Request gate
Declare unresolved signal properties, receive compatible tools, commit to a complete batch, and execute. The host checks coverage and prevents revisiting measured properties.
Declare → Offer → Commit → Execute
Agentic4PDM
Host-Governed Evidence Admissibility
for Industrial Fault Diagnosis
1FPT Corporation2Independent Researcher3Aalborg University
When a diagnosis remains ambiguous, what should an agent investigate—and how much should each result influence its decision? Agentic4PDM combines case-specific investigation with a deterministic host that governs how evidence is acquired and used.
01 Compare the signal with a labeled reference from every class.
02 Select measurements for the ambiguity of the current case.
03 Check evidence roles before proposing a diagnosis.
48Signal properties
25Default tools
3Host checks
ADAPTIVE INVESTIGATION
Agent reasoning Host governance
Video demonstration
See Agentic4PDM in action
A three-minute narrated overview: the problem, the idea, how the agent reaches a decision under CCGA, and the results.
Narrated overview · 3:11 · up to 1080p · Download MP4 ↓
The framework
Freedom to investigate.
Discipline in evidence use.
LLM agents identify the questions that remain unresolved. Contract-checked gated adjudication (CCGA) controls which measurements may run, how results may support a class, and whether the proposed diagnosis is admitted.
Packet & influence gate
Test each tool on known-label controls. Usable results become corroborative if every control is recovered, or advisory otherwise. Failed or ambiguous packets are excluded.
Checked support or advisory information
Finalization
Admit a proposal only if its class is in the initial candidate set and a quantitative measurement round is complete. Otherwise, return it for revision within the step budget.
A class + its admission status
Standard mode
Pretrained embeddings and labeled references, plus SVM and XGBoost predictions as hypotheses to investigate. The main results use this mode.
Training-free mode
Pretrained representations and labeled references, with no task-specific fitted diagnostic classifier. Evaluated separately in the appendix.
Read the full abstract
We study industrial fault diagnosis as a classification task that assigns sensor time-series windows to predefined health states or fault types. Scarce labels and changing operating conditions make repeated training of specialized classifiers costly. Yet structures such as periodicity, trends, and transients can remain diagnostically informative and be examined through reusable analytical tools. LLM agents can select and interpret these tools, but fixed rules or training-derived guidance may not resolve each signal’s ambiguity, while unconstrained interpretation may overstate weak or conflicting results. We introduce Agentic4PDM, a framework that separates adaptive investigation from control over evidence use. LLM agents compare each signal with labeled references from every candidate class, identify unresolved distinctions, and propose follow-up measurements. A deterministic software controller, called the host, enforces contract-checked gated adjudication (CCGA): it checks requests, requires commitment to a complete tool batch before results are observed, and assigns evidence roles based on tool checks against labeled references. These roles distinguish corroborative support from advisory information. The Judge Agent proposes a diagnosis, while the controller checks protocol compliance and records the investigation. Agentic4PDM supports training-free diagnosis using pretrained representations and labeled references without task-specific diagnostic models, and a standard setting that additionally investigates trained-model predictions. Experiments on selected bearing and motor diagnosis challenges and an unfiltered industrial test set supplied by our customer show gains in aggregate macro F1 over the evaluated diagnostic and agentic baselines. Ablations further support the value of the governed workflow.
Empirical results
Three complementary tests
Agentic4PDM achieves the highest aggregate macro-F1 among the evaluated methods across public benchmark challenges and the full external customer test set.
Regime coverage · GRS
59.20%Unweighted public-dataset averageHard cases · BFH
47.47%Unweighted public-dataset averageCustomer test · Bel
73.95%All 330 test samplesMacro-F1 (%) · standard mode · values from the paper
2,048 observations / window50 training samples / classQwen3.6-27B for all agentic methods
Explore the detailed benchmark results
A closer look
What the investigation changes
Beyond aggregate scores, the paper examines prediction revisions, sample-specific plans, the contribution of evidence-handling controls, and whether simpler uses of the same evidence can match the workflow.
01 / Beyond fitted-model predictions
Agreement is a starting point
When SVM and XGBoost initially agree, Agentic4PDM can still correct their shared errors. On PU and MCC5 GRS, the consensus analysis reports a net accuracy gain of 7.97 percentage points, after accounting for new errors introduced.
Fitted-model predictions serve as hypotheses for further investigation.
02 / Sample-specific investigation
Different cases, different questions
Agentic4PDM produces 31 distinct plans on PU and 50 on MCC5, compared with one fixed plan for FETA and limited variation in ZARA.
Evidence acquisition varies with the case. This observation does not by itself show that plan diversity causes the performance gains.
03 / Component ablation
Both evidence access
and governance matter
On PU GRS, replacing class-balanced retrieval or removing both governance gates substantially lowers performance. All variants use the same evaluation setting.
| Variant | Accuracy | Macro-F1 | Δ F1 |
|---|---|---|---|
| Full workflow | 47.24 | 38.54 | — |
| Global top-K retrieval | 37.20 | 29.89 | −8.65 |
| No Analyst reports | 46.35 | 37.72 | −0.82 |
| Both gates removed | 36.31 | 30.53 | −8.01 |
| No request gate | 39.96 | 33.35 | −5.19 |
| No packet & influence gate | 42.60 | 36.70 | −1.84 |
| No finalization | 45.40 | 38.02 | −0.52 |
Scores are percentages. Δ F1 is the change in percentage points from the full workflow.
04 / Simpler alternatives
Simple shortcuts do not reproduce the workflow
Could the same result come from voting over the evidence the Judge already has, or from one strong measurement? The paper tests both. Neither matches the full workflow.
Flat pluralityAgentic4PDM
MCC5+5.83
PU+9.45
UO+2.38
Coverage: 77.7% of samples on MCC5, 84.1% on PU, 89.5% on UO (samples where the vote gives a prediction rather than a tie).
| Tool | MCC5 | PU | UO | Mean |
|---|---|---|---|---|
| Windowed profilebest on MCC5 | 36.15 | 35.22 | 30.17 | 33.85 |
| Crossing-run statisticsbest on PU | 22.89 | 37.94 | 16.73 | 25.85 |
| Distribution KSbest on UO | 31.78 | 32.78 | 60.02 | 41.53 |
| Agentic4PDMfull workflow | 52.93 | 38.54 | 86.12 | 59.20 |
Outlined cell: the dataset where the tool was selected. Crossing-run statistics reaches 37.94 on PU but falls to 22.89 on MCC5 (−15.05) and 16.73 on UO (−21.21).
Neither fixed aggregation of the initial evidence nor one strong measurement reproduces the full workflow. This is consistent with the value of choosing and combining measurements for the current case.
Resource use & evaluation scope
Agentic4PDM is intended for on-demand investigation of difficult cases. The paper reports higher token use alongside stronger aggregate diagnostic performance.
| Method | Tokens | Time (s) | LLM calls | Est. API charge |
|---|---|---|---|---|
| ZARA | 7.83k | 98.20 | 4.0 | $0.0129 |
| FETA | 72.67k | 1,078.59 | 8.0 | $0.0575 |
| Agentic4PDM | 102.36k | 1,645.79 | 17.4 | $0.0989 |
Timer scopes differ between implementations. Charges use the paper’s assumptions: $0.60 per million input tokens and $3.60 per million output tokens. These are estimates, not observed bills or total deployment costs.
Training-free uses labeled references. It omits fitted diagnostic classifiers while retaining pretrained representations.
Admission is a protocol flag. Every sample receives a class; the flag is not a probability of correctness.
The evaluation has a defined scope. GRS/BFH are selected challenges. Bel is one full external customer test set, not a fleet-wide deployment study.
Authors
Research team
FPT Corporation · Independent Researcher · Aalborg University
BibTeX citation
@misc{nguyen2026agentic4pdmhostgovernedevidenceadmissibility,
title={Agentic4PDM: Host-Governed Evidence Admissibility for Industrial Fault Diagnosis},
author={Nam V. Nguyen and Hoang Van Truong and Hieu D. D. and Duy Khuong Nguyen and Quang Pham and Tung Kieu},
year={2026},
eprint={2610.17786},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2610.17786},
}arXiv preprint arXiv:2610.17786 (v1, 13 October 2026), cs.AI.