovered Ĝ and no graph. τ wq (τ )d (τ ), where d LambdaMART (LightGBM) is the episode-adaptive learned max{downtime risk − µd , 0} and wq (τ ) = max{q(τ ) − µq + policy. We compare it against a conceptual hierarchy. Severity- 1, 1} weights asset priority q. Semiconductors (SE). 10 nodes spanning fab and downonly is an intentionally information-light monitoring baseline. True-root is an attribution-as-action baseline that acts on stream. UCI SECOM [29] calibration. Disruptions: qualitythe injected source. Train-selected static chooses the single hold spike (2–5×), rework-yield drop (10–30%), lead-time 2–5d training-set node with highest average net value and reuses it extension (+7–28d). Lags: 3–7d (quality→inventory), P + Over 21 days, L = λ w (τ )S (τ )+ e s p τ for every test episode. Because the selection uses simulator- (inventory→delay). P + + + derived counterfactual labels for every training candidate, it is λd τ wp (τ )D (τ ), with λs = λd = 1; S (τ ) and D (τ ) a strong full-information non-adaptive benchmark rather than are inventory below and delay above baseline; wp (τ ) = an ordinary logged-data operational policy. Constant-buffer max{p(τ ) − µp + 1, 1} weights product priority p. Critical Materials (CM). 9 nodes. USGS [30] calibration. always intervenes on the archetype-specific loss-driving buffer. It is a strong, domain-informed, oracle-like structural reference Multi-week propagation (4–8 week purity-to-availability lag). because that node is supplied in advance. A perfect-value Disruptions: purity decline (3–10 percentage points), export stockpile draw (30–60%). Over 26 weeks, oracle selects the highest-value option in C(e) ∪ {no action}, restriction P (20–50%), + + with no-action value zero, and provides an upper bound using Le = τ wc (τ )S (τ ), where S (τ ) is allocation shortfall realized counterfactual value. Equal-weight is retained as an above baseline and wc (τ ) = max{c(τ ) − µc + 1, 1} weights end-use criticality c. adaptive heuristic outside the six-level hierarchy. Cross-Archetype Transfer. Feature alignment uses quantile matching within seven aligned KPI classes (inverse-CDF V. C RITICAL SCM-B ENCH V 1 mapping); lag rescaling by median injected lag; loss rescaling CriticalSCM-Bench combines a hand-designed to zero mean/unit variance; frozen ranker evaluation. Reported: temporal DAG, autoregressive trajectories with DI→SE and DI→CM. seasonality/noise/missingness, sampled disruptions, and paired counterfactual labels. Calibration magnitudes draw from Backblaze drive data, electronics lifecycle data, SECOM, and USGS; graph and intervention ground truth remain synthetic. Calibration fixes the marginals of input nodes only. Downstream service-risk dynamics follow from graph propagation and are not empirically calibrated: realized dispersion at those sinks exceeds the calibrated standard deviation by roughly 590× (DI), 302× (SE), and 6× (CM). Benchmark magnitudes at the loss-bearing nodes should therefore be read as graph-determined, not data-derived. Train/test base seeds are disjoint (200 episodes each). V1 collapses sub-tier networks, omits multi-product/sharedresource competition, and keeps graphs fixed within episodes.

VI. E XPERIMENTS The main ranking study uses 5 seeds with 200 disjoint train and test episodes per archetype. Unless stated otherwise, seedlevel point estimates are medians. Reported seed-level intervals are interquartile ranges (IQRs), whereas paired episode-level analyses use bootstrap confidence intervals (CIs). Secondarystudy budgets are stated below. Because the ranking question is separate from detector selection, the main study holds detection timing at the controlled proxy tdet = tinj + 5. This is not a detector-specific threshold crossing; the timing-sensitivity study sweeps the offset. Ranking quality uses normalized discounted cumulative gain at 5 (NDCG@5), while the separate detector study uses a FAR budget of 1 event/month.

A. Decision-Aware Ranking Table II is the principal decision-ranking comparison. The zero floor in Table II acts as a perfect episode-level no-action reference whenever a selected action has negative raw value. The table therefore measures ranking relevance against that reference, not realized value from a deployable act-orabstain policy. Absolute normalized values are small because the feasibility scalar floors a large share of candidate-episodes; §VI-F quantifies this, bounds the interpretation of magnitude, and gives the forced-action and score-abstention accounting. Relative to the full-information train-selected static benchmark, LambdaMART’s median gain is 16.2%/5.7%/11.1% (DI/SE/CM). It does not dominate stronger references: constantbuffer is 5.1% better on DI and the equal-weight adaptive heuristic 3.5% better, both relative to LambdaMART, which leads only on SE and CM and by modest margins. Table II therefore supports a bounded conclusion: episode adaptation adds value in some regimes, but greater model complexity is not uniformly beneficial. Paired bootstrap confidence intervals and sign-flip tests use 1,000 test episodes with disjoint episode seeds across the five runs. They support LambdaMART over static on SE (mean +.0015, 95% CI [.0003,.0027], p = .015) and CM (+.0062, [.0013,.0104], p = .008), but not DI (+.0061, [-.0063,.0182], p = .323). Against constant-buffer, DI is significantly worse (-.0055, [-.0086,-.0028], p < .001), while paired tests favor LambdaMART on SE and CM. Thus the learned model has statistically supported incremental benchmark value on SE and CM, but not on DI; whether those gains justify operational complexity is outside this study. The most consistent result is the distinction between attribution and action. Severity-only obtains 7.2%/6.1%/32.9% of LambdaMART’s value, while true-root is near zero on DI/SE and below LambdaMART on CM. The true-root result is partly a property of the feasibility model rather than of root-cause action as such: a just-disrupted node is frequently driven to the feasibility floor, contributing to near-zero DI value (§VI-F). LambdaMART reaches 57%/55%/88% of the perfect-value oracle. The buffer is value-optimal in only 19.5%/18.0%/9.5% of episodes, so a fixed buffer is strong on average without being episode-optimal.

and CM and lower recall on DI. In a controlled graph-swap (3 seeds × 150; ranker retrained per regime), recovered Ĝ versus benchmark G gives net value 0.060/0.060 (DI), 0.059/0.048 (SE), and 0.038/0.038 (CM); no graph gives 0.036/0.035/0.024. These secondary values are not directly comparable to Table II. Structural recall and downstream decision value are different quantities, and the recovered graph does not consistently reduce value here. C. Learning Curve With 25/50/100/200 training episodes (3 seeds, 200 test), LambdaMART median net value is .028/.058/.067/.068 (DI), .039/.041/.046/.052 (SE), and .016/.022/.023/.053 (CM). Medians improve with sample count, but the path is not smooth. Seed-level trajectories remain noisy, and CM changes sharply only at 200. The curves show that the policy is sample-sensitive; they do not show convergence. D. Feature Ablation Leave-one-out (3 seeds × 150 episodes): Lifecycle feasibility carries the clearest signal. Removing lifecycle_feasibility causes the largest median drop (-.050/-.012/-.020 for DI/SE/CM); removing edge_lag gives -.028/-.015/-.013. Other removals are smaller or can improve value, which suggests redundancy rather than a feature set in which every variable is uniformly helpful. E. Cost-Scale Sensitivity We sweep ca ∈ {.1, .2, .3, .5} (3 seeds × 150). LambdaMART exceeds constant-buffer only at .1 on DI, is equal or slightly higher on SE, and leads CM at .1–.2 but loses at .3–.5. Buffer-optimal frequency falls with cost. The cost model is an explicit benchmark factor, not an empirical monetary estimate; the sweep evaluates policy dependence on this factor. Monetary calibration remains external to v1. F. Feasibility and Intervention Sensitivity

The feasibility scalar is a synthetic intervention-effect proxy, not an empirical recovery probability. We therefore rerun the injected-root versus loss-optimal agreement audit under four paired specifications: no attenuation (ϕ = 1), mild attenuation (γ = 0.375, ϕmin = 0.1), the current setting (γ = 0.75, ϕmin = 0.1), and a higher floor (γ = 0.75, ϕmin = 0.3). Root/optimal B. Detection and Causal-Structure Results disagreement spans 89.7–99.7% (DI), 97.3–100% (SE), and Table III reports detector coverage and delay under the FAR 57.0–74.3% (CM), remaining high across all four specifications constraint. with greater variation on CM. On DI, LSTM-AE provides greater event coverage and The four specifications are not fully independent because substantially shorter censored delay than EWMA under the floor saturation makes portions of the attenuation sweep same FAR budget. Detector selection should therefore consider inert. Since ϕ uses calibrated rather than larger realized coverage and lead time jointly. The detector study uses a longer dispersion (§V), 51.2%/47.8%/51.9% of DI/SE/CM candidatemonitoring window than the intervention-loss windows in §V. episodes are floored, rising to 87.8% (DI) and 75.0% (CM) CM EWMA detects fewer than half the events in every seed, so for disrupted candidates. This contributes to near-zero DI trueits within-seed censored medians equal the 120-week horizon. root value and limits interpretation of magnitude. The audit Using a PCMCI significance threshold αPCMCI = 0.01, supports persistence of root/action disagreement, not empirical PCMCI recovers 0.75 / 0.44 / 0.67 of true edges on DI / SE / calibration of its size. CM (medians over five seeds). Granger recovers 0.58 / 0.56 / Table IV summarizes the effects of partial and delayed 0.78 on the same seeds, so Granger has higher recall on SE interventions.

TABLE II D ECISION - AWARE RANKING (5 SEEDS × 200 TEST EPISODES ). C ELLS : MEDIAN NORMALIZED NET VALUE / NDCG@5; BOLD MARKS THE BEST NON - ORACLE NET VALUE . VALUES ARE NORMALIZED BY NO - ACTION LOSS AND ROUNDED TO THREE DECIMALS ; TEXT PERCENTAGES USE UNROUNDED SUMMARIES . L ABELS FLOOR NEGATIVES AT ZERO . §VI-F REPORTS FORCED - ACTION AND SCORE - ABSTENTION ACCOUNTING . Policy LambdaMART Constant-buffer Train-selected static Equal-weight Severity-only True-root Perfect-value oracle

DI

SE

CM

0.075 / 0.267 0.079 / 0.244 0.064 / 0.262 0.077 / 0.274 0.005 / 0.100 0.000 / 0.083 0.131 / –

0.049 / 0.249 0.047 / 0.217 0.047 / 0.217 0.048 / 0.246 0.003 / 0.103 0.000 / 0.098 0.090 / –

0.050 / 0.569 0.045 / 0.393 0.045 / 0.393 0.039 / 0.155 0.016 / 0.420 0.024 / 0.550 0.057 / –

TABLE III D ETECTION AT FAR ≤ 1/ MONTH (5 SEEDS × 200). C ELLS : EVENT RATE / CONDITIONAL / CENSORED DELAY. D ELAYS ARE WITHIN - SEED MEDIANS OVER DETECTED / ALL EVENTS ( MISSES = 120 STEPS ), THEN MEDIANS ACROSS SEEDS ; UNITS ARE DAYS (DI/SE) OR WEEKS (CM). B OLD MARKS THE FRONTIER . Detector

DI rate / cond. / cens. (d)

SE rate / cond. / cens. (d)

CM rate / cond. / cens. (w)

EWMA+MAD Iso. Forest LSTM-AE USAD MTAD-GAT

44% / 6.0 / 120 28% / 15.5 / 120 60% / 4.0 / 11.5 48% / 5.0 / 120 57% / 4.0 / 10

48% / 5.0 / 120 30% / 11.0 / 120 88% / 2.0 / 2 78% / 3.0 / 4 73% / 2.0 / 3

43% / 5.0 / 120 62% / 8.0 / 12.0 69% / 3.0 / 8 55% / 4.0 / 24.5 55% / 4.0 / 34.5

TABLE IV I NTERVENTION SENSITIVITY (3 SEEDS × 150 TRAIN / TEST ). C ELLS : L AMBDA MART / CONSTANT- BUFFER / SEVERITY- ONLY NORMALIZED VALUE ; PARENTHESES SHOW L AMBDA MART RETENTION VERSUS FULL CLAMP. Intervention

DI

SE

CM

Full at detection .060/.071/.002 (100%) .048/.047/.004 (100%) .038/.026/.015 (100%) 75% effect .045/.052/.002 (75%) .033/.032/.003 (68%) .026/.023/.011 (70%) 50% effect .020/.030/.001 (33%) .019/.017/.001 (39%) .020/.017/.006 (52%) Five-step delay .039/.056/.000 (64%) .030/.031/.003 (62%) .025/.024/.006 (66%) 75% + delay .030/.040/.000 (49%) .019/.020/.002 (40%) .020/.020/.004 (54%)

Partial effects alter value and ordering: constant-buffer remains stronger on DI and becomes competitive on delayed SE. Policy ordering is therefore not invariant to intervention fidelity. A further stochastic-response stress test introduces efficacy variation, outright failure, delay, ramping, and decay (5 regimes, 5 seeds). LambdaMART retention relative to the fullclamp benchmark falls to 0.51/0.29/0.24 under the moderate regime and 0.13/0.00/0.06 under the weak temporary regime for DI/SE/CM. No candidate has positive net value in 45– 99% of episodes, and CM’s optimal node changes in up to 63%. Scenario-matched retraining on noisy labels does not consistently outperform the frozen full-clamp ranker.

rather than a regime-agnostic deployment rule. On full-clamp episodes, this rule retains 30–52% of the corresponding clippedlabel headline value. Across stochastic regimes, it abstains in 71–100% of episodes and remains approximately break-even (minimum mean test value −0.004).

G. Circular-Evaluation Control

To test dependence on counterfactual labels, a ranker trained once on hard clamps is frozen and evaluated under unseen intervention specifications against scenario-matched retraining (5 seeds × 100 train/test). Clipped labels supply a zero-valued benchmark reference but assume exact episode-level knowledge of whether an action’s Across the four mismatched idealizations, frozen/scenariorealized net value is negative. They do not supply a deployable matched median value ratios range from 1.02 to 1.13 on DI and no-action rule. Forced action has mean value −0.13/ − 0.12/ + from 0.93 to 1.21 on SE; the CM mean across specifications 0.00 under full clamp (DI/SE/CM), and roughly half of those is 0.83. On CM, the frozen ranker loses about half its value episodes lack a positive-net candidate; under stochastic response relative to matched retraining under five-step delay (ratio 0.49). the share reaches 99%. Separately for each archetype, seed, and Dependence on counterfactual labels is therefore mild on DI response regime, a training-selected rule chooses the top-score and SE but material on CM, and the circular-evaluation concern cutoff that maximizes summed normalized raw net on training is reduced rather than removed: label mismatch, OOD, and episodes, then freezes it for test. Because the cutoff is selected transfer tests all remain inside the same autoregressive simulator separately for each assumed response regime, this is diagnostic family.

(spare_buffer_level), obtaining 24.9 (0.046 normalIn a paired leave-one-disruption-template-out OOD test, the ized), while spare availability (spare_availability) same fitted model retains 95%/92%/62% in-domain net value obtains 16.2. This episode is illustrative rather than general on DI/SE/CM (0.069→0.065, 0.052→0.048, 0.047→0.029). evidence, but it shows that the largest anomaly, injected source, NDCG retention is 96%/95%/78%. Constant-buffer OOD net and best intervention can differ. value is 0.072/0.046/0.049, exceeding LambdaMART on DI and N. Guarded Operator-Explanation Layer: Secondary Study CM. Thus CM has a substantial held-out-disruption weakness. This secondary study asks whether a language model can verbalize a fixed DACRI decision, whether evidence constraints I. Schema Transfer improve first-pass acceptance, and whether validation provides Frozen DI→SE and DI→CM rankers retain 75% (0.036 vs fallback. Packets omit counterfactual labels, and the schema 0.048) and 65% (0.025 vs 0.038) of in-domain value. Transfer has no recommendation field. The validator checks decision is a secondary stress test. The larger CM drop is consistent preservation, packet entities, edges, lags and numbers, and with several schema and time-scale differences. Because those rejects verified root-cause or promised-outcome claims; rejected differences move together, this experiment does not isolate lag drafts use a packet template. scale as the cause. We build 90 packets (30 per archetype, 11 carrying a noaction decision) and generate three times per packet under J. Controlled Lag-Scale Test a generic and an evidence-constrained prompt: 540 calls To probe the lag explanation directly, we ran a controlled using the recorded API identifier deepseek-v4-flash. test that holds topology, weights, calibration, cost, disruptions, Supplementary Artifact A contains generation settings, prompt and the feature schema fixed and varies only the edge-lag versions, packet schema, validator rules, and complete outputs. scale (factor f ∈ {0.5, 0.75, 1.5, 2.0} about the trained regime), For this explanation study only, no-action is locked when the evaluating a frozen ranker against one retrained at each f top ranker score falls below the pre-specified 10th percentile of (5 seeds). Across the four factors, no monotone degradation an out-of-fold training top-score distribution. This rule uses no pattern is observed. CM degrades only on the stretch side— test labels and is distinct from the scenario-matched diagnostic frozen-to-retrained value falls to 0.79/0.60 at f = 1.5/2.0, the cutoff in §VI-F. day-to-week direction—while lag compression does not hurt No output redirected the fixed decision, and none of the and DI/SE show no consistent effect. Thus lag scale alone does 66 no-action outputs recommended intervention. Evidencenot explain the cross-archetype transfer drop in this controlled constrained prompting improves first-pass acceptance from test. 74.1% to 90.4%, with 59 discordant pairs favoring it against 15 favoring generic (exact McNemar p = 2.55 × 10−7 ). EvidenceK. Real-Data Detection Sanity Check constrained drafts achieve 100% structured grounding under the We use Backblaze Drive Stats Q1 2024 (25.2M drive-days, specified checks, compared with 99.3% for generic prompting; 978 failures). Isolation Forest detects pre-failure anomalies in guarded outputs reach 100% after deterministic validation and 39% of failed drives at FAR 0.057 per 1k drive-days, median 21 fallback. All 26 rejections in the evidence-constrained condition days before failure; EWMA detects 4%. The Backblaze result arise from one banned claim: describing the intervention target supports the plausibility of early detection and the detection-toas the verified root cause, an attribution–intervention conflation. decision motivation; detector ordering differs from simulation Generation stability is the weakest outcome: neither prompt and therefore remains deployment-specific. yields three identical generations for any packet; fallback raises exact stability only to 2.2%. Prose varies across generations, so L. Detection-Timing Sensitivity deployment-oriented reporting should disclose this variability. At detection offsets 3/5/10/14 time steps—days for DI/SE Checks cover structured claims and numeric tokens, not and weeks for CM—(3 seeds × 150), LambdaMART value usefulness, style, or complete semantic faithfulness. All packets is .059/.060/.011/.008 (DI), .059/.048/.027/.010 (SE), and .079/.038/.013/.001 (CM). Because intervention begins at detection, late flags leave less recoverable loss. Here the TABLE V G UARDED EXPLANATION LAYER (90 PACKETS × 3 GENERATIONS PER mechanism is direct: detection timing changes the value that PROMPT; 540 CALLS ), IN PERCENT. D RAFT ACCEPTANCE PRECEDES remains available to the policy. The effect is severe, especially FALLBACK ; G UARDED INCLUDES 9.6% TEMPLATE FALLBACK . G ROUNDING on CM. COVERS ENTITIES , EDGES , LAGS , AND NUMBERS ; STABILITY REQUIRES H. Out-of-Distribution Generalization

IDENTICAL TEXT ACROSS GENERATIONS .

M. Illustrative Intervention Walkthrough Consider DI episode 2000001. A fault-rate spike is injected at drive failure rate (drive_failure_rate) at t = 49 and detected at t = 54; no-action loss is 540.1. Severityonly selects recovery throughput (rework_throughput) and true-root selects drive_failure_rate; both have zero net value. LambdaMART selects spare-buffer level

Template Generic Ev.-constr. Guarded Decision preserved Draft accepted Structurally grounded Banned claim absent Template fallback Exactly stable over 3 gens

100 — 100 100 0 100

100 74.1 99.3 74.4 — 0

100 90.4 100 90.4 — 0

100 90.4 100 100 9.6 2.2

are simulator-derived, and no human audit was completed; this remains a constrained secondary component. VII. D ISCUSSION Attribution is not equivalent to intervention selection: severity-only and true-root miss value, although feasibilityfloor saturation partly drives the true-root result. Adaptation is statistically supported over the full-information static benchmark on SE and CM, but not DI, where constant-buffer wins. Operational complexity was not evaluated. The oracle gap leaves further policy headroom. The stress tests map the benchmark’s validity envelope. Intervention fidelity, cost, timing, and graph specification alter value and ordering, while CM is the weakest held-out-disruption regime. These are explicit, replaceable benchmark factors rather than hidden deployment estimates: the results locate where adaptive ranking helps within the controlled environment and where structural policies remain preferable. CriticalSCM-Bench supports controlled comparative evaluation, not deployment certification. Independent simulators and retrospective operational data must test transportability, monetary costs, and organizational feasibility. The guarded explanation study evaluates a constrained renderer’s safety and grounding, not human factors. Artifact availability. Supplementary Artifact A—the benchmark, the evaluation harness, configurations, seeds, evaluation scripts, prompts, validator rules, and every stored output behind the reported numbers—is available at https://github.com/dyshang/dacri-criticalscm-bench

VIII. C ONCLUSION DACRI starts from a simple distinction: finding the source of a disruption is not the same as choosing where to act. CriticalSCM-Bench makes that distinction measurable through a paired, cost-aware objective. Within the benchmark, severity and true-root often miss recoverable value. LambdaMART improves a full-information static benchmark on SE and CM, but loses to a strong structural rule on DI. Partial, delayed, cost, and OOD tests identify regimes with statistically supported benchmark gains from episode-specific ranking and those favoring structural rules. The contribution is a falsifiable benchmark for deciding where to act, not universal learnedpolicy superiority. Independently implemented simulators and retrospective operational data provide the next tests of transportability. R EFERENCES [1] Executive Order 14017, “America’s Supply Chains,” Federal Register, 2021. [2] U.S. Congress, “CHIPS and Science Act,” Public Law 117-167, 2022. [3] U.S. Department of Energy, “2023 Critical Materials Assessment,” 2023. [4] A. Dolgui, D. Ivanov, and B. Sokolov, “Ripple effect in the supply chain: an analysis and recent literature,” Int. J. Prod. Res., vol. 56, no. 1–2, pp. 414–430, 2018. [5] T. Y. Choi and Z. Wu, “Triads in Supply Networks: Theorizing Buyer– Supplier–Supplier Relationships,” J. Supply Chain Manage., vol. 45, no. 1, pp. 8–25, 2009. [6] D. Ivanov and A. Dolgui, “A digital supply chain twin for managing the disruption risks and resilience in the era of Industry 4.0,” Prod. Plan. Control, vol. 32, no. 9, pp. 775–788, 2021.

[7] C. C. Sherbrooke, “METRIC: A Multi-Echelon Technique for Recoverable Item Control,” Oper. Res., vol. 16, no. 1, pp. 122–141, 1968. [8] X. Su, L. Zeng, B. Shao, and B. Lin, “Data-driven optimization for production planning with multiple demand features,” Kybernetes, vol. 54, no. 1, pp. 110–133, 2025, doi: 10.1108/K-04-2023-0690. [9] X. Li, Y. Hu, X. Su, and B. Shao, “A Data-Driven Newsvendor Problem with Shifting Demand: A Deep Autoregressive Model with Attention Mechanism,” J. Eng. Sci. Technol. Rev., vol. 16, no. 3, pp. 74–83, 2023. [10] B. Shao, X. Su, X. Li, and X. Zhuo, “Mixed-frequency data-driven forecasting port throughput: a novel attention-DeepAR-MIDAS model,” Int. J. Shipping and Transport Logistics, vol. 20, no. 3, pp. 338–358, 2025, doi: 10.1504/IJSTL.2025.147408. [11] F. T. Liu, K. M. Ting, and Z.-H. Zhou, “Isolation Forest,” in Proc. IEEE ICDM, 2008, pp. 413–422. [12] J. Audibert et al., “USAD: Unsupervised Anomaly Detection on Multivariate Time Series,” in Proc. ACM KDD, 2020, pp. 3395–3404. [13] H. Zhao et al., “Multivariate Time-Series Anomaly Detection via Graph Attention Network,” in Proc. IEEE ICDM, 2020, pp. 841–850. [14] H. Ren et al., “Time-Series Anomaly Detection Service at Microsoft,” in Proc. ACM SIGKDD, 2019, pp. 3009–3017. [15] S. Tariq, M. Baruwal Chhetri, S. Nepal, and C. Paris, “Alert Fatigue in Security Operations Centres: Research Challenges and Opportunities,” ACM Comput. Surv., vol. 57, no. 9, 2025. [16] T. Cao, M.-H. Trinh, A. Deng, Q.-N. Nguyen, K. Duong, N.-M. Cheung, and B. Hooi, “Are Anomaly Scores Telling the Whole Story? A Benchmark for Multilevel Anomaly Detection,” in Proc. IEEE/CVF CVPR Workshops, 2026. [17] Y. Zhao, S. Zhang, Y. Wu, Y. Sun, Y. Sun, D. Pei, C. Bansal, and M. Ma, “Triage in Software Engineering: A Systematic Review of Research and Practice,” arXiv:2511.08607, 2025. [18] A. Gulay and L. Maglaras, “Alignment of Cybersecurity Incident Prioritisation with Incident Response Management Maturity Capabilities,” arXiv:2410.02259, 2024. [19] L. Porcelli, “A Feature Engineering Approach for Business ImpactOriented Failure Detection in Distributed Instant Payment Systems,” arXiv:2510.21710, 2025. [20] C. Granger, “Investigating Causal Relations by Econometric Models and Cross-spectral Methods,” Econometrica, vol. 37, no. 3, pp. 424–438, 1969. [21] J. Runge, “Discovering contemporaneous and lagged causal relations in autocorrelated nonlinear time series datasets,” in Proc. UAI/PMLR, 2020, vol. 124, pp. 1388–1397. [22] R. Pamfil et al., “DYNOTEARS: Structure Learning from Time-Series Data,” in Proc. AISTATS, 2020, pp. 1595–1605. [23] Z. Liang, L. van der Laan, and A. Alaa, “Hybrid Meta-learners for Estimating Heterogeneous Treatment Effects,” in Proc. AISTATS, 2026. [24] P. Lewis et al., “Retrieval-Augmented Generation for KnowledgeIntensive NLP Tasks,” in Advances in Neural Information Processing Systems, vol. 33, 2020. [25] P. Manakul, A. Liusie, and M. Gales, “SelfCheckGPT: ZeroResource Black-Box Hallucination Detection for Generative Large Language Models,” in Proc. EMNLP, 2023, pp. 9004–9017, doi: 10.18653/v1/2023.emnlp-main.557. [26] S. Min et al., “FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation,” in Proc. EMNLP, 2023, pp. 12076–12100, doi: 10.18653/v1/2023.emnlp-main.741. [27] Backblaze, “Drive Stats: Hard Drive Reliability Test Data.” [Online]. Available: https://www.backblaze.com/cloud-storage/resources/hard-drive -test-data. Accessed: Jul. 28, 2026. [28] C. W. Babbitt et al., “Disassembly-based bill of materials data for consumer electronic products,” Scientific Data, vol. 7, Art. no. 251, 2020. [29] UCI Machine Learning Repository, “SECOM Dataset.” [Online]. Available: https://archive.ics.uci.edu/ml/datasets/SECOM. Accessed: Jul. 28, 2026. [30] U.S. Geological Survey, Mineral Commodity Summaries 2025, ver. 1.2, Mar. 2025, doi: 10.3133/mcs2025.

Related documents

• An integrated multi-criteria and evolutionary optimization framework for supply chain disruption risk prioritization and mitigation.#275528
• Critical path method networks and their use in claims analysis.#497197
• Towards a decision aid model to build a production plan integrating a holistic vision of the automotive supply chain#468357
• Causal Machine Learning: A Deductive–Inductive Framework for Sociological Research#453500
• Causal Discovery in Linear Models with Unobserved Variables and Measurement Error#128301