Pith. sign in

REVIEW 3 major objections 6 minor 37 references

High reward does not prove an agent learned the task's state

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 07:13 UTC pith:MPMKXRAK

load-bearing objection A genuinely useful measurement paper: it separates reward success from latent-state learning with an exact oracle and a held-out structural predictor, and it deserves a serious referee despite a load-bearing linear-probe assumption. the 3 major comments →

arxiv 2607.11953 v2 pith:MPMKXRAK submitted 2026-07-11 cs.LG

When Does Reward Teach State? A Hidden-Automaton Instrument and the Group-Language Boundary

Jim Allchin This is my paper
classification cs.LG MSC 68Q4568T0768T0568Q70
keywords reinforcement learninglatent statereward shortcutdeterministic finite automatonlinear probingperception gapplanning gapgroup language
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that in a reinforcement-learning agent, earning high reward and learning the task's latent state are two different things that can be measured separately, and that the tie between them depends on optimizer strength, task structure, and observation informativeness. It builds a white-box testbed from hidden deterministic finite automata, where the experimenter knows the exact latent state and optimal return, so reward becomes a normalized score and the agent's internal representation can be probed exactly. The paper's key finding is that weak reinforcement learning can settle on reward shortcuts with the state probe at chance, and that permutation (group-language) structure in the automaton — computable before any training — flags these perception gaps on held-out automata at precision 0.86. It also shows that reward alone cannot distinguish a perception gap (state not recoverable) from a planning gap (state recoverable but unused), so high reward is not evidence that the state was learned. A sympathetic reader would care because this gives a concrete, auditable way to tell whether a stateful agent actually tracks the state of the workflow it is controlling.

Core claim

The paper's central claim is that reward success and latent-state learning are distinct, separately measurable quantities, and that their coupling is governed by three axes: optimizer strength, task structure (above all permutation/group-language structure), and observation informativeness. Using a hidden deterministic finite automaton (DFA) testbed where the agent observes a symbol stream, occasionally chooses the next symbol, and gets a single terminal reward for acceptance, the experimenter can compute the optimal return and the exact latent state at every step — yielding an oracle-normalized reward score and a direct linear probe of the agent's representation. The paper shows that under

What carries the argument

The hidden-automaton instrument: express the task as a hidden deterministic finite automaton (DFA) with partial control and sparse terminal reward, so the exact optimal return and the latent state at every timestep are computable by the experimenter. This yields an oracle-normalized reward score and a linear probe of the policies' pre-head features against the true state, with an off-policy supervised capacity check as a control. The structural predictor is the permutation (group-language) property — every symbol induces a bijection on the reachable states, so no input ever merges states — which a credit-assignment analysis (bounded by 'state-influence' and geometric decay at a state-blind p

Load-bearing premise

The perception gap is defined as 'latent state not linearly recoverable, though representable' via a linear probe on pre-head features plus a saturated off-policy supervised capacity check; if the agent stored the state in non-linear features that the linear probe misses, the perception-vs-planning distinction could be an artifact of probe class rather than a property of the agent.

What would settle it

Run the full instrument on a single permutation automaton with saturated supervised capacity where a sufficiently strong and budgeted PPO+GAE agent achieves near-optimal oracle-normalized reward, and check whether a non-linear probe (e.g., a small MLP on the same features) recovers the latent state with high accuracy while the paper's linear probe stays near chance. If so, the 'perception gap' is a probe-class artifact. Alternatively, find one non-permutation automaton that the paper's structural predictor flags as hard but where the agent learns the state perfectly with saturated capacity, wh

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • High reward is not evidence that the latent state was learned; state-aware evaluation must include an internal probe, not just behavioral reward.
  • A perception gap and a planning gap demand opposite fixes: make the latent state observable or latchable to close the first, and change the policy or reward to exploit the second.
  • Any finite-state workflow can be audited white-box by recovering its automaton and applying the same oracle and probe; the paper demonstrates this on real checksum logic and process logs.
  • The permutation/group-language property is a cheap pre-training red flag for state-blindness, but its one-directionality means non-group automata still need individual auditing.
  • Observation informativeness has a sharp threshold: a label-free auxiliary recovers state as soon as observations carry any state signal (β≈0.1), suggesting small observation leaks can bootstrap representation learning.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the linear-probe operationalization is right, the perception-gap concept could be ported to other settings where a known hidden variable exists, such as model-based planning or world-model evaluation in supervised sequence models, giving a measurement of 'state blindness' beyond RL.
  • The sharp β≈0.1 threshold hints that even negligible state leakage in observations can ignite representation learning, which is a testable design principle for environments where state is otherwise hidden.
  • The failure of the converse — non-permutation automata are heterogeneous — suggests there is no single structural predictor for all perception gaps; a broader theory would need to account for reward-shortcut stability beyond group structure.
  • The LLM-agent port suggests the same credit-starvation mechanism may limit large language models on accumulation-heavy tasks; the authors' open corner is a stronger optimizer, which is a concrete next experiment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper introduces a white-box instrument for measuring whether an RL agent that earns reward has also learned the task's latent state. Tasks are hidden deterministic finite automata (DFAs) with partial control and sparse terminal reward; since the automaton is known, the experimenter can compute an exact oracle for normalized return and can probe the exact latent state on every step without showing it to the agent. The central claim is that reward success and latent-state learning are distinct, separately measurable quantities, and that their coupling is governed by three axes: optimizer strength, task structure (in particular permutation/group-language structure), and observation informativeness. The paper reports a pre-registered A2C-vs-PPO control, a held-out validation of the structural predictor on 153 fresh automata (precision 0.86 for permutation structure flagging perception gaps), and a distinction between perception gaps (latent state not linearly recoverable) and planning gaps (state recoverable but not exploited), concluding that reward alone cannot localize the failure.

Significance. If valid, this is a significant contribution: it makes goal-misgeneralization and reward-shortcut worries measurable rather than purely behavioral, provides an exact oracle and exact probe target for any FSM-approximable task, and identifies a pre-training-computable warning signal (permutation/group-language structure) with a mechanistic credit-assignment explanation. The paper's strengths are substantial: the §5.2 pre-registered 2×2 de-confound, the episode-split probe protocol with selectivity controls, the saturated supervised-capacity checks, the 153-automaton held-out evaluation with Wilson CI and threshold robustness, and the explicit demotion of the non-transferring leverage threshold. The manuscript is unusually transparent about its own corrections and limitations.

major comments (3)
  1. [§4, §5.5, Fig. 2] The perception gap is operationalized as 'state not linearly recoverable, though representable': a task is labeled a perception gap when an on-policy linear probe recovers <0.5 balanced accuracy while an off-policy supervised capacity check saturates. The capacity check only shows that the architecture can be trained, under supervision, to linearly decode state; it does not rule out a nonlinearly encoded but present state in the RL-trained policy. A nonlinear probe with matched capacity could reclassify some of the 103 permutation automata, and the headline held-out precision 0.86 is computed against exactly this linear-probe label. Fig. 2's caption goes further and calls perception gaps 'state absent,' which is stronger than 'not linearly recoverable.' I recommend running nonlinear probes (e.g., a matched MLP probe) with the same episode split and selectivity control, and re-scoring the
  2. [§4, §5.4, §6] The planning-gap construct is defined as 'state recoverable but not exploited,' yet the interventional probe described in §4 as distinguishing a representation the policy consults from one merely present is reported only for 'the seven policies where an auxiliary installed the state' (§6). It is not applied to the planning-gap cases such as ends_with_1000 (§5.4), where a high probe gap accompanies low reward. Without the interventional check, a high linear-probe reading plus low reward could also be consistent with a state representation that is partially used but insufficient for control, or with a probe recovering state from features that do not influence action selection. The perception/planning distinction would be on firmer ground if the interventional probe were run on naturally arising planning-gap policies.
  3. [§5.5, Appendix D, Table 3] The held-out validation uses a stratified sample designed to populate both predictor classes, and the text notes that the predicted-hard class is only about 1% of uniform random diagnostic DFAs. Consequently the precision 0.86 is a conditional positive-predictive value within the constructed sample, not an estimate for a natural deployment distribution. The paper is transparent about the sampling, but the practical claim that permutation structure is a 'high-precision warning signal' would benefit from an explicit statement of how precision varies with the natural prevalence of permutation automata and with the within-class ratio of low- versus high-leverage permutation automata; otherwise a reader may over-generalize the 0.86 figure.
minor comments (6)
  1. [Fig. 2 caption] "Perception gap (state absent)" conflicts with the formal definition in §4/§5.5 as "not linearly recoverable." Please align the caption wording with the operational definition.
  2. [Table 2 and §5.4 text] The text says the recovered fraction shrinks "from ≈1 at k=2,3," but Table 2 reports PPO/ceiling = 0.83 at k=2 and 0.99 at k=3. Please reconcile the textual summary with the table.
  3. [Appendix A] The pre-registration is described as fixed before running, but no timestamp, versioned document, or independent registration identifier is given. Adding a date and/or hash would strengthen the claim.
  4. [§5.4 (Spearman correlations)] The correlations between behavioral signals and probe gap (Spearman −0.29, −0.25) are reported without confidence intervals or exact sample size. Given the small curated suite, bootstrap CIs would help calibrate the "nearly uncorrelated" claim.
  5. [§5.5, Appendix D] The 120-automaton extension uses 3 seeds per automaton. Please report seed-level label stability (e.g., how many automaton labels would change if labels were computed from individual seeds versus the mean), since the Wilson CI treats automata as independent units.
  6. [General wording] The phrase "pre-training structural warning signal" may be misread as referring to pretrained models; consider "pre-RL" or "pre-training-time" structural warning signal.

Circularity Check

0 steps flagged

No significant circularity: the oracle, probe labels, and structural predictor are independent measurements.

full rationale

The paper's derivation chain is self-contained and its central predictive claim is not circular. The hidden DFA and exact oracle are defined before any RL and are used only as measurement instruments; the latent state is known exactly from the automaton, so the probe labels are read from the agent's activations, not from the predictor. The structural predictor (permutation/group-language structure) is computed from the transition function δ before any training and is validated on 153 fresh automata against probe-derived labels, with capacity controls (off-policy supervised probes saturate) and explicit checks for control-relevant state quotients. The in-sample leverage threshold was fitted to 15 curated tasks, but the paper transparently reports that it does not transfer and demotes it, so it is not presented as a held-out prediction. The Appendix B credit-assignment result is a conditional mathematical derivation under an explicit state-blind assumption, not an equation that reduces the empirical finding to its own input. No load-bearing self-citation or author-imported uniqueness theorem appears; the algebraic references (Schützenberger, Krohn–Rhodes) are external classical results used for classification, not to force the empirical conclusion. The linear-probe operationalization concern raised by the skeptic is a legitimate construct-validity limitation, but it is not a circularity: the paper defines 'perception gap' as 'state not linearly recoverable, though representable' and is careful to use that definition in the main claims, even if Figure 2's caption uses the shorthand 'state absent.' That is an overstatement of the operational definition, not a reduction of the prediction to its inputs. Overall, no step exhibits a fitted parameter being renamed as a prediction, no equation equals itself by construction, and no conclusion is forced by a self-citation chain.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

No new physical or ontological entities are introduced. The perception gap, planning gap, oracle, and probes are operational measurement constructs; the controllable-hold and observation-channel interventions are design features, not postulates.

free parameters (3)
  • p_control (partial-control probability) = 0.5
    Chosen by hand for all main experiments. The exact oracle normalizes for it, but the reward/state coupling results are conditional on this experimental knob; it is not fitted to data.
  • perception-gap probe threshold = probe recovery < 0.5
    Definitional threshold used in Section 5.5 to label a perception gap. Sensitivity is checked at 0.4 and 0.6 (precision 0.84 and 0.88), so it is not a sharply fitted value.
  • in-sample control-leverage threshold (L <= 0.50 hard) = 0.50/0.519 boundary
    Fitted to the 15 curated tasks to obtain in-sample ROC-AUC 1.00. The held-out test shows it does not transfer, and the paper demotes it to a within-group severity proxy. Listed for transparency.
axioms (5)
  • domain assumption The latent DFA state is a deterministic function of the fully observed symbol prefix.
    Sections 1 and 3: the agent sees the full emitted prefix and control flags, so state tracking is a memory/computation problem, not POMDP belief tracking; this licenses exact-state probing.
  • standard math Known DFA and control distribution make optimal acceptance probability computable by backward induction.
    Section 3 oracle equation V(q,r) = ...; finite-horizon Bellman optimality.
  • standard math Policy-gradient score-function identity with an action-independent baseline.
    Appendix B Proposition 1 proof; standard REINFORCE identity.
  • domain assumption At a state-blind policy, the random-symbol operator on a permutation automaton is doubly stochastic with Dobrushin contraction lambda < 1, giving geometric credit decay.
    Appendix B Proposition 2; holds at initialization and for policies not yet tracking the state. The paper argues relaxing it requires the state-tracking that the sparse signal must install.
  • domain assumption A saturated off-policy supervised probe means the architecture can represent the state, so low on-policy linear recovery is a learning gap rather than a capacity limit.
    Sections 5.1 and 5.5 capacity controls; load-bearing for labeling perception gaps and interpreting the held-out validation.

pith-pipeline@v1.3.0-alltime-deepseek · 16856 in / 18891 out tokens · 187065 ms · 2026-08-02T07:13:34.813074+00:00 · methodology

0 comments
read the original abstract

Does a reinforcement-learning agent that earns high reward represent its task's latent state, or only a reward-correlated shortcut? Usually this is unanswerable: the "true state" is undefined. We make it exactly answerable with a white-box instrument: express the task as a hidden deterministic finite automaton (DFA), let the agent observe a symbol stream and intermittently choose the next symbol under partial control, with a sparse terminal reward for acceptance. Knowing the automaton gives two things for free: the optimal return (so reward becomes an interpretable normalized score) and the exact latent state at every step (so we can probe the agent's representation without ever showing it). Reward success and latent-state learning become separately measured quantities whose coupling is governed by three axes. Optimizer strength: under weak on-policy RL the agent can lock onto a reward-shortcut, earning reward with the state probe at chance; whether it escapes is set by optimizer strength, budget, and task structure -- a pre-registered control isolates the optimizer (PPO+GAE), which escapes fastest but only partially and with high seed variance. Task structure: permutation (group-language) structure is a warning sign computable from the transition function before any training, and held out on 153 capacity-controlled fresh automata it flags perception gaps at precision 0.86 (89 of 103), in one direction only. Observation informativeness: a label-free auxiliary is vacuous when observations carry no state and recovers it sharply once they carry any (a threshold, saturating near beta=0.1). The payoff is a distinction reward-only evaluation cannot make: a perception gap (state not linearly recoverable, though representable) versus a planning gap (state recoverable but not exploited). Reward alone cannot localize which; high reward alone is not evidence the state was learned.

Figures

Figures reproduced from arXiv: 2607.11953 by Jim Allchin.

Figure 1
Figure 1. Figure 1: Three controllable axes of the reward↔state coupling, all from values reported below. (a) At a fixed budget the state-probe gap climbs monotonically with optimizer competence, closing to the privileged-auxiliary ceiling once oracle-potential shaping supplies credit directly (a separate reward-signal change, not an optimizer step; §5.2)—the chance probe is a reward-shortcut the agent escapes with a stronger… view at source ↗
Figure 2
Figure 2. Figure 2: Behavior cannot separate what the probe separates. Each point is an RL-evaluated task. Horizontal axis: reward gap 1−valn (what a behavioral diagnostic sees). Vertical axis: probe gap (recoverable latent state). At a given large reward gap the probe ranges from chance (perception gap, bottom right) to near-ceiling (planning gap, top right). automata: an initial 33 (5 seeds) and a 120-automaton extension (3… view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references · 6 canonical work pages

  1. [1]

    Langosco, J

    L. Langosco, J. Koch, L. Sharkey, J. Pfau, L. Orseau, D. Krueger. Goal Misgeneralization in Deep RL. ICML, 2022. arXiv:2105.14111

    Pith/arXiv arXiv 2022

  2. [2]

    Shah et al

    R. Shah et al. Goal Misgeneralization: Why Correct Specifications Aren't Enough for Correct Goals. 2022. arXiv:2210.01790

    Pith/arXiv arXiv 2022

  3. [3]

    B. Liu, J. T. Ash, S. Goel, A. Krishnamurthy, C. Zhang. Transformers Learn Shortcuts to Automata. ICLR, 2023. arXiv:2210.10749

    Pith/arXiv arXiv 2023

  4. [4]

    M. P. Sch\"utzenberger. On Finite Monoids Having Only Trivial Subgroups. Information and Control, 1965

    1965

  5. [5]

    Krohn, J

    K. Krohn, J. Rhodes. Algebraic Theory of Machines, I. Transactions of the American Mathematical Society, 116:450--464, 1965

    1965

  6. [6]

    Strobl et al

    L. Strobl et al. What Formal Languages Can Transformers Express? A Survey. TACL, 12:543--561, 2024. arXiv:2311.00208

    Pith/arXiv arXiv 2024

  7. [7]

    Chang, Y

    Y. Chang, Y. Bisk. Language Models Need Inductive Biases to Count Inductively. 2024. arXiv:2405.20131

    Pith/arXiv arXiv 2024

  8. [8]

    Huang et al

    X. Huang et al. A Formal Framework for Understanding Length Generalization in Transformers. ICLR, 2025. arXiv:2410.02140

    Pith/arXiv arXiv 2025

  9. [9]

    Huang, Z

    Y. Huang, Z. Wen, A. Singh, Y. Chi, Y. Chen. Transformers Provably Learn Chain-of-Thought Reasoning with Length Generalization. NeurIPS, 2025. arXiv:2511.07378

    arXiv 2025

  10. [10]

    K. Li, A. K. Hopkins, D. Bau, F. Vi\'egas, H. Pfister, M. Wattenberg. Emergent World Representations. ICLR, 2023. arXiv:2210.13382

    Pith/arXiv arXiv 2023

  11. [11]

    Nanda, A

    N. Nanda, A. Lee, M. Wattenberg. Emergent Linear Representations in World Models of Self-Supervised Sequence Models. 2023. arXiv:2309.00941

    Pith/arXiv arXiv 2023

  12. [12]

    Hewitt, P

    J. Hewitt, P. Liang. Designing and Interpreting Probes with Control Tasks. EMNLP, 2019. arXiv:1909.03368

    Pith/arXiv arXiv 2019

  13. [13]

    Alain, Y

    G. Alain, Y. Bengio. Understanding Intermediate Layers Using Linear Classifier Probes. 2016. arXiv:1610.01644

    Pith/arXiv arXiv 2016

  14. [14]

    Gelada, S

    C. Gelada, S. Kumar, J. Buckman, O. Nachum, M. Bellemare. DeepMDP. ICML, 2019. arXiv:1906.02736

    Pith/arXiv arXiv 2019

  15. [15]

    Zhang, R

    A. Zhang, R. McAllister, R. Calandra, Y. Gal, S. Levine. Learning Invariant Representations for RL without Reconstruction (DBC). ICLR, 2021. arXiv:2006.10742

    Pith/arXiv arXiv 2021

  16. [16]

    T. Ni, B. Eysenbach, R. Salakhutdinov. Recurrent Model-Free RL Can Be a Strong Baseline for Many POMDPs. ICML, 2022. arXiv:2110.05038

    Pith/arXiv arXiv 2022

  17. [17]

    Jaderberg et al

    M. Jaderberg et al. Reinforcement Learning with Unsupervised Auxiliary Tasks (UNREAL). ICLR, 2017. arXiv:1611.05397

    Pith/arXiv arXiv 2017

  18. [18]

    Toro Icarte et al

    R. Toro Icarte et al. Using Reward Machines for High-Level Task Specification and Decomposition in RL. ICML, 2018

    2018

  19. [19]

    D. Angluin. Learning Regular Sets from Queries and Counterexamples. Information and Computation, 1987

    1987

  20. [20]

    Oncina, P

    J. Oncina, P. Garc\'ia. Inferring Regular Languages in Polynomial Updated Time. In Pattern Recognition and Image Analysis, World Scientific, 1992

    1992

  21. [21]

    Del\'etang, A

    G. Del\'etang, A. Ruoss, J. Grau-Moya, et al. Neural Networks and the Chomsky Hierarchy. ICLR, 2023. arXiv:2207.02098

    Pith/arXiv arXiv 2023

  22. [22]

    Bhattamishra, K

    S. Bhattamishra, K. Ahuja, N. Goyal. On the Ability and Limitations of Transformers to Recognize Formal Languages. EMNLP, 2020. arXiv:2009.11264

    Pith/arXiv arXiv 2020

  23. [23]

    Weiss, Y

    G. Weiss, Y. Goldberg, E. Yahav. On the Practical Computational Power of Finite Precision RNNs for Language Recognition. ACL, 2018. arXiv:1805.04908

    Pith/arXiv arXiv 2018

  24. [24]

    J. J. Michalenko, et al. Representing Formal Languages: A Comparison Between Finite Automata and Recurrent Neural Networks. ICLR, 2019. arXiv:1902.10297

    Pith/arXiv arXiv 2019

  25. [25]

    Toro Icarte, et al

    R. Toro Icarte, et al. Learning Reward Machines: A Study in Partially Observable Reinforcement Learning. Artificial Intelligence, 323:103989, 2023. arXiv:2112.09477

    Pith/arXiv arXiv 2023

  26. [26]

    R. I. Brafman, G. De Giacomo. Regular Decision Processes: A Model for Non-Markovian Domains. IJCAI, 2019

    2019

  27. [27]

    M. L. Littman, R. S. Sutton, S. Singh. Predictive Representations of State. NeurIPS, 2001

    2001

  28. [28]

    de Haan, D

    P. de Haan, D. Jayaraman, S. Levine. Causal Confusion in Imitation Learning. NeurIPS, 2019. arXiv:1905.11979

    Pith/arXiv arXiv 2019

  29. [29]

    A. Pan, K. Bhatia, J. Steinhardt. The Effects of Reward Misspecification: Mapping and Mitigating Misaligned Models. ICLR, 2022. arXiv:2201.03544

    Pith/arXiv arXiv 2022

  30. [30]

    Skalse, N

    J. Skalse, N. Howe, D. Krasheninnikov, D. Krueger. Defining and Characterizing Reward Gaming. NeurIPS, 2022. arXiv:2209.13085

    Pith/arXiv arXiv 2022

  31. [31]

    B. F. van Dongen. BPI Challenge 2012 (loan application process). 4TU.ResearchData, 2012. doi:10.4121/uuid:3926db30-f712-4394-aebc-75976070e91f

    work page doi:10.4121/uuid:3926db30-f712-4394-aebc-75976070e91f 2012

  32. [32]

    B. F. van Dongen. BPI Challenge 2015 Municipality 1 (building-permit applications). 4TU.ResearchData, 2015. doi:10.4121/uuid:a0addfda-2044-4541-a450-fdcc9fe16d17

    work page doi:10.4121/uuid:a0addfda-2044-4541-a450-fdcc9fe16d17 2015

  33. [33]

    B. F. van Dongen. BPI Challenge 2017 (loan application process). 4TU.ResearchData, 2017. doi:10.4121/uuid:5f3067df-f10b-45da-b98b-86ae4c7a310b

    work page doi:10.4121/uuid:5f3067df-f10b-45da-b98b-86ae4c7a310b 2017

  34. [34]

    B. F. van Dongen. BPI Challenge 2019 (purchase-order/procurement process). 4TU.ResearchData, 2019. doi:10.4121/uuid:d06aff4b-79f0-45e6-8ec8-e19730c248f1

    work page doi:10.4121/uuid:d06aff4b-79f0-45e6-8ec8-e19730c248f1 2019

  35. [35]

    Mannhardt

    F. Mannhardt. Hospital Billing -- Event Log. 4TU.ResearchData, 2017. doi:10.4121/uuid:76c46b83-c930-4798-a1c9-4be94dfeb741

    work page doi:10.4121/uuid:76c46b83-c930-4798-a1c9-4be94dfeb741 2017

  36. [36]

    de Leoni, F

    M. de Leoni, F. Mannhardt. Road Traffic Fine Management Process. 4TU.ResearchData, 2015. doi:10.4121/uuid:270fd440-1057-4fb9-89a9-b699b47990f5

    work page doi:10.4121/uuid:270fd440-1057-4fb9-89a9-b699b47990f5 2015

  37. [37]

    Mannhardt

    F. Mannhardt. Sepsis Cases -- Event Log. 4TU.ResearchData, 2016. doi:10.4121/uuid:915d2bfb-7e84-49ad-a286-dc35f063a460

    doi:10.4121/uuid:915d2bfb-7e84-49ad-a286-dc35f063a460 2016