Pith. sign in

REVIEW 5 major objections 5 minor 34 references

Evaluation Blindness: How Silent Measurement Failures Corrupt AI Systems from Training to Deployment

T0 review · 5 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read The paper proposes that silent AI failures are one structural property: the measurements that should catch them look healthy.

desk verdict Useful unifying framework and a solid GRPO case study, but the 53% 'silent majority' claim doesn't survive contact with the labeling methodology. read the letter →

arxiv 2608.02786 v1 pith:BN56XD2I submitted 2026-08-03 cs.LG

Priyanka Bajaj (Independent Researcher) This is my paper · ORCID
classification cs.LG
keywords evaluationblindnesssilentfailuresLLMproductionmonitoringfailuretaxonomyrewardhackingGRPOimportancesamplingbudgetAIlifecyclemeasurement
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Many AI failures are invisible at the moment they happen: the loss curve falls, reward climbs, error rates stay flat, and no alert fires, while the system is quietly degrading. The paper's central proposal is that these are not many separate bugs but one structural property, which it names evaluation blindness: a measurement function produces readings indistinguishable from a healthy state while the system is failing, with no auxiliary signal flagging the gap. It formalizes the property and argues that the same definition unifies training-time corruption (reward hacking, benchmark contamination, a concrete bug in a widely used reinforcement-learning fine-tuning implementation) with deployment-time silent failures. If the paper is right, measurement infrastructure should be treated as a correctness concern at every lifecycle stage, and acceptable failure rates should be set per use-case risk rather than by benchmark scores. An empirical taxonomy of 50 incidents, 36 from verifiable public sources, reports that 53% of the verifiable incidents were silent.

What carries the argument

The central object is Definition 1's evaluation-blindness condition together with the detectability predicate $\delta(s,\mathcal{M})$. A measurement set $\mathcal{M}=\{M_1,\dots,M_n\}$ fails to detect state $s$ when $\delta(s,\mathcal{M})=0$, meaning no deployed measurement lands outside its reference distribution $D_\text{ok}$; adding the requirement that no auxiliary signal flag the gap makes the failure silent rather than merely undetected. This predicate carries the argument because it converts "silent failure" from a metaphor into a checkable property of a measurement set, and it is what lets the paper classify each failure as structurally or incidentally blind and then stage the same property at training time and deployment time.

What would settle it

Take the 36 real incidents, have independent annotators who have not seen the paper's labels and do not know whether an incident came from a court filing, a regulator, or a postmortem, and relabel detectability from the raw narratives. If the silent fraction does not reproduce, or varies substantially by source type, the 53% silent-majority claim is an artifact of the convenience sample.

Watch

Extended reading notes

Core claim

The paper claims that a failing AI system in failure class $F$ is evaluation-blind with respect to measurement function $M$ when $M$'s readings are distributionally indistinguishable from the non-failing reference while the system is failing and no auxiliary signal flags the gap. It distinguishes structural blindness (no function of the available signals can detect the failure) from incidental blindness (a known monitor is simply absent), and defines an explicit detectability predicate $\delta(s,\mathcal{M})$. On this basis it presents four training-time case studies, including a real gradient-corruption bug in an open-source reinforcement-learning fine-tuning implementation whose loss curves looked normal for the entire affected run, and a six-class deployment taxonomy in which 19 of 36 verifiable public incidents (53%) were silent. The unifying consequence the author draws is that training-time and deployment-time measurement failure are the same problem at different lifecycle stages.

Load-bearing premise

The central claims rest on treating a convenience sample of 50 documented incidents as representative enough to support a general silent-majority rate, and on assuming the only training-time measurements in play are loss, reward, and benchmark scores.

Editorial extensions

If this is right

  • Production monitoring for LLM systems should include distributional signals such as output-length distribution, semantic similarity to baseline outputs, and retrieval freshness, not just error rates and latency percentiles.
  • Training dashboards built only from loss, reward, and benchmark scores are insufficient; RL fine-tuning runs need out-of-loop measurements such as gold-label preference evals, membership-inference checks, and diversity probes.
  • Decision-critical use cases must be able to detect failures at roughly the 0.1% level, which makes the monitoring and guardrail requirements explicit before model selection.
  • Safety and compliance failures should be treated as missing verification layers between model output and consequential action, not as model behavior that alignment training can eliminate.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The structural-versus-incidental distinction suggests a practical design rule: for each failure class, teams can enumerate the minimal measurement set that makes $\delta=1$, turning the taxonomy into a coverage checklist.
  • If the 53% silent rate holds beyond this sample, public incident databases are systematically biased toward alertable failures, so the true silent share of production AI failures is likely higher, not lower.
  • The failure-budget idea could be extended to treat silent failures as explicit error-budget violations, so that an undetected failure consumes budget even before it causes user-visible harm.
  • The same formal definition may transfer to other autonomous systems with monitoring stacks, such as robots or medical devices, where the question "can any deployed measurement distinguish failing from healthy?" is equally load-bearing.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper introduces 'evaluation blindness' as a structural property of measurement functions: a measurement M exhibits it with respect to failure class F when M(s) is indistinguishable from the non-failing reference while the system is failing and no auxiliary signal flags the gap. It formalizes this with a detectability predicate, claims the property unifies training-time failures (reward hacking, an importance-sampling bug in TRL, benchmark contamination, mode collapse) with deployment-time failures, and proposes a six-class failure taxonomy (C1-C6) validated against 50 incidents (36 from verifiable public sources), reporting that 53% of the 36 verifiable incidents were silent. It closes with a per-use-case failure budget framework and releases data, code, and schema. The central conceptual contribution is a unifying vocabulary for a real phenomenon, but the formal definition is not fully operational as written, the empirical 'silent majority' claim is not supported by the sampling and labeling procedure, and there is an internal contradiction about which failure class is structurally silent.

Significance. If the formal definition were operationalized and the empirical claims properly scoped, this would be a useful contribution: it connects training-time and deployment-time monitoring failures, provides a concrete case study (TRL PR #6594) that is checkable from public artifacts, offers a classification scheme with remediation implications, and releases reproducible artifacts. The failure budget framework, while simple, is a reasonable organizational tool. The paper is honest about several limitations, and the existence of released code and data is a strength. However, as it stands, the load-bearing quantitative claim (a 'silent majority' of public incidents) rests on a convenience sample and labels that conflate absence of documentation with absence of detection, and the formal definition does not yet support the structural/incidental distinction that is invoked to classify the GRPO bug.

major comments (5)
  1. [Abstract; Section 2.3; Section 5.6 and Table 3] The abstract and Section 2.3 state that the Operational class (C6) is '100% silent by structural definition', but Table 3 gives C6 Detectability as 'Delayed' and Section 5.6 gives its detectability profile as 'Delayed'. The class that is 100% blind by definition is C4 (Evaluation), as stated in Section 2.3. This internal contradiction affects the taxonomy's consistency and must be fixed by correcting the abstract and any downstream claims that rely on C6 being structurally silent.
  2. [Section 2.1, Definition 1 and Section 2.2] The formal predicate is not operational as written. Definition 1 compares M(s), a single point, to D_ok, a reference distribution, via 'distributional indistinguishability up to tolerance epsilon', but no test statistic, sample size, or procedure for constructing D_ok for arbitrary M is given, and epsilon is unquantified. Consequently, δ(s,M) in Section 2.2 cannot be evaluated for a concrete system, and the paper's later statements that a particular failure 'is structural' or 'is incidental' cannot be checked against the definition.
  3. [Section 2.1, Definition 2 and Section 6] Definition 2 relies on M*, the set of 'all structurally available measurement functions', but M* is nonconstructive: no characterization is given of which measurements count as available for a given observable-signal set. This matters directly for Case Study 1 (Section 6.1), where the GRPO bug is labelled 'structural' because the analysis restricts M_train to {loss, reward, benchmark} and excludes per-token audits, gradient checks, or specification comparison. Without a justification that those measurements are not 'structurally available', the structural classification is true by construction rather than by analysis.
  4. [Section 8.1 and Section 8.2] The headline claim that 53% (19/36) of verifiable public incidents are silent is not supported by the sampling and coding procedure. The sample is a convenience set of severe, publicly documented incidents with no sampling frame. Court documents and regulatory filings record legal and factual findings, not whether an internal alert fired; coding the absence of a documented alert as 'silent' conflates 'no alert existed' with 'alert existed but was not reported'. The finding also lacks any confidence interval: for 19/36, a binomial 95% CI spans roughly 36-68%, which includes values below 50%. The robustness check with 14 synthetic composites does not address this concern, because those composites were constructed by the same author to cover under-represented classes.
  5. [Section 8.1 and Section 8.4] The inter-rater reliability study reports perfect agreement on 20 class labels (C1-C6), but it does not report agreement on the detectability classification ('silent' vs 'immediate' vs 'delayed'), which is the exact variable driving the paper's central empirical finding. The single-annotator limitation is acknowledged in Section 8.4, but the key label remains validated only by the author. The paper should either provide reliability evidence for the detectability labels or explicitly demote the prevalence claim to a dataset statistic rather than a general property of public incidents.
minor comments (5)
  1. [Abstract] There is a typographical error in the abstract: 'We identifyevaluation blindness' should read 'We identify evaluation blindness'.
  2. [Section 4.1] Section 4.1 says the 'key empirical finding is that 50% of incidents are silent', while Section 8.2 reports 53% for the 36 verifiable incidents (and 52% including synthetic composites). These figures should be reconciled.
  3. [Section 7.3] The calibration paragraph mentions 'Basel model risk management requirements' but cites only DORA, the EU AI Act, and an FCA review; a specific reference to the Basel framework should be added if this calibration claim is retained.
  4. [Section 8.3] The classifier results are described as '98% label consistency'; since the pattern set was developed on the same 50 incidents, the term 'resubstitution accuracy' or 'training-set consistency' would be more precise and would avoid the impression of held-out evaluation.
  5. [Table 2] Table 2 lists MTTD as 'Numeric Hours (0.1 to >1000)', but the dataset statistics in Table 7 report medians without clarity on how missing MTTD values were handled; a sentence in Section 8.2 explaining the treatment of unrecorded MTTD would help.

Circularity Check

2 steps flagged · score 6.0 of 10

Silent-majority prevalence and the GRPO 'structural blindness' label both reduce partly to the paper's own definitions.

  1. self definitional [Section 5.4 (C4 definition and detectability profile); Section 8.2 (Table 7, silent majority); Section 2.3 (admission)]
    "C4 failures are the deployment-time realisation of evaluation blindness: the deployed Mdeploy has δ(s,Mdeploy)=0 not because the failure is subtle, but because the measurement infrastructure is structurally absent or misconfigured. ... Detectability profile..Silent. ... The C4 Evaluation failure class is 100% blind by definition (the measurement infrastructure is the failed component). ... Silent failures are concentrated in C4 (100% of 5 real incidents)."

    C4 is defined as the failure of the measurement infrastructure itself, so δ(s,Mdeploy)=0 and a 'silent' detectability label follow from the class definition rather than from independent incident evidence. Those five C4 real incidents are then counted inside the 53% silent figure. Removing them leaves 14 of 31 verifiable incidents silent (45%), so the 'silent majority' headline depends on a by-definition component. The paper's own sentence '100% blind by definition' is the explicit reduction of the empirical claim to its definitional input.

  2. self definitional [Section 6 intro (Mtrain restriction); Section 2.3 (structural classification of GRPO bug); Table 4 (detection methods)]
    "Training-time evaluation blindness occurs when the measurement set available during model training (Mtrain = {loss, reward, benchmark score}) fails to detect that the training process is producing a corrupted policy. ... The GRPO IS ratio bug (Section 6) is structural — no training-time measurement could have flagged the gradient corruption without external specification comparison. ... Detection Requires ... Spec comparison or per-token ratio audit."

    The paper fixes Mtrain to contain only loss, reward, and benchmark score, and then uses that restricted set to declare the GRPO bug 'structural' under Definition 2, which instead requires checking all structurally available measurements M*. Because Table 4 itself lists 'per-token ratio audit' as a detection method, the bug is detectable in principle and should be classified as incidental, not structural. The 'no training-time measurement could have flagged' assertion is true only relative to the paper's own restriction of Mtrain, making the structural label an artifact of the definitional choice rather than of the observable signals.

full rationale

The paper's formal framework is largely self-contained: Definition 1 and the detectability predicate are stated in the paper, the four training-time case studies rest on external literature and a real TRL pull request, and the failure budget framework is a free-standing proposal. There is an overlapping-annotator citation (Pandey [21]) but it is disclosed and is used as supporting related work and IRR validation, not as the derivation of the central claim, so it does not by itself raise circularity. The two concrete circular reductions are definitional. First, the C4 class is defined as broken or absent measurement infrastructure, so 'silent' is true by definition for every C4 incident; yet those incidents are included in the central empirical claim that 53% of verifiable incidents are silent. Second, the GRPO bug is called 'structural' only after Mtrain has been restricted to loss, reward, and benchmark, which excludes the per-token ratio audit that Table 4 says would detect it. These are partial, not total, circularity: the unification concept and most of the taxonomy still have independent empirical content, but the headline silent-majority figure and one structural-classification claim reduce partly to the paper's own definitions.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on an idealized measurement-set model, an unquantified tolerance parameter, hand-chosen risk rates and severity weights, and unverified external incident narratives. No new physical or ontological entities are introduced; the taxonomy classes are analytical constructs rather than entities with independent falsifiable handles.

free parameters (3)
  • failure budget rates rho(FC_A..FC_D) = 1.0, 5.0, 20.0, 100.0 per 1,000 requests
    Hand-chosen thresholds in Table 5, stated as calibrated to EU AI Act, DORA and FCA practice, but no derivation or validation from actual failure data is provided.
  • severity weights w(critical, high, medium, low) = 3, 2, 1, 0.5
    Assigned in Section 7.2 with no empirical basis; these weights drive the WARNING utilization example in Table 6.
  • distributional indistinguishability tolerance epsilon = unspecified
    Introduced in Definition 1 but never quantified, so the formal detectability predicate is not evaluable on real measurements.
assumptions (5)
  • domain assumption A single system state s can be compared to a reference distribution D_ok via distributional indistinguishability up to tolerance epsilon.
    Definition 1 applies a distributional comparison to a single state without specifying repeated sampling or an operational test.
  • domain assumption The 36 public incident narratives accurately report whether alerts fired and how the incident was discovered.
    Section 8 labels incidents from court documents, regulatory filings, and press reports; the detectability ground truth is taken at face value from external narratives.
  • domain assumption The TRL PR #6594 bug occurred as described and is the root cause of corrupted GRPO gradients.
    Section 6.1 relies on an external pull request and the DeepSeek-V3 specification; neither is independently verified inside this paper.
  • domain assumption Detectability labels assigned by the author and one overlapping annotator are reliable.
    Section 8.1 reports kappa = 1.00 on a 20-item subset, but the annotator is the author of a cited concurrent paper and the full set was labeled by a single author.
  • domain assumption Risk class thresholds from EU AI Act, DORA, and FCA practice support the specific numerical failure budget rates.
    Section 7.3 asserts calibration without showing how the regulatory provisions imply the specific rates 1, 5, 20, and 100 per 1,000 requests.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Evaluation Blindness: How Silent Measurement Failures Corrupt AI Systems from Training to Deployment." pith.science (2026). https://pith.science/paper/BN56XD2I

@misc{pith2026260802786,
  author       = {Pith},
  title        = {Pith review of: Evaluation Blindness: How Silent Measurement Failures Corrupt AI Systems from Training to Deployment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BN56XD2I}},
  note         = {Machine review of arXiv:2608.02786}
}
read the original abstract

AI systems can fail silently. The failure propagates through training loops, evaluation pipelines, and production monitoring stacks until downstream harm makes it visible. This paper introduces evaluation blindness: a measurement function M exhibits evaluation blindness with respect to failure class F when it produces readings indistinguishable from a healthy state while the system is actually failing, with no auxiliary signal flagging the gap. The problem surfaces at two lifecycle stages the literature has treated separately. At training time, reward models are gamed, importance-sampling corrections are silently miscalculated, and benchmark contamination inflates fine-tuning evaluations, all while loss curves look healthy and gradient updates proceed normally. At deployment time, monitoring fails to catch six classes of production failure, including an Operational category that is 100% silent by structural definition. We provide a formal detectability predicate unifying both stages. Four training-time case studies trace concrete breakdowns, including a real implementation bug in TRL PR #6594 where gradients are corrupted as loss decreases normally. A six-class taxonomy validated against 50 real-world incidents from court documents and regulatory filings finds that 53% of verifiable public failures were silent. A failure budget framework ties acceptable failure rates to use-case risk class. The implication is direct: measurement infrastructure is a correctness concern across the full AI lifecycle, not just at evaluation time. Data, code, and taxonomy schema are at https://github.com/priyanka25aug/llm-failure-taxonomy.

Figures

Figures reproduced from arXiv: 2608.02786 by the authors.

Figure 1
Figure 1. Incident distribution (left) and detectability profile (right). Safety & Compliance and Integration each account for 26% of incidents. 53% of 36 verifiable public incidents are silent — the core finding motivating distributional monitoring as a first-class engineering concern [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗
Figure 2
Figure 2. Mean Time to Detection (MTTD) by failure class, log scale. Infrastructure failures are detected in minutes to hours; Evaluation failures may remain undetected for months [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 27 canonical work pages

  1. [1]

    AI, algorithmic, and automa- tion incidents and controversies (AIAAIC)

    AIAAIC. AI, algorithmic, and automa- tion incidents and controversies (AIAAIC). https://www.aiaaic.org/, 2023

    work page 2023

  2. [2]

    Amershi, A

    S. Amershi, A. Begel, C. Bird, R. DeLine, H. Gall, E. Kamar, N. Nagappan, B. Nushi, 15 and T. Zimmermann. Software engineer- ing for machine learning: A case study. In Proceedings of the 41st International Con- ference on Software Engineering: Software Engineering in Practice, pages 291–300. IEEE Press, 2019

    work page 2019

  3. [3]

    Y . Bai, A. Jones, K. Ndousse, A. Askell, A. Chen, N. DasSarma, D. Drain, S. Fort, D. Ganguli, T. Henighan, et al. Constitu- tional AI: Harmlessness from AI feedback. arXiv preprint arXiv:2212.08073, 2022

    arXiv 2022

  4. [4]

    S. R. Bowman and G. E. Dahl. What will it take to fix benchmarking in natural lan- guage understanding? InProceedings of NAACL-HLT, pages 1843–1855, 2021

    work page 2021

  5. [5]

    Air canada chatbot liable for misin- formation on bereavement fares

    British Columbia Civil Resolution Tri- bunal. Air canada chatbot liable for misin- formation on bereavement fares. Tribunal Decision No. SC-2023-005226, 2024

    work page 2023

  6. [6]

    Reg- ulation (EU) 2022/2554 on digital oper- ational resilience for the financial sector (DORA)

    European Parliament and Council. Reg- ulation (EU) 2022/2554 on digital oper- ational resilience for the financial sector (DORA). Official Journal of the European Union, 2022

    work page 2022

  7. [7]

    Reg- ulation (EU) 2024/1689 of the european parliament and of the council laying down harmonised rules on artificial intelligence (Artificial Intelligence Act)

    European Parliament and Council. Reg- ulation (EU) 2024/1689 of the european parliament and of the council laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Jour- nal of the European Union, 2024

    work page 2024

  8. [8]

    Artificial in- telligence in financial services: Review of firms’ approaches to consumer duty com- pliance

    Financial Conduct Authority. Artificial in- telligence in financial services: Review of firms’ approaches to consumer duty com- pliance. FCA Thematic Review TR24/1, 2024

    work page 2024

Show all 34 references
  1. [9]

    L. Gao, S. Biderman, S. Black, L. Gold- ing, T. Hoppe, C. Foster, J. Phang, H. He, A. Thite, N. Nabeshima, et al. Reward hacking in RLHF: Silent corruption of pol- icy training through proxy optimisation. arXiv preprint arXiv:2606.03238, 2025

    2025 arXiv

  2. [10]

    Greshake, S

    K. Greshake, S. Abdelnabi, S. Mishra, C. Endres, T. Holz, and M. Fritz. Not what you’ve signed up for: Compromising real- world LLM-integrated applications with indirect prompt injection.arXiv preprint arXiv:2302.12173, 2023

    2023 arXiv

  3. [11]

    Huang, W

    L. Huang, W. Yu, W. Ma, W. Zhong, Z. Feng, H. Wang, Q. Chen, W. Peng, X. Feng, B. Qin, and T. Liu. A sur- vey on hallucination in large language models: Principles, taxonomy, challenges, and open questions.arXiv preprint arXiv:2311.05232, 2023

    2023 arXiv

  4. [12]

    Fix GRPO importance sampling ratio: Replace per-token with sequence-mean in KL bias correction (PR #6594)

    Hugging Face TRL Contributors. Fix GRPO importance sampling ratio: Replace per-token with sequence-mean in KL bias correction (PR #6594). GitHub Pull Request, https://github.com/ huggingface/trl/pull/6594, 2024

    2024

  5. [13]

    Z. Ji, N. Lee, R. Frieske, T. Yu, D. Su, Y . Xu, E. Ishii, Y . J. Bang, A. Madotto, and P. Fung. Survey of hallucination in natu- ral language generation.ACM Computing Surveys, 55(12):1–38, 2023

    2023

  6. [14]

    Kumar, N

    A. Kumar, N. Patel, M. Rodriguez, and J. Thompson. Measurement gaps in pro- duction AI: When evaluation frameworks fail to detect system degradation.arXiv preprint arXiv:2607.09999, 2025

    2025 arXiv

  7. [15]

    W. Kwon, Z. Li, S. Zhuang, Y . Sheng, L. Zheng, C. H. Yu, J. E. Gonzalez, H. Zhang, and I. Stoica. Efficient mem- ory management for large language model serving with PagedAttention, 2023

    2023

  8. [16]

    Liang, R

    P. Liang, R. Bommasani, T. Lee, D. Tsipras, D. Soylu, M. Yasunaga, Y . Zhang, D. Narayanan, Y . Wu, A. Kumar, et al. Holistic evaluation of language models.arXiv preprint arXiv:2211.09110, 2022

    2022 arXiv

  9. [17]

    X. Liu, H. Chen, M. Yang, and J. Zhou. The blind spot problem: Characteris- ing undetectable failures in deployed lan- guage model systems.arXiv preprint arXiv:2606.09863, 2025

    2025 arXiv

  10. [18]

    McGregor

    S. McGregor. Preventing repeated real world AI failures by cataloging incidents: The AI incident database. InProceedings of the AAAI Workshop on Investigating and Preventing AI Safety Concerns, 2021

    2021

  11. [19]

    Ouyang, J

    L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. L. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray, et al. Train- ing language models to follow instructions with human feedback.Advances in Neu- ral Information Processing Systems, 35: 27730–27744, 2022

    2022

  12. [20]

    Paleyes, R.-G

    A. Paleyes, R.-G. Urma, and N. D. Lawrence. Challenges in deploying ma- chine learning: A survey of case stud- ies.ACM Computing Surveys, 55(6):1–29, 2022. 16

    2022

  13. [21]

    M. Pandey. Evaluating agentic AI in the wild: Failure modes, drift patterns, and a production evaluation framework, May

  14. [22]

    Perez and I

    F. Perez and I. Ribeiro. Ignore previous prompt: Attack techniques for language models. InProceedings of the Workshop on Trustworthy NLP (TrustNLP), 2022

    2022

  15. [23]

    Post office Horizon IT inquiry: In- terim report

    Post Office Horizon IT Inquiry. Post office Horizon IT inquiry: In- terim report. https://www. postofficehorizoninquiry. org.uk/, 2024

    2024

  16. [24]

    M. T. Ribeiro, T. Wu, C. Guestrin, and S. Singh. CheckList: Beyond accuracy: Behavioral testing of NLP models with CheckList. InProceedings of the 58th An- nual Meeting of the Association for Com- putational Linguistics, pages 4902–4912, 2020

    2020

  17. [25]

    Sculley, G

    D. Sculley, G. Holt, D. Golovin, E. Davy- dov, T. Phillips, D. Ebner, V . Chaudhary, M. Young, J.-F. Crespo, and D. Dennison. Hidden technical debt in machine learning systems. InAdvances in Neural Informa- tion Processing Systems, volume 28, 2015

    2015

  18. [26]

    Shankar, Y

    S. Shankar, Y . Halpern, E. Breck, J. At- wood, J. Wilson, and D. Sculley. Evaluat- ing machine learning systems with miss- ing, noisy, and biased data.arXiv preprint arXiv:2006.05051, 2020

    2006 arXiv

  19. [27]

    Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y . K. Li, Y . Wu, et al. DeepSeekMath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

    2024 arXiv

  20. [28]

    Y . Shen, J. Wang, L. Zhang, and W. Chen. Silent failures in production LLM systems: A taxonomy of undetected deployment fail- ures.arXiv preprint arXiv:2606.14589, 2025

    2025

  21. [29]

    W. Shi, A. Ajith, M. Xia, Y . Huang, D. Liu, T. Blevins, D. Chen, and L. Zettle- moyer. Detecting pretraining data from large language models.arXiv preprint arXiv:2310.16789, 2024

    2024 arXiv

  22. [30]

    Stiennon, L

    N. Stiennon, L. Ouyang, J. Wu, D. M. Ziegler, R. Lowe, C. V oss, A. Radford, D. Amodei, and P. F. Christiano. Learning to summarise with human feedback.Ad- vances in Neural Information Processing Systems, 33:3008–3021, 2020

    2020

  23. [31]

    United States District Court, S.D.N.Y . Mata v. Avianca, Inc., no. 22-cv-1461 (pkc). Sanctions Opinion, June 2023, 2023

    2023

  24. [32]

    A. Wei, N. Haghtalab, and J. Steinhardt. Jailbroken: How does LLM safety training fail?arXiv preprint arXiv:2307.02483, 2023

    2023 arXiv

  25. [33]

    Zheng, W.-L

    L. Zheng, W.-L. Chiang, Y . Sheng, S. Zhuang, Z. Wu, Y . Zhuang, Z. Lin, Z. Li, D. Li, E. P. Xing, et al. Judging LLM-as- a-judge with MT-bench and chatbot arena. Advances in Neural Information Process- ing Systems, 36, 2024. 17 A Full Taxonomy Schema The complete taxonomy def...

    2023

  26. [2026]

    URL https://arxiv.org/ abs/2605.01604

    arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.