Point by point response to reviewers
Manuscript systems-4601389  |  Revision dated 29 September 2026
Laboratory Evaluation of Mechanics Informed Cross Fault Risk Indicators for Injection Decision Support
Dear Editor and Reviewers,
We have revised the manuscript and performed additional analyses of the recovered experimental records. This response identifies every change and distinguishes it from requested evidence that remains unavailable. Yellow highlighting marks revised manuscript text and replacement figure captions; newly added or corrected supplementary material is also highlighted. Locations below use the revised section, figure, and table numbers because pagination and line numbering have changed.
The original cohort contains 18 stage events within three specimens. We now add six independently archived slow-creep stage peaks from a fourth specimen, evaluated without external refitting. The total is 24 trajectories, below the requested doubling to 36. Field validation has not been added. The independent outcomes reveal policy-transfer failures despite high domain support. We also replace ambiguous archived event partitions with a fully reproducible raw-pump reconstruction.
Reviewer 1
R1.1  Abstract, comment 1
Comment: At line 14, what is meant by “governed states”?
We removed the undefined expression from the rewritten abstract. The text now describes persistent Watch, Mitigate and Hold indicators and a separate empirical support flag. Section 2 defines their meanings and the proposed advisory mapping. These outputs do not constitute control permissions, and no operator or actuator workflow was tested.
Revision location: Abstract; Sections 2.1–2.2; Figure 8. Status: Addressed.
R1.2  Abstract, comment 2
Comment: At line 16, what is meant by “mechanics aligned variables”?
We replaced this phrase with an explicit description of seven variables representing hydraulic loading, stress changes, preslip, slip velocity and acoustic activity. Section 3.3 and Table 3 define the individual quantities. “Mechanics informed” now refers to physically interpretable feature selection, not enforcement of governing equations in the model. The revision audit also corrected two previously imprecise definitions: the acoustic feature is the logarithm of one plus the trailing mean AE rate, and the nominal 30 s features use 30 samples spanning 29 s.
Revision location: Abstract; Section 3.3; Equation (1); Table 3. Status: Addressed.
R1.3  Abstract, comment 3
Comment: The closing abstract claim about an Industry 5.0 architecture is too general and does not present the direct implications of the study.
The abstract now concludes with the direct evidential implication: interception, early alert exposure and empirical source support must be reported separately before considering an advisory application. It includes the weak Hold result and the new smooth-source deletion sensitivity, for which final-30 s support ranges from 19.7% to 88.6%. We removed the claim that these measurements validate an Industry 5.0 operational architecture and explicitly state that they do not establish automatic control, field readiness, population generalization or CCUS network benefits.
Revision location: Abstract; Section 6. Status: Addressed.
R1.4  Abstract, comment 4
Comment: Rewrite the entire abstract to better present the goal, obtained results, implications and authors’ contributions.
We rewrote the abstract around the problem, analytical contribution, dataset, methods and principal results. It identifies the study as a retrospective analysis of 18 pressure-stage events nested within three laboratory specimens; explains the compact representation, source-calibrated persistent indicators and separate Mahalanobis support flag; and reports interception counts, warning times, early alert exposure, support coverage and sensitivity. The conclusion is limited to further advisory testing and identifies the missing independent replication and field validation.
Revision location: Abstract. Status: Addressed.
R1.5  Keywords, comment 1
Comment: Replace “digital twin”, “predictive risk management” and “subsurface injection” because they already appear in the title.
These three keywords have been removed. The revised list is CCUS supply chains; fault reactivation; domain shift; selective prediction; event based validation; human oversight; and Industry 5.0. The title has also been changed to reflect the laboratory evaluation rather than digital-twin readiness.
Revision location: Title; Keywords. Status: Addressed.
R1.6  Introduction, comment 1
Comment: Convert the question about estimating the operational state of an injection node before disruption into an understandable declarative sentence.
We replaced the question with: “These formulations motivate estimating the local condition of an injection node before declaring a loss of service capacity.” The subsequent paragraphs specify that the present evaluation concerns archived laboratory risk indicators and does not calculate injection capacity or network disruption.
Revision location: Section 1, opening paragraph and scope paragraph. Status: Addressed.
R1.7  Introduction, comment 2
Comment: Resolve the apparent contradiction between an injection-node digital shadow and a “digital twin ready” module.
We agree that the terminology overstated the demonstrated maturity. The revised manuscript consistently calls the implemented work an offline analytical prototype. It neither receives automatic live updates, as a digital shadow would, nor provides the operational feedback and integration of a deployed twin. “Digital twin ready” has been removed from the title and the contribution claim. A future local advisory interface is explicitly separated from the evaluated analysis.
Revision location: Title; Section 1, contribution paragraph; Sections 2.2 and 5.5; Figure 1. Status: Addressed.
R1.8  Introduction, comment 3
Comment: Resolve the scope inconsistency between a node-level contribution and RQ4 referring to a higher-level CCUS supply-chain twin.
The empirical scope is now stated as retrospective fault-slip risk assessment using laboratory specimens, with a local injection-node advisory application proposed for future development. RQ4 asks how a local interface could distinguish urgency from empirical support and which components remain conceptual. CCUS planning is retained only as motivation. No supply-chain optimizer, API, capacity mapping, network integration or network benefit is claimed as an implemented or evaluated result.
Revision location: Section 1, scope paragraph and RQ4; Sections 2.1 and 5.3; Figure 1. Status: Addressed.
R1.9  Introduction, comment 4
Comment: Remake Figure 1 because it appears to be generated.
Figure 1 has been redrawn as a deterministic diagram with editable/vector source. It explicitly separates the implemented offline workflow from proposed live acquisition, advisory actions, physical constraints and network use. The caption identifies the study boundary and states that no live API, actuator, network optimizer or field digital twin was implemented.
Revision location: Figure 1 and caption; figure source files in the revision archive. Status: Addressed.
R1.10  Conceptual and Systems Framework, comment 1
Comment: Clarify what “Unsupported” means and reformulate the statement that it requires operator verification.
“Unsupported” now has a precise, limited definition: the seven-variable state has Mahalanobis distance D(t) greater than the selected empirical source cutoff. It indicates low similarity under the chosen scaling and covariance representation; it does not prove failure, sensor malfunction or incorrect prediction. Passing the gate likewise does not certify safety. In the proposed interface, an unsupported flag requests review while independent physical alarms and approved limits remain applicable.
Revision location: Sections 2.1 and 3.7; Figure 8. Status: Addressed.
R1.11  Conceptual and Systems Framework, comment 2
Comment: Define “permitted authority class” and explain how it is calculated.
We removed “permitted authority class” and use “advisory action class.” It is a proposed categorical mapping from displayed risk tier and support flag, not a learned quantity, numerical safety score or computed permission to actuate. The classes and their mapping are now stated explicitly. All outputs remain advisory because neither operator decisions nor control actions were evaluated.
Revision location: Sections 2.1–2.2; Figure 8; Tables 1 and 7. Status: Addressed.
R1.12  Conceptual and Systems Framework, comment 3
Comment: Explain what “Digital twin readiness” means.
We removed the readiness claim rather than assign an unsupported maturity level. The manuscript now distinguishes the offline prototype from a live digital shadow and a deployed digital twin. It lists live synchronization, prospective validation, independent specimens, operator testing, physical operating limits and integration as future requirements. No technology-readiness level or certified deployment status is assigned.
Revision location: Title; Sections 1, 2.2 and 5.5; Table 10. Status: Addressed.
R1.13  Conceptual and Systems Framework, comment 4
Comment: Explain how the three Industry 5.0 principles relate to readiness and why they require a particular readiness level.
The revised text makes clear that these principles do not establish a universal readiness threshold. They motivate future design choices: resilience motivates visibility of risk and model limitations; sustainability motivates examining missed events and unnecessary alert activity; and human-centered operation motivates accountable operator review. Their practical benefits would require separate intervention, environmental and human-factors evaluation. None is claimed to have been demonstrated by this laboratory analysis.
Revision location: Sections 1, 2.2 and 5.4. Status: Addressed.
R1.14  Conceptual and Systems Framework, comment 5
Comment: Define “Permitted action class” and list the possible classes.
The revised advisory classes are routine monitoring for Normal; enhanced monitoring for Watch; operator assessment of possible rate modification for Mitigate; operator assessment of possible interruption for Hold; and review required when the support flag is unsupported. Section 2.2 and Figure 8 give this proposed mapping. These classes cannot authorize pressure or rate changes, and an unsupported flag must not suppress independent safety limits or alarms.
Revision location: Section 2.2; Figure 8. Status: Addressed.
R1.15  Materials and Methods, comment 1
Comment: Explain clearly what “target domain” means where the final source scaler and covariance model are applied to untouched target data.
Section 3.4 now defines source and target by specimen group before they are used in the support-gate description. In rough-to-smooth transfer, the source is 12 events from the two rough specimens and the target is six events from the smooth specimen. These roles reverse in smooth-to-rough transfer. Scaling, fitting, intercept calibration, score thresholds and support calibration use the source records only; the held-out target group supplies the reported evaluation outcomes. We also clarify that event-based source cross-validation is not independent-specimen validation.
Revision location: Sections 3.4 and 3.7; Table 2. Status: Addressed.
R1.16  Materials and Methods, comment 2
Comment: State whether Equations (1)–(5) are proposed by the authors and give their sources if they are not.
The revision assigns provenance to each equation. Equation (1) is our selected state representation, assembled from measured and derived quantities; Equation (3) is our empirical-quantile policy definition; and Equation (5) defines our descriptive support fraction. They are not claimed as new governing laws. Equation (2) uses the established logistic inverse link with a source-fitted additive intercept correction, and Equation (4) is classical Mahalanobis distance with Ledoit–Wolf covariance shrinkage. Appropriate direct references have been added. Equation (2)’s extra parenthesis and the distinction between raw and scaled regression state have also been corrected.
Revision location: Sections 3.3–3.7; Equations (1)–(5); References. Status: Addressed.
R1.17  Materials and Methods, comment 3
Comment: Justify why Table 4 uses source-negative quantiles 0.90, 0.95 and 0.99 instead of other values.
We now identify them as illustrative design settings rather than uniquely optimal or physically prescribed values. Before persistence and subject to quantile ties, they represent increasingly extreme source-negative scores, with nominal upper-tail fractions of approximately 10%, 5% and 1%. These are not target false-alarm guarantees, acceptable field risks or economic budgets. We removed any implication of an independently preregistered optimum.
We additionally evaluated every combination of these three quantiles and persistence of 1, 2, 3 or 5 s at each horizon in both directions: 72 configurations using source-only threshold calibration. No target outcome was used to choose a new preferred setting. The results expose sensitivity rather than justify one universal policy: rough-to-smooth Watch remained 6/6, whereas Mitigate ranged from 2/6 to 6/6 and Hold from 0/6 to 5/6.
Revision location: Section 3.5; Table 4; Section 4.2; Supplementary Tables S1 and S4; policy_grid.csv. Status: Addressed.
R1.18  Materials and Methods, comment 4
Comment: Explain how “pre-horizon intervention burden” should be interpreted, rather than only defining it.
We explain it as early alert exposure relative to the nominal horizon within the selected 180 s window. For example, 42.2 s/event means that the indicator was active for that average duration before the 30 s horizon; it does not mean that an intervention was performed or that 42.2 s of production was lost. Greater early activity might be useful or unnecessary depending on an untested response policy. Zero burden means no pre-horizon activation in those selected windows, not absence of false alarms. Equivalent limitations are stated for duty fraction and repeated in the relevant captions.
Revision location: Section 3.8; Table 5; Figures 5 and 10; Sections 4.2 and 5.7. Status: Addressed.
R1.19  Results, Discussion and Conclusions, comment 1
Comment: At least double the six smooth and twelve rough events and redo results, discussion and conclusions because the sample is extremely small.
We expanded the empirical evaluation using a genuinely independent published laboratory record. Wang et al. (2020), Figshare version 3 (DOI 10.6084/m9.figshare.11948739.v3), supplies SC2, one additional polished Bentheim specimen under slower pressurization. Its six nonoverlapping stage-wise slow-creep velocity peaks have complete feature histories. The accompanying SC1 record was excluded because its AE and hydraulic histories match the already-used smooth specimen. This duplication check prevents counting the same experiment twice.\n\nBoth original source policies and support gates were applied to all six SC2 peaks without refitting, external recalibration or threshold selection. Rough-source Watch intercepted 5/6, whereas Mitigate and Hold intercepted 0/6; smooth-source interception was 4/6, 6/6 and 3/6. Results, discussion, abstract and conclusions now explain the failures and early alert exposure, despite 99.7–100.0% overall state support.\n\nThe total therefore increases from 18 trajectories in three specimens to 24 in four specimens. This is genuine expansion, but it does not fulfill the requested doubling to at least 36, and the original rough/smooth groups are not silently enlarged by pooling a different loading protocol. Additional archives were screened and excluded for duplicate records, missing or incompatible channels, and ambiguous overlapping events. We request consideration of the added independent test within the narrowed laboratory scope; further independent experiments remain necessary to fulfill the numerical request.
Revision location: Abstract; Sections 3.10, 4.7, 5.2, 5.7 and 6; Supplementary Section S11 and Table S12; external-data screening audit. Status: Partly addressed.
R1.20  Results, Discussion and Conclusions, comment 2
Comment: Wilson intervals such as 61.0%–100.0% are too wide for precise claims, and three-decimal PR AUC values overstate statistical precision.
We removed precise population-performance claims and now prioritize counts such as 6/6 and 12/12. The original 6/6 Wilson interval remains 61.0%–100.0%; reporting fewer decimal places cannot remove this uncertainty. More fundamentally, successive stages share specimen history, so the Wilson intervals are now labelled nominal binomial event-count reference intervals rather than confidence limits for a new fault population. Event-bootstrap ranges are explicitly conditional on the observed specimens and do not account for between-specimen uncertainty.
AP and Brier values in the main results are rounded to two decimal places. We also identify the reported discrimination statistic correctly as noninterpolated average precision and supply descriptive event-resampling ranges and per-specimen summaries. Exact numerical values are retained only in reproducibility files, where they represent computational precision rather than inferential certainty.
Revision location: Sections 3.8 and 4.1–4.2; Tables 5 and 9; Section 5.7; Supplementary Table S10. Status: Addressed.
R1.21  Results, Discussion and Conclusions, comment 3
Comment: Explain why Table 5 reports 16.7% interception for the 10 s Hold tier if that tier is advisory and near-failure evidence is limited.
The 16.7% is the observed 1/6 interception of a candidate alert rule, not an operating authorization or evidence of a successful shutdown system. It is retained to report that the rule missed five selected smooth-target events. All operational action classes, including Hold, now remain advisory and untested. Table 5 also explicitly reports ungated indicator performance; the domain-support flag is separate.
The added sensitivity analysis reinforces this distinction. At q99, rough-to-smooth Hold intercepted 2/6 events with 1 s persistence, 1/6 with 2 s, and none with 3 or 5 s. This is negative evidence about the reliability of short-horizon transfer, and the conclusions explicitly reject automatic Hold use on the present evidence.
Revision location: Sections 3.5, 3.8, 4.2 and 4.5; Table 5; Figure 9; Section 6; Supplementary Table S4. Status: Addressed.
R1.22  Results, Discussion and Conclusions, comment 4
Comment: Provide a plausible justification for the 99th-percentile primary support cutoff; Figures 6–7 and Table 8 depend strongly on percentile, raising concern that q99 was selected to obtain desired results.
We no longer present q99 as a validated primary operating threshold or independent optimum. It is retained as the illustrative reference of the original analysis, with approximately 1% exceedance in the pooled cross-fitted source sample only. No prospective preregistration or external selection record is available. The revised interpretation expressly acknowledges this and reports the complete q95–q99.5 range without selecting a replacement cutoff from target outcomes.
We also went beyond percentile sensitivity. Complete recalibration after deletion of a source event changes final-30 s support to 90.6%–93.9% for the rough source and 19.7%–88.6% for the smooth source. Alternative distance gates, temporal thinning and whole-rough-specimen calibration produce additional changes. We now state that the absolute coverage is materially calibration-dependent; the gate is an exploratory source-support diagnostic, not a safety boundary or universal authority criterion.
Revision location: Sections 3.7 and 4.3–4.4; Figures 6–7; Tables 6 and 8; Section 5.7; Supplementary Section S5 and Tables S5–S8. Status: Addressed.
R1.23  Results, Discussion and Conclusions, comment 5
Comment: Remake Figure 8 because it appears to be generated.
Figure 8 has been redrawn as a deterministic advisory decision matrix with editable/vector source. It lists explicit categories and separately identifies risk tier, empirical support and the proposed operator-review action. The caption states that the mapping is conceptual, was not tested with operators or actuators, and cannot override independent physical constraints. It no longer implies that a supported state grants validated automation authority.
Revision location: Figure 8 and caption; Sections 2.2 and 4.5; figure source files in the revision archive. Status: Addressed.
R1.24  Results, Discussion and Conclusions, comment 6
Comment: Redo Figure 10 because directly comparing PR AUC across rough-to-smooth and smooth-to-rough transfer can confuse readers given six versus twelve events.
We replaced the pooled AP–Brier display with separate panels for the target groups, identifying six smooth events from one specimen and twelve rough events from two specimens. The figure reports observed interception counts and early alert exposure, including Hold and both reference policies; it does not visually rank the directions by discrimination. AP remains descriptive in Table 5.
The revised methods also clarify that identical 180-row windows produce equal label prevalence between directions at a fixed horizon, whereas prevalence differs across the 60, 30 and 10 s horizons. Unequal event and specimen replication affects uncertainty and representativeness even when prevalence is equal. No cross-direction superiority test or population comparison is claimed.
Revision location: Figure 10 and caption; Sections 3.8 and 4.5; Table 5. Status: Addressed.
R1.25  Results, Discussion and Conclusions, comment 7
Comment: There is no validation on real data; “Digital Twin Ready” suggests a maturity requiring real-data validation, so the title should be changed.
We adopted the reviewer’s proposed remedy by changing the title to “Laboratory Evaluation of Mechanics Informed Cross Fault Risk Indicators for Injection Decision Support” and removing the digital-twin-readiness claim. The source measurements are real laboratory observations. We additionally evaluated an independently archived slow-pressurization specimen, with six stage peaks, using both unchanged source policies. The new results reveal calibration limitations rather than operational maturity.\n\nThis addresses the title and maturity concern through the explicitly offered laboratory scope. It does not constitute field validation. We inspected public field-data candidates, including Mont Terri, and documented missing equivalent dynamic stress/slip inputs, clock reconciliation and event-definition requirements. Prospective field and operator evaluation remains future work; no readiness level is assigned.
Revision location: Title; Abstract; Sections 3.10, 4.7, 5.5, 5.7 and 6; Supplementary Section S11; field-data compatibility audit. Status: Addressed.
Reviewer 2
R2.1  Event definition and labels
Comment: Explain the peak-slip identification rule, manual annotation, multiple peaks, independence of pressure stages and exclusion of long quiet periods from false-alarm evaluation.
We now provide an executable reconstruction from the raw measurements rather than relying on ambiguous archived stage labels. On the original 1 Hz grid, a trailing five-sample mean pump rate above 0.1 mL/min for at least five consecutive samples identifies physical injection ramps. The first qualifying sample defines the ramp start. The earliest maximum slip velocity between that start and the next ramp start defines the event peak; the final search ends at the record end. The current reconstruction adds or removes no event manually and reproduces all 18 original peaks, all 3240 labels and the seven model inputs (maximum feature error below 1.5×10⁻¹¹). All nine combinations of thresholds 0.05/0.10/0.20 mL/min and persistence 3/5/10 s at five-sample smoothing retain the same peaks.\n\nThe audit found that archived injection_stage/inj_start_index values are midpoint partitions between selected peaks, not physical ramp onsets. The revised method and Supplementary Table S2 replace their use with the measured pump-based search intervals. Historical manual decisions are unknown, and the reconstruction was developed during revision with the frozen catalogue available; we do not claim to have recovered the original authoring history or created an independent validation set. Tied velocity maxima use the earliest index. Low-threshold unsmoothed variants only fragment existing stages into overlapping peaks, which are not counted as new events.\n\nLabels use lead 180,…,1, with yₕ=1 exactly for 0<lead≤h; the peak row is excluded. Every evaluated row has a complete 30-sample history. Pressure stages share specimen history and are repeated observations. We also evaluate actual interstage exposure and add fixed post-peak exclusions of 180/210/240 s that do not depend on historical event-end labels. Both analyses are limited laboratory exposure checks, not months of field surveillance.
Revision location: Sections 3.2–3.3, 3.9 and 4.6; Supplementary Tables S2, S9 and S11; raw-to-event reconstruction and sensitivity scripts. Status: Addressed.
R2.2  Thresholds and persistence
Comment: Provide stronger justification for the thresholds and persistence rules defining Watch, Mitigate and Hold.
We explain the rules as laboratory design settings and separate their intended ordering from any claim of operational optimality. Quantiles 0.90, 0.95 and 0.99 require increasingly extreme source-negative scores before persistence; the 5, 3 and 2 s runs trade suppression of brief crossings against delay at progressively shorter horizons. The implementation is explicit: scores at or above threshold accumulate a consecutive run, and a below-threshold score resets the indicator; there is no latching. The original event analysis resets runs at the event-window boundary.
We added a source-calibrated grid of all three quantiles crossed with persistence of 1, 2, 3 and 5 s for each horizon and transfer direction. The 72 configurations show that short-horizon rough-to-smooth detection is sensitive to design choices, while reverse-transfer detection can remain unchanged despite substantial variation in early alert activity. Exact thresholds and all outcomes are supplied. The original settings were retained; target outcomes did not select a new optimum.
Revision location: Sections 3.5 and 4.2; Table 4; Supplementary Tables S1 and S4; exact_thresholds.csv and policy_grid.csv. Status: Addressed.
R2.3  Mahalanobis gate robustness
Comment: Strengthen gate robustness given seven dimensions, few event-level observations and strong autocorrelation; report variable contributions, source-distance distributions and event-level cutoff variability; compare alternatives such as one-class SVM, robust density, conformal-style support or leave-one-specimen-out gating.
We added source and target distance distributions, each held-out event’s upper-tail percentiles, signed coordinate contributions to D², complete recalibration after deleting each source event, and source thinning at 5, 10 and 30 s. The signed decomposition cⱼ=δⱼ(Pδ)ⱼ sums exactly to D²; terms may be negative because of covariance cross-terms and are not causal importance scores. Median within-event lag-one distance correlations are 0.947 for rough events and 0.963 for smooth events, so row counts are explicitly not treated as independent calibration samples.
We also implemented source-calibrated diagonal robust-distance and RBF one-class SVM comparisons, plus two-fold leave-one-rough-specimen-out calibration. At q99, final-30 s support is 91.7%/36.4% for Mahalanobis, 100.0%/63.1% for diagonal robust distance, and 97.2%/31.9% for one-class SVM (rough-to-smooth/smooth-to-rough). Crucially, deleting one smooth-source event and fully recalibrating gives a much wider 19.7%–88.6% range. The manuscript now reports this instability and weakens the interpretation accordingly.
A leave-one-smooth-specimen-out analysis is impossible with only one smooth specimen. We did not add a formal conformal coverage claim because the available stage and second-level observations do not establish the required independent/exchangeable calibration units. The additional estimators are contrasting exploratory diagnostics, not tuned method rankings or new validation datasets. Complete code, scores, cutoffs and environment metadata are supplied.
Revision location: Section 3.7; Section 4.4; Sections 5.2 and 5.7; Supplementary Section S5, Tables S5–S8 and Figures S1–S3; gate_*.csv. Status: Addressed.
R2.4  Interpretation of burden and duty
Comment: Use more cautious operational interpretation of event-window burden and duty; state in Table 5 and Figures 5 and 10 that these are not months-long quiet-operation false-alarm burden, and if possible add non-event exposure or synthetic quiet-period tests.
We revised the metric definitions, discussion and specified captions to state that burden and duty describe alert activity within selected 180 s windows. They do not measure executed interventions, throughput loss, downtime or field false-alarm burden. Zero event-window burden is not interpreted as zero activity elsewhere.
We additionally evaluated recovered continuous laboratory records with the original source models, calibration, thresholds, persistence settings and support cutoff frozen. Excluding each principal event from 180 s before its peak through its catalogued post-slip endpoint plus a 30 s feature-memory washout leaves 2,097 s of smooth-target exposure and 3,806 s of rough-target exposure. Watch is active for 6 s (0.29%; one episode) and 61 s (1.60%; five episodes), respectively; Mitigate for 0 and 34 s (0.89%; three episodes); and Hold for 0 s in both. All alert seconds pass the original support gate.
We report 0 and 60 s washout sensitivities because the results depend materially on recovery exclusion. These intervals are archived as interstage locked and have low pump rates, but were not independently certified free of minor slip or AE activity. Approximately 1.6 h of retrospective laboratory exposure cannot establish a field false-alarm rate or replace months of prospective monitoring. No synthetic data are represented as empirical validation.\n\nA further sensitivity analysis removes dependence on the archived post_end annotation by excluding fixed intervals through 180, 210 or 240 s after each peak. At 210 s, eligible exposure is 1430 s for the smooth target and 3028 s for the rough targets. Smooth policies have no activity; rough Watch and Mitigate are active for 33 and 16 s (1.09% and 0.53%). All three exclusion durations are reported, without interpreting one as a validated recovery time (Supplementary Table S11).
Revision location: Sections 3.8–3.9, 4.6 and 5.7; Table 5; Figures 5 and 10; Supplementary Table S9; interstage exposure files. Status: Addressed.
R2.5  Reproducibility
Comment: Provide an actionable reproducibility package with anonymous repository link, version, random seed, software environment and exact threshold files.
The accompanying archive is now Systems_4601389_Revision_Analysis_v3_20260929.zip (version 3.0.0). It includes original frozen inputs and outputs, a raw-pump-to-event reconstruction, exact feature/label verification, all sensitivity analyses, the independent SC2 evaluation, source data with attribution, exact source thresholds, environment metadata, seeds, citation metadata and a complete SHA-256 inventory. A standard-library verifier checks missing, changed and unexpected files. The original bootstrap and revision estimator seed is 20260914; gate calculations are deterministic. Both original and revision environments are distinguished, including the previously omitted SciPy dependency.\n\nThe original dataset DOI and the independent Figshare versioned DOI are verified. The new code archive includes repository deposit metadata and CITATION.cff, but an authenticated repository connection is still required to create its own reviewer-access link. We do not reuse a source-data DOI as an identifier for the revision package. The repository-link part of this request remains pending; the complete versioned archive is available for journal upload. Original annotation intent is unknown, but raw-to-catalogue computational reproducibility no longer depends on the midpoint stage partitions.
Revision location: Data Availability Statement; Supplementary Sections S2, S8 and S11; version 3 archive README, CITATION.cff, repository metadata, SHA256SUMS and verifier. Status: Partly addressed.
R2.6  Formatting and evidence status
Comment: Review formatting, wording and journal compliance; correct Equation (2), remove duplicate Table 5 title, and distinguish conceptual Figures 1 and 8 and Table 10 from implemented and empirically evaluated artifacts.
We corrected Equation (2)’s extra closing parenthesis, explicitly identified its scaled input, and retained the equations as editable Word equations. The duplicate Table 5 caption has been removed. We reviewed wording, definitions, reference attribution, statistical precision and consistency between methods, results and conclusions. The audit also identified substantive corrections to interpolation and acoustic-feature definitions, which are now disclosed rather than treated as cosmetic edits.
Figures 1 and 8 were redrawn with deterministic sources and explicitly identify the proposed interface and action mapping. Table 10 is labelled a future development roadmap. Tables 1 and 7 distinguish implemented offline code and retrospective verification from untested controls, live ingestion, physical fallback, operator effects and network integration. The manuscript no longer equates conceptual requirements or statistical support with demonstrated operational safety or readiness.
Revision location: Equations (1)–(5); Sections 2–3; Figures 1 and 8; Tables 1, 5, 7 and 10; marked manuscript. Status: Addressed.
Editorial requirements
E-I  Ensure all references are relevant to the content of the manuscript.
We reviewed the bibliography against the statements it supports. CCUS network and digital-twin papers now support background and proposed future integration only; they are not presented as evidence of network benefits or deployment readiness. We clarified the original experimental/data provenance and the status of the related unpublished manuscript. We corrected the geothermal article link and added directly relevant sources for the logistic link, calibration hierarchy, Mahalanobis distance, Brier score, Wilson interval and the one-class SVM actually implemented in the revision. The authors’ state, threshold and coverage definitions are distinguished from established statistical formulas.
Location: Introduction; Sections 3.3–3.8 and 5.3–5.6; References.
E-II  Highlight revisions so editors and reviewers can see the changes.
The marked manuscript highlights revised and added text in yellow, including corrected methods, captions, conclusions and reference changes. Replacement figures are identified by highlighted captions. A clean version is also provided for reading; the marked version is the revision record.
Location: Marked manuscript and clean manuscript.
E-III  Provide a cover letter responding to reviewers and explaining the revisions point by point.
The accompanying cover letter describes the change in scope and the additional analyses. This response addresses all 25 numbered comments in the first report and all six comments in the second report, identifies the revised locations, and distinguishes completed revisions from requests that remain only partly addressed or unfulfilled.
Location: Cover letter; this point-by-point response.
E-IV  Critically assess reviewer-recommended references and include only those that enhance the manuscript.
Neither report names a specific bibliographic reference to add. The second reviewer proposes classes of support estimators. We implemented a one-class SVM comparison and cite its original methodological source. Other additions address actual equations or evaluation methods. We have not added citations merely to increase the bibliography or treated a citation as a substitute for the requested analyses.
Location: Section 3.7; References; supplementary gate-robustness analysis.
E-V  Explain any comments that cannot be addressed in an appeal.
We explain the remaining unfulfilled requests explicitly. A newly evaluated independent specimen increases the evidence to 24 trajectories across four specimens, below the requested doubling to 36. Field validation remains absent. Public candidate archives were screened, including actual channel and timing inspection, and cannot be represented as compatible validation without additional reconstruction. A repository connection is still required to create a review-access link. The original-event definition is now reproducible directly from raw pump measurements; unknown historical annotation intent is retained as a provenance limitation. We request assessment within the revised laboratory scope while acknowledging that the editor may require further experiments.
Location: Responses R1.19, R1.25 and R2.5; Sections 3.10, 4.7 and 5.7; Data Availability Statement.
Explanation of requests that remain unfulfilled
Requested doubling of the experimental sample
Independent SC2 data add six stage peaks from one new specimen, bringing the total to 24 trajectories across four specimens. This remains below the requested 36, and the original rough/smooth cohort is reported separately. Further independent, compatible experiments are required.
Prospective field validation
No compatible field test of the unchanged seven-feature model has been completed. Public candidates were inspected and their missing channel, clock, calibration and event-definition requirements documented. The reviewer’s alternative title/scope remedy is implemented, but no field-readiness claim is made.
Repository deposit link
Version 3 analysis files, citations, exact thresholds, environment records, checksums and deposit metadata are prepared. A connected repository account is required to create the actual review-access URL; this remains pending.
The new SC2 evaluation supplies independent experimental evidence, while the sensitivity analyses reuse the original specimens. The combined evidence remains below the requested sample doubling and does not establish field performance. The revised title and conclusions reflect these limits. A repository deposit link also remains necessary; no link or DOI for the new code is invented.