Agentic4PDM: Host-Governed Evidence Admissibility for Industrial Fault Diagnosis

• Diagnosis must adapt when operating conditions shift, even though labeled fault examples needed to retrain and revalidate classifiers are scarce.Source: §1 · p. 2
• Agentic4PDM uses LLM agents to compare signals with labeled references from every candidate class, identify unresolved distinctions, and propose follow-up measurements under Host-governed CCGA.Source: §1 · p. 2
• CCGA separates adaptive investigation from evidentiary authority: the Host checks requests, assigns result roles using reference preflight, and checks final proposals.Source: §1 · pp. 2–3
• Agentic4PDM achieves the strongest overall macro-F1 performance across bearing and motor tasks, with the highest three-dataset averages on BFH and GRS and the best result on the customer test set.Source: §1 · p. 3
• Agentic4PDM evaluates complementary bearing and motor settings for predictive performance and governed-investigation behavior.Source: §1 · p. 3

• A fixed classifier head predicts from learned boundaries without acquiring further evidence for ambiguous inputs, whereas Agentic4PDM uses labeled references and tools for investigation.Source: §2 · p. 3
• Prior agents use offline rules, training-derived features, or self-assessed confidence for evidence selection or aggregation, which may not resolve case-specific ambiguity.The cited approaches include ARGOS [23], ZARA [39], and FETA [55].Source: §2 · p. 3
• Agentic4PDM requires the agent to declare unresolved properties and commit to a covering tool batch before results are revealed.Source: §2 · p. 3
• A deterministic Host checks measurements against labeled references retrieved for the case, independently determining each result’s permitted influence.Source: §2 · pp. 3–4
• Unlike governance focused on permissions, safety, security, or argument acceptability, CCGA governs adaptive diagnostic evidence acquisition and use through final class admission.Related governance examples include, [11] [14] [57], and.Source: §2 · p. 4

3.1 Framework Overview

• The framework has an evidence-construction phase followed by CCGA, with the Judge choosing investigations and the Host controlling measurement execution, evidence influence, and admission.Source: §3.1 · p. 4

3.2 Problem Definition

• The system takes a query, its class-balanced reference bank, and read-only context, and returns a class plus an action-and-result record that can mark the class unadmitted.Source: §3.2 · p. 4

3.3 Agents and Host

• Separate LLM roles select smoothing and channels, compare stable and deviant signal regions, and synthesize evidence; standard mode also supplies SVM and XGBoost predictions, while training-free mode omits them.Source: §3.3 · p. 5

3.4 Query-Adaptive Evidence Construction

• The Normal and Abnormal Analysts compare stable and deviant regions, respectively, against class references to produce complementary candidate reports with supporting and contradicting observations.Source: §3.4 · p. 5

3.5 CCGA: Host-Governed Adjudication

• CCGA begins with frozen Analyst evidence and governs the Judge’s subsequent investigation through request, packet-and-influence, and finalization checks.Source: §3.3 · p. 5, §3.5 · p. 6

3.5.1 Request Gate

• The request gate requires the Judge to declare unresolved measurable properties, select compatible offered tools, and commit to a complete batch before execution.The Judge must support requests with prior evidence; the Host checks batch conditions before running tools.Source: §3.5.1 · p. 6

3.5.2 Packet and Influence Gate: Controlling What Each Result Can Support

• A reference preflight assigns usable tool results corroborative or advisory roles: corroborative results provide checked support, while advisory results may guide hypotheses but should not alone overturn another result.The Judge reports agreement, conflict, or incomparable results and cannot stop while valid results conflict and tool budget remains.Source: §3.5.2 · p. 7

3.5.3 Finalization

• The Host admits a proposal only if it remains within the initial candidate set and at least one round of quantitative measurement is complete; otherwise it reports the final-step class as not admitted.A not-admitted flag indicates failed protocol checks, not an estimate of diagnostic accuracy.Source: §3.5.3 · p. 7

4.1 Experimental Protocol

• PU, UO, and MCC5 evaluations use 50 training samples per class and selected GRS or BFH windows; Bel evaluates all 330 customer test samples without selection.Signals are segmented into non-overlapping windows of 2048 observations; methods use Qwen3.6-27B, and performance is measured by accuracy and macro F1.Source: §4.1 · p. 8

4.2 Main Results

• Agentic4PDM has the highest average performance across tests, beating the strongest baseline +7.5% on GRS, ZARA by +12% and FETA by +19.53% on BFH, and the strongest baseline by more than +12% on Bel.The reported run-to-run standard deviations are at most approximately 2.3 points on BFH and 1.2 points on GRS.Source: §4.2 · pp. 8–9, Figure 2 · p. 9

4.3 Agent Behaviour Analysis

• On PU and MCC5 GRS, Agentic4PDM corrects 19.9 points of SVM errors and 19.3 points of XGBoost errors, while also introducing new errors.When SVM and XGBoost initially agree, it corrects 17.2 points of shared errors and introduces 9.2 points of new errors, yielding a net gain of 7.97 points.Source: §4.3 · p. 10
• Flat plurality trails Agentic4PDM by 5.83 macro-F1 points on MCC5 and 9.45 on PU; the best single tool also transfers poorly across datasets.Crossing-run Statistics leads on PU at 37.94% macro-F1 but scores 22.89% on MCC5 and 16.73% on UO; Agentic4PDM outperforms each dataset's best single tool.Source: §4.3 · p. 10
• Agentic4PDM produces 31 distinct plans on PU and 50 on MCC5, compared with FETA's fixed plan and ZARA's 3 and 7 plans, respectively.Its most frequent plan covers 47% of PU and 51% of MCC5 samples; the analysis supports adaptive capacity but does not establish that plan diversity causes performance gains.Source: §4.3 · pp. 10–11

4.4 Ablation Study

• Global top-K retrieval (A2) drops accuracy 10.04 points and macro-F1 8.65 points; removing Analyst reports (A3) drops them 0.89 and 0.82 points, respectively.Class-balanced retrieval keeps every class represented by references, while Analyst reports retain candidate classes, observations, and contradictions for the Judge.Source: §4.4 · p. 11
• Removing both CCGA gates drops accuracy 10.93 points and macro-F1 8.01 points; removing finalization drops accuracy 1.84 points and macro-F1 0.52 points.Without the request gate, accuracy falls 7.28 points and macro-F1 5.19 points; without the packet and influence gate, they fall 4.64 and 1.84 points, respectively. The gates play complementary roles.Source: §4.4 · pp. 11–12

4.5 Computational Cost Analysis

• Agentic4PDM uses 102.36k tokens per query and costs $0.0989, versus FETA's 72.67k and $0.0575 and ZARA's 7.83k and $0.0129; it targets on-demand, not real-time, diagnosis.Relative to FETA, the additional estimated charge is $0.0414 per query, alongside macro-F1 gains of 18.82 points on GRS, 19.53 on BFH, and 22.47 on Bel.Source: Figure 5 · p. 12, §4.5 · p. 12

5 Conclusion

• CCGA lets agents select measurements for unresolved distinctions while the host controls each result’s evidentiary influence through reference checks rather than LLM confidence.Unchecked outputs receive less evidentiary authority than corroborative results.Source: §5 · p. 12
• Across two bearing benchmarks, an induction-motor benchmark, and customer data, Agentic4PDM achieves the strongest aggregate diagnostic performance among evaluated baselines.The analyses and ablations also support the value of the governed investigation and CCGA controls.Source: §5 · p. 12
• The paper proposes consolidating tools and broadening their functionality to reduce selection and context costs, and adding follow-up measurement or inspection suggestions for unresolved cases.The current system records whether a diagnosis is admitted but does not yet recommend follow-up actions.Source: §5 · p. 12

A Context Management for Multi-Turn Diagnosis

• Image retention keeps recent images active, permits a limited extension with recent earlier images, and archives older images with descriptions and source records.The examples’ nimg = 3 and next = 2 values are illustrative, not experimental settings.Source: A Context Management for Multi-Turn… · p. 18
• Conversation summarization retains the opening and latest turns and replaces intervening turns with a structured summary of key observations and decisions.The opening preserves task context, while recent turns preserve the reasoning leading to the next action.Source: A Context Management for Multi-Turn… · p. 18
• The supplement presents image retention and conversation summarization as separate mechanisms for reducing historical material in each request while retaining diagnostic context.Their usefulness depends on image descriptions and summaries preserving information needed for later decisions.Source: A Context Management for Multi-Turn… · pp. 18–19

B Preventing Label Leakage through Client--Server Data Isolation

• Neutral upload and workspace paths, plus removal of the query label column, reduce exposure of source fault categories and specimen names during inference.Labeled references and dataset context remain permitted inputs, while the query’s true label is kept separately for post-inference scoring.Source: B Preventing Label Leakage through… · pp. 19–20
• The isolation addresses accidental disclosure through source paths and query columns within the inference data flow, not filesystem access control or cryptographic confidentiality.Evaluation metadata and diagnostic logs remain with the trusted Host.Source: B Preventing Label Leakage through… · p. 20

C.1 Performance on BFH, GRS, and Bel

• Table 2 compares standard Agentic4PDM with deep-learning, time-series foundation-model, and agentic baselines across public BFH/GRS datasets and the full customer Bel test set.It reports accuracy and macro-F1, with an unweighted public-dataset average; the supplied caption does not provide the table’s cell values.Source: §C.1 · p. 20, Table 2 · p. 20

C.2 Training-Free Comparison with ZARA and FETA

• Training-free Agentic4PDM outperforms ZARA and FETA on all four challenge-average measures, despite omitting fitted-model predictions.Standard mode improves mean GRS accuracy and macro-F1 by 2.23 and 1.76 percentage points and mean BFH accuracy by 0.66 points, while training-free mode has higher mean BFH macro-F1 (48.38% versus 47.47%).Source: Table 3 · p. 20, §C.2 · p. 21

C.3 Performance Across Three Runs

• Across three observed executions, standard Agentic4PDM has a smaller spread in average BFH macro-F1 than ZARA, while average GRS scores vary little for all three methods.The comparisons do not establish variability over a larger number of runs.Source: §C.3 · p. 21, Figure 6 · p. 21

C.4 Agentic4PDM versus Ensembles of Its Evidence Sources

• On matched covered samples, Agentic4PDM exceeds flat source plurality by 5.83 macro-F1 points on MCC5, 9.45 on PU, and 2.38 on UO.The largest difference is on PU: 37.90% for Agentic4PDM versus 28.45% for plurality; the comparison does not isolate individual workflow mechanisms.Source: §C.4 · p. 22, Table 4 · p. 22

C.5 Evidence Guidance

• Corroborative packets occur less often than advisory packets on MCC5, PU, and UO, but have higher correctness on MCC5 and UO and nearly equal correctness on PU.Corroborative versus advisory coverage is 30.30% versus 83.68% on MCC5, 64.35% versus 86.53% on PU, and 68.15% versus 96.85% on UO; correctness is 55.56% versus 34.26% on MCC5, 48.06% versus 47.86% on PU, and 82.24% versus 64.64% on UO.Source: §C.5 · p. 23, Table 5 · p. 23, Table 6 · p. 23

D Can the Best Single Tool on Each Dataset Replace Agentic4PDM?

• A tool comparison predicts a sample only from unique channel winners and a unique largest class-vote count; ties or no valid vote become unknown.Unknown predictions contribute false negatives to macro-F1, and the offline comparison does not override each ToolCard’s inference permissions.Source: D Can the Best Single Tool on Each… · p. 23
• 59.20% macro-F1: the full workflow exceeds each dataset’s best default by 16.78, 0.60, and 26.10 percentage points on MCC5, PU, and UO, respectively.Distribution KS has the highest equal-dataset mean among the 25 defaults, at 41.53%.Source: D Can the Best Single Tool on Each… · p. 24, Table 7 · p. 24
• The winning default depends on the dataset: crossing-run statistics ranks first on PU but twentieth on UO and fifteenth on MCC5.The windowed amplitude profile ranks first on MCC5, fourth on PU, and fifth on UO.Source: D Can the Best Single Tool on Each… · p. 24
• The audit reuses saved predictions and historical agent implementations, so its rows are configuration-level comparisons rather than a controlled ablation of adaptive tool selection.Tool selection used the evaluated samples, not a separate selection split.Source: D Can the Best Single Tool on Each… · pp. 23–24

E Backbone Substitution: What the Agent Adds to a Representation

• The archived C2S experiment replaces the frozen retrieval encoder and embedding cache; standard-mode heads are fitted to C2S embeddings, while neither backbone supplies fitted-model predictions in TF.The LLM, decoding, prompts, budgets, thresholds, and BFH/GRS manifests are shared, so these are different backbones rather than repeat executions.Source: E Backbone Substitution: What the… · p. 24, Figure 7 · p. 25
• Figure 8 reports ACC and macro-F1 as percentages and shows TF-minus-standard differences for unweighted three-dataset averages on the same frozen BFH and GRS windows.Figure 7 identifies the substituted encoder, cache, retrieval representation, and standard fitted heads; other settings and windows are shared.Source: Figure 8 · p. 25, Figure 7 · p. 25
• Agentic4PDM exceeds both fitted heads on both backbones and both challenges, though its C2S gain over the stronger head is larger while its absolute macro-F1 is lower.This separates improvement over a given representation from the representation’s overall quality.Source: E Backbone Substitution: What the… · p. 26, Figure 9 · p. 26
• The comparison does not isolate the causal effect of host gates or quantify run-to-run variability.The displayed agent-versus-head comparisons are descriptive.Source: E Backbone Substitution: What the… · p. 26

F Empirical Time and Token Use

• Agentic4PDM records 1,645.79 s, 102.36k tokens, and 17.4 LLM calls per query, versus FETA’s 1,078.59 s and ZARA’s 98.20 s and four calls.FETA’s retrieval accounts for 1,044.51 s of its recorded time; the methods’ timer scopes differ, so these are descriptive implementation measurements rather than a controlled latency benchmark.Source: Table 8 · p. 26, F Empirical Time and Token Use · p. 27
• 18.82 percentage points on GRS and 19.53 points on BFH: Agentic4PDM’s three-dataset mean macro-F1 gains over FETA accompany an estimated USD 0.0414 additional API charge per query.The estimated charges are USD 0.0989 per query for Agentic4PDM and USD 0.0575 for FETA; Figure 10 estimates token-based API charges, not observed bills or total deployment costs.Source: Figure 10 · p. 27, F Empirical Time and Token Use · p. 28
• On the unfiltered customer Bel test set, Agentic4PDM’s macro-F1 exceeds ZARA’s by 29.70 points, with scores of 73.90% versus 44.20%.ZARA has the lowest estimated API cost, at USD 0.0129 per query; the comparison highlights that lower online cost does not necessarily coincide with higher diagnostic performance.Source: Figure 10 · p. 27, F Empirical Time and Token Use · p. 28

G How consistent is performance across operating conditions?

• Across PU and MCC5, Agentic4PDM has higher mean accuracy than FETA and ZARA with comparatively limited variation across the displayed operating conditions.Its accuracy spans 10 points on PU and 5 on MCC5, compared with ZARA’s 23 and 10 points, respectively; FETA varies relatively little on PU but has substantially lower accuracy on both datasets.Source: G How consistent is performance across… · p. 28, Figure 11 · p. 28

H.1 Datasets, Signal Modalities, and Split Boundaries

• PU, UO, and MCC5 cover three-class bearing or four-class motor diagnosis, while Bel is a separate binary bearing evaluation using customer-supplied data.The datasets include differing fault settings, operating conditions, and sampling-rate availability.Source: §H.1 · p. 29, Table 9 · p. 29

H.1.1 Frozen Manifests and Per-Class Support

• Test samples and labeled references come from different recordings in all four datasets, and the preceding sample used by LBAgents is never part of labeled training.The complete Bel test set has 330 samples: 150 normal and 180 abnormal; Table 10 gives frozen BFH/GRS test-set sizes and per-class counts alongside Bel.Source: §H.1 · p. 29, §H.1.1 · p. 29, Table 10 · p. 30

H.2 Shared Evaluation Protocol and Inventory of Compared Methods

• The comparison inventory includes deep-learning, foundation-model, and agentic families; each of four frozen foundation-model backbones is paired with KNN, SVM, and XGBoost.The non-agent methods produce predictions without adaptive evidence construction or Host-governed adjudication.Source: §H.2 · p. 30, Table 11 · p. 31

H.3 Configuration of Agentic4PDM

• The reported Agentic4PDM configuration uses fixed operating choices, including a GRS geometry target of G=250 and response-stability thresholds of 0.70 and 0.50.These are not learned parameters or claims of optimality.Source: §H.3 · p. 31, Table 13 · p. 32

H.4 Training Settings of the Deep-Learning Baselines

• The deep-learning baselines comprise CNN-TCC, QCNN, and WDCNN, while foundation-model baselines pair four frozen encoders with classical classifier heads.The listed settings are documented in Tables 14 and 15.Source: §H.4 · p. 32, §H.5 · p. 32, Table 14 · p. 33

H.6 Implementation Settings of the Agentic Baselines (FETA and ZARA)

• FETA ranks channels and retrieves DTW exemplars for channel-level reasoning, whereas ZARA uses pairwise feature knowledge, candidate pruning, and decision synthesis.Both implementations use shared evaluation conditions while retaining their distinct evidence workflows.Source: §H.6 · p. 33, Table 16 · p. 34

I Benchmark Construction: How the BFH and GRS Challenge Sets Were Built

• Because full-pool agentic evaluation is costly, the benchmarks use frozen selection registries that exclude the evaluated agent and fixed manifests.The registries are challenge-construction devices, not ensembles or claims that every member is a competitive baseline.Source: I Benchmark Construction: How the BFH… · p. 34

I.1 Challenge Set I --- BFH: A Failure-Severity Stress Test

• BFH selects the hardest samples by frozen-model error, so it can over-sample recurring failure patterns and under-cover other conditions and signal presentations.The challenge answers whether an agent recovers samples that many frozen registry models fail on.Source: §I.1 · p. 35

I.2 Challenge Set II --- GRS: A Representative-Coverage Diversity Test

• GRS groups samples within class and condition by embedding geometry and frozen-model response, then uses stability tiers to represent within-category variation.Its fixed thresholds merge response-signature groups according to per-model prediction stability, not agreement between models.Source: §I.2 · pp. 35–36
• GRS retains one observed representative per nonempty signal/response group, while BFH selects samples with high frozen-registry error to test concentrated failure recovery.Together, the challenges target failure severity and broader case coverage, respectively.Source: I Benchmark Construction: How the BFH… · p. 34, §I.1 · p. 35, §I.2 · pp. 36–37

I.3 Evaluation Scale and Class Composition

• Both challenges provide descriptive margins, not unbiased estimates of accuracy on the full pool or in deployment; BFH selection also uses registry errors, including errors of fitted heads in the comparison.Matched samples enable paired comparisons but do not remove BFH selection bias.Source: §I.3 · p. 37
• GRS class imbalance follows the one-representative-per-group coverage rule; every retained class has at least 175 samples, and the largest within-dataset majority-to-minority ratio is 2.38:1.Macro-F1 is reported alongside accuracy so each class contributes equally.Source: §I.3 · p. 37
• BFH uses 100 samples per dataset to isolate recovery on concentrated failures, while the roughly order-of-magnitude-larger GRS spans classes, conditions, signal geometry, and response regimes.The two challenges are reported separately because neither substitutes for the other.Source: §I.3 · p. 37

J Summary of Notation for the Method and Inference Path

• Table 17 collects notation needed for the protocol and sample-specific inference path, while appendix-only construction symbols are defined locally.BFH and GRS construction notation is defined in Section I.Source: J Summary of Notation for the Method… · p. 38, Table 17 · p. 38

K.1 Retrieval Protocol: Class-Balanced Reference Selection

• Retrieval ranks labeled references within each class by scaled-embedding cosine similarity and retains the highest-scoring examples without filtering by operating condition.With k=1, the reference bank contains one exemplar per class; the method is not a KNN classifier.Source: §K.1 · p. 39

K.2 Request Gate: Property Registry

• The Host maps each class-independent registered signal property to eligible ToolCards; properties constrain which cards may be offered but do not reveal scores or choose for the Judge.The registry contains 48 properties shared across datasets.Source: §K.2 · p. 39, Table 18 · p. 39

K.3 Packet and Influence Gate: From Measurement to Influence Role

• A packet is usable only when the tool succeeds and the highest-priority nonempty channel tier agrees on a winning class; usable packets become corroborative or advisory according to reference preflight.Excluded packets remain in the ledger but cannot support a class.Source: §K.3 · pp. 40–41

K.4 Publication: What the Host Serializes and Reports

• The Host serializes the reported class together with an admitted or not-admitted status, distinguishing decisions that passed protocol checks from those that did not.A non-admitted class is the Judge’s last-step class.Source: §K.4 · p. 41

L Tool Registries: What the Agents May Execute

• The 25 default quantitative ToolCards are authored offline, while the inference-time agent selects from the fixed registry and never writes code.Tools measure named signal properties; reference preflight, rather than authoring, determines whether results count as checked evidence.Source: L Tool Registries: What the Agents May… · pp. 41–42

L.1 General-Purpose Inspection Tools

• The Host validates requested actions for syntax, scope, authorization, and quantitative bindings before dispatch.The Judge’s interface includes seven general tool actions.Source: §L.1 · pp. 42–43, Table 21 · p. 42

L.2 Smoothing Families and Host-Fixed Parameters

• The SAgent selects one label-blind smoothing family per channel from a fixed registry, while the Host chooses parameters from a reviewed grid.LLM-proposed parameter values are not accepted.Source: §L.2 · p. 43, Table 22 · p. 43

L.3 Channel Priority Tiers and Host Guards

• The Host enforces channel-selection safeguards by restoring all channels if the agent keeps only one and restoring diagnostically strong channels under a score threshold.A kept channel scoring at least max(2.5, 0.7 smax) is restored if removed; the strongest-scoring channel or one with a burst train is set to high priority.Source: §L.3 · p. 43

L.4 Quantitative Tools Available to the Judge

• Only the 25 default quantitative ToolCards are available, each with fixed parameters; availability follows selector and request constraints, not invocation frequency.The residual-level card is limited to fault-existence questions.Source: §L.4 · pp. 43–44, Table 24 · p. 44

L.5.1 Tool 1: Threshold-Crossing Run Statistics (T04)

• Threshold-crossing run statistics compare consecutive above- or below-median run durations while discarding absolute positions.Run lengths of 32 samples or more share the final histogram bin.Source: §L.5.1 · p. 45

L.5.2 Tool 2: Envelope-Spectrum Similarity (T07)

• The envelope-spectrum tool compares log-scaled power-spectrum shape of Hilbert-transform amplitude envelopes, capturing modulation and burst repetition with reduced phase dependence.It uses normalized FFT bins and does not identify physical defect frequencies in Hz.Source: §L.5.2 · p. 45

M Prompt Templates for Each Agent Role

• The archived runtime templates use placeholders for injected figures, labels, and paths, while the Host interface, acceptance rules, and required output fields remain part of each contract.The listings reproduce archived runtime role contracts.Source: M Prompt Templates for Each Agent Role · p. 46

M.1 Evidence-Preparation Agents

• Evidence-preparation agents choose smoothing per channel and assign channel priorities, while the Host fixes stable and deviant regions rather than using the historical segmentation agent.The CSAgent keeps uncertain channels at medium and removes only uniform cyclic channels without modulation, deviation, or useful reference context.Source: §M.1 · pp. 46–47

M.2 Analysts and LBAgents

• Analysts compare stable baselines or focused deviations across candidate references; tied top similarity bands are ambiguous and cast no class vote.Lookback agents compare preceding-window context but cannot override current-window comparisons or create a class vote.Source: §M.2 · pp. 48–49

M.3 Judge and Host Contract

• The Judge must resolve material uncertainty with a focused investigation, never finalize alongside a non-final tool call, and cannot rely on blocked, unavailable, or unpreflighted results.It must name the candidates, discriminating question, and observable, then request only the needed comparison, inspection, or registered action.Source: §M.3 · p. 50
• The Host validates JSON and dispatches the requested action, admitting finalize only after authorization, budget, and host finalization checks.Source: §M.3 · p. 50
• If no admissible finalize request arrives before the step limit, the Host publishes the Judge’s last-step class from H1 and marks it as not admitted.Finalize is a Judge request, not something the Host publishes; the governed configuration additionally requires the Host’s finalization check.Source: §M.3 · p. 50
• The Judge declares unresolved properties and commits to a batch; the Host checks offered-card use, property coverage, and that no property was measured in an earlier round.The Host offers cards only for declared properties; failed tools or packets without a winning class are excluded and support no class.Source: §M.3 · p. 50
• The Host admits a proposed class only if it is a starting candidate and at least one measurement round is complete, then records the class, confidence, reasoning, and recoverable evidence/action trace.It does not validate free-text causal claims.Source: §M.3 · p. 51

N End-to-End Trace of One Sample, Stage by Stage

• Host-tuned median_mean_ewm separated noisy vibration traces around +0.05 versus −0.02, while Savitzky–Golay exposed a phase offset in the current channels.The agent selected transform families, but the Host fixed their parameters before comparison; the raw vibration traces were indistinguishable in the displayed view.Source: N End-to-End Trace of One Sample,… · pp. 52–53
• The Host marked 660 of 2048 observations stable and 1312 deviant (normal_ratio 0.32, abnormal_ratio 0.64), and classified recurrence as strong (recurrence_score: 0.4161).It detected recurring impulse clusters and cautioned against dismissing sparse abnormal coverage as a healthy isolated transient.Source: N End-to-End Trace of One Sample,… · p. 54
• Although the agent dropped phase_current_2, the Host’s mask-based scoring identified its prominence z = 15.6 and restored it for downstream comparisons and measurements.The agent ranked vibration_1 high and phase_current_1 medium, while the Host could override channel judgments using segmentation scores.Source: N End-to-End Trace of One Sample,… · p. 54

N.2 — Comparative Reasoning: two Analysts, and the tie that starts everything

• Class 0 was ruled out, but Classes 1 and 2 remained unresolved: the abnormal Analyst ranked Class 2 highest yet withheld a verdict because its vibration match could be coincidental.The two Analysts’ reports left fault type as the live question, with Ht = {1, 2}.Source: §N.2 · pp. 56–57

N.3 — The Request Gate: asking for a measurement without being able to shop for one

• The Host offers matching ToolCards with methods and limitations but no scores, class rankings, or predicted class, so the Judge cannot anticipate results.The Judge requested extrema morphology, roughness across scales, and recurrence match; five cards were offered because three measured roughness.Source: §N.3 · pp. 58–59
• The Host requires the Judge to declare unresolved properties, select a batch covering every property, and commit before any tool runs.Measured properties cannot be requested again, preventing an unwelcome result from being retried under new settings or selectively omitted.Source: §N.3 · p. 60
• All three packets executed validly, but only two had a decided winner: two packets selected Class 2, while the third had conflicting channel winners.Arbitration requires agreement among the highest-weight channels, not a large margin: a tiny separation can be decided, while disagreement yields no winner.Source: §N.3 · p. 61
• Reference preflight passed one ToolCard and failed two; only a passing packet can rise from advisory to corroborative, and one failed control is decisive.The replay tests whether the unchanged tool recovers every known control-class label; under strict_all_controls, two-thirds correct earns no partial credit.Source: §N.3 · pp. 61–62
• The Host reports Class 2, matching the inner_race_fault ground truth, after assigning evidence roles and closing the round when both usable packets name Class 2.An excluded packet has no influence, while an advisory packet may inform a hypothesis; all packets retain provenance in the ledger.Source: §N.3 · pp. 62–63

N.5 — Finalization: publishing a class from the checked record

• The Host publishes Class 2 only after checking candidate-set membership and a completed measurement round, and sets confidence to medium rather than the Judge’s requested high.The published record preserves the candidate set and validation results so the class can be traced to earlier packets and reports.Source: §N.5 · pp. 63–64

O Side-by-Side Comparison: Agentic4PDM, FETA, and ZARA on the Same Sample

• The side-by-side comparison is framed as turn-by-turn records for Paderborn test sample 140691, whose true class is 2, with explanatory Host summaries of eligibility and admission.The comparison includes Agentic4PDM, FETA, and ZARA on the same sample.Source: O Side-by-Side Comparison:… · p. 64

Host

• On the sample, the vibration agent alone names true class 2, but largest-sum fusion outputs class 0 without examining the channel disagreement.The vibration answer loses two votes to one.Source: Host · pp. 64–65
• 40.4%: vibration alone scores above FETA’s fused 35.2% on 906 samples when both current channels disagree with vibration.The current-channel pair wins 355 / 355 times in that disagreement condition.Source: Host · p. 65

ZARA

• ZARA’s stage-2 pruning removes true class 2, leaving classes 1 and 0; no later stage can restore it.The subsequent feature selection uses no number from the sample.Source: ZARA · p. 65

Host

• Before seeing results, the Judge committed to autocorrelation, local extrema, and fractal scaling after ruling out 18 other tool families.The offer kept scores and predicted class hidden; STFT time–frequency was over the cost limit.Source: Host · p. 67
• The selected tools examine repeating bursts, peak density and spacing, and roughness across scales.Roughness comparisons can be affected by finite length and lag support; smoothing can erase weak faults, and multiscale complexity is finite-length sensitive and less directly interpretable.Source: Host · p. 67
• Autocorrelation was corroborative after recovering reference-control labels, while local extrema and fractal scaling were advisory because they failed that check.The Host assigned these evidence roles; the user could read but not change the ratings.Source: Host · p. 67
• The Judge proposed Class 2 with medium confidence despite predictor outputs of Class 0, because neither comparator gave Class 0 a clear vote.All three measurements favored Class 2, and autocorrelation passed its reference check.Source: Host · p. 67
• The Host published Class 2 with medium confidence, matching the true class in this trace.The completed measurement round left Class 2 as a candidate.Source: Host · p. 67
• The true class remained live after the Analyst stage in 89.1% of 906 samples (ZARA: 68.0%).Source: Host · p. 67