pith. sign in

arxiv: 2607.01254 · v1 · pith:PUEW2YYLnew · submitted 2026-06-04 · 💻 cs.CY · econ.GN· q-fin.EC

The Benchmark Ceiling: Human Judgment, Evaluation Scarcity, and the Political Economy of AI Capability Measurement

Mark Esposito , Liu Zhang , Ali Ansari This is my paper

Pith reviewed 2026-07-04 00:15 UTC · model grok-4.3

classification 💻 cs.CY econ.GNq-fin.EC
keywords benchmark ceiling problemAI evaluationhuman judgment scarcitybenchmark validityevaluation signalpolitical economy of AIcapability measurement
0
0 comments X

The pith

As AI models saturate easy benchmark items, valid measurement signal concentrates in scarce hard-tail items authored by elite experts.

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

The paper argues that benchmark validity depends on the quality of human judgment embedded in their construction, and this quality is structurally scarce. As foundation models reach ceiling performance on existing suites, discriminating signal concentrates in the hardest items, which require elite expert judgment to design. A formal model shows that benchmark scores depreciate as measurement instruments when capability rises or contamination occurs, with valid signal concentrating in hard-tail items whose replacement cost rises convexly. Platform data from over one thousand credentialed professionals documents a scarcity premium for high-judgment evaluation labor. The argument concludes with political economy and governance implications for how AI capability is measured and controlled.

Core claim

Benchmarks are public signals of latent model quality whose precision depends endogenously on benchmark validity. As frontier capability rises and contamination or strategic optimization increases, fixed benchmarks depreciate, and valid signal concentrates in hard-tail items whose replacement cost rises convexly with capability. Private benchmark producers underinvest in validity relative to the social optimum.

What carries the argument

The formal model of benchmark signal depreciation, which shows that valid signal concentrates in hard-tail items and that their replacement cost rises convexly with frontier capability.

If this is right

  • Valid signal concentrates in hard-tail items authored by a thin stratum of highly expert evaluators.
  • The replacement cost of such high-validity items rises convexly with frontier capability.
  • Private benchmark producers underinvest in validity relative to the social optimum.
  • High-judgment, low-codifiability evaluation labor commands a scarcity premium in professional markets.

Where Pith is reading between the lines

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

  • Scarcity of elite evaluators may create a bottleneck that slows the overall rate of new benchmark development.
  • Entities controlling access to credentialed professionals could gain disproportionate influence over AI capability assessments.
  • Governance proposals for frontier AI may need to address allocation or subsidization of expert judgment resources directly.

Load-bearing premise

The quality of human judgment embedded in benchmark construction is structurally scarce in ways that standard scaling narratives obscure, and private benchmark producers underinvest in validity relative to the social optimum.

What would settle it

Data showing that the marginal cost or time to produce new high-validity benchmark items remains constant or falls as model capabilities advance would falsify the convex rise in replacement costs.

read the original abstract

Benchmarks are the primary instruments through which AI capability is measured, compared, and governed. This paper argues that the validity of frontier AI benchmarks is a function of the quality of human judgment embedded in their construction, and that this quality is structurally scarce in ways that standard scaling narratives obscure. As foundation models approach ceiling performance on existing evaluation suites, discriminating signal concentrates in the hardest benchmark items, precisely those requiring elite expert judgment to design. We term this the benchmark ceiling problem: the progressive exhaustion of evaluation signal as models saturate the easy majority of items while the difficult tail, authored by a thin stratum of highly expert evaluators, remains the only source of genuine discrimination. The paper develops this argument in three steps. First, we present a formal model of benchmark signal depreciation. Benchmark scores are public signals of latent model quality, but their precision depends endogenously on benchmark validity. As frontier capability rises and as contamination or strategic optimization increases, fixed benchmarks depreciate as measurement instruments. The model shows that valid signal concentrates in hard-tail items, that the replacement cost of such items rises convexly with frontier capability, and that private benchmark producers underinvest in validity relative to the social optimum. Second, drawing on platform data from micro1 covering over one thousand credentialed professionals, we document the scarcity premium associated with high-judgment, low-codifiability evaluation labor. Third, we develop the political economy and governance implications.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper claims that as foundation models saturate easy benchmark items, discriminating signal concentrates in hard-tail items requiring scarce elite expert judgment to design, creating a 'benchmark ceiling.' It develops this via (1) a formal model of endogenous benchmark signal depreciation showing convex replacement costs for hard items and underinvestment by private producers relative to the social optimum, (2) platform data from micro1 on over 1,000 credentialed professionals documenting a scarcity premium for high-judgment evaluation labor, and (3) political-economy and governance implications.

Significance. If the formal model and data hold, the paper would usefully highlight a structural constraint on AI capability measurement that scaling narratives overlook, with downstream effects on how benchmarks are produced and governed. The combination of modeling, empirical labor-market evidence, and policy discussion is a strength if the model is made verifiable.

major comments (2)
  1. [Formal model (post-abstract)] The formal model of benchmark signal depreciation (described in the abstract and section outline as showing convex replacement costs and underinvestment) is asserted without any equations, functional forms, assumptions, or derivations. This is load-bearing for the central claim, as it is impossible to assess whether the convexity and externality results follow from the premises or are definitional.
  2. [Empirical platform data section] The empirical claim of a scarcity premium for high-judgment labor rests on platform data from micro1 covering over one thousand professionals, but no details on sampling, measurement of 'high-judgment' tasks, controls, or statistical results are provided, preventing evaluation of whether the data supports the underinvestment conclusion.
minor comments (1)
  1. [Data section] Clarify the identity and selection criteria of the 'micro1' platform if it is not a standard public dataset.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for these constructive comments, which identify places where the manuscript requires greater transparency to allow proper evaluation of its core claims. We address each point below and will incorporate the requested details in revision.

read point-by-point responses

  1. Referee: [Formal model (post-abstract)] The formal model of benchmark signal depreciation (described in the abstract and section outline as showing convex replacement costs and underinvestment) is asserted without any equations, functional forms, assumptions, or derivations. This is load-bearing for the central claim, as it is impossible to assess whether the convexity and externality results follow from the premises or are definitional.

    Authors: The referee is correct that the manuscript currently describes the model at a conceptual level without presenting the equations, assumptions, or derivations. This omission prevents verification of the claimed results. In the revised manuscript we will add a dedicated formal section that states the model primitives, functional forms, and full derivations, showing explicitly how convex replacement costs and private underinvestment relative to the social optimum follow from the setup. revision: yes

  2. Referee: [Empirical platform data section] The empirical claim of a scarcity premium for high-judgment labor rests on platform data from micro1 covering over one thousand professionals, but no details on sampling, measurement of 'high-judgment' tasks, controls, or statistical results are provided, preventing evaluation of whether the data supports the underinvestment conclusion.

    Authors: We agree that the current text provides insufficient methodological information on the micro1 data. The revised version will expand the empirical section to report the sampling frame, the operational definition and measurement of high-judgment tasks, any covariates or controls employed, and the key statistical results (including coefficients and robustness checks) that document the scarcity premium. revision: yes

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper outlines a formal model of benchmark signal depreciation whose key results (concentration of signal in hard-tail items, convex replacement costs, underinvestment relative to social optimum) are presented as following from stated assumptions on validity, contamination, and externalities. No equations appear in the provided text, so no reduction of any prediction to fitted parameters or definitional relations can be exhibited. The platform-data step on scarcity premium is empirical and independent of the model. The political-economy implications are downstream. No self-citation, uniqueness theorem, or ansatz smuggling is invoked. The derivation chain is therefore self-contained against external benchmarks and receives the default non-finding.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no details on free parameters, axioms, or invented entities used in the formal model of benchmark depreciation.

pith-pipeline@v0.9.1-grok · 5795 in / 1045 out tokens · 33118 ms · 2026-07-04T00:15:39.175780+00:00 · methodology

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

32 extracted references · 32 canonical work pages · 1 internal anchor

  1. [1]

    Handbook of labor economics , volume=

    Skills, tasks and technologies: Implications for employment and earnings , author=. Handbook of labor economics , volume=. 2011 , publisher=

    work page 2011

  2. [2]

    Readings in industrial economics: Volume two: Private enterprise and state intervention , pages=

    Economic welfare and the allocation of resources for invention , author=. Readings in industrial economics: Volume two: Private enterprise and state intervention , pages=. 1962 , publisher=

    work page 1962

  3. [3]

    The Quarterly journal of economics , volume=

    The skill content of recent technological change: An empirical exploration , author=. The Quarterly journal of economics , volume=. 2003 , publisher=

    work page 2003

  4. [4]

    Transactions of the Association for Computational Linguistics , volume=

    Beat the AI: Investigating adversarial human annotation for reading comprehension , author=. Transactions of the Association for Computational Linguistics , volume=. 2020 , publisher=

    work page 2020

  5. [5]

    International conference on machine learning , pages=

    Pythia: A suite for analyzing large language models across training and scaling , author=. International conference on machine learning , pages=. 2023 , organization=

    work page 2023

  6. [6]

    2025 , publisher=

    Item response theory: Foundations for psychologists and social scientists , author=. 2025 , publisher=

    work page 2025

  7. [7]

    arXiv preprint arXiv:2603.00932 , year=

    No Last Mile: A Theory of the Human Data Market , author=. arXiv preprint arXiv:2603.00932 , year=

    work page arXiv

  8. [8]

    The structured premium: Benchmarking expert wages in AI training markets [Working paper] , author=

    work page

  9. [9]

    2025 , doi =

    AI Act Regulation (EU) 2024/1689 – Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence and amending Regulations (EC) No 300/2008, (EU) No 167/2013, (EU) No 168/2013, (EU) 2018/858, (EU) 2018/1139 and (EU) 2019/2144 and Directives 2014/90/EU, (EU) 2016/797 and (EU) ...

    work page 2024

  10. [10]

    2023 , number =

    Executive Order 14110: Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence , author =. 2023 , number =

    work page 2023

  11. [11]

    International Conference on Learning Representations , volume=

    Time travel in llms: Tracing data contamination in large language models , author=. International Conference on Learning Representations , volume=

    work page

  12. [12]

    International Conference on Learning Representations , year=

    Measuring Massive Multitask Language Understanding , author=. International Conference on Learning Representations , year=

    work page

  13. [13]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    Stop uploading test data in plain text: Practical strategies for mitigating data contamination by evaluation benchmarks , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    work page 2023

  14. [14]

    Proceedings of the 2021 conference of the North American chapter of the Association for Computational Linguistics: human language technologies , pages=

    Dynabench: Rethinking benchmarking in NLP , author=. Proceedings of the 2021 conference of the North American chapter of the Association for Computational Linguistics: human language technologies , pages=

    work page 2021

  15. [15]

    Advances in neural information processing systems , volume=

    Superglue: A stickier benchmark for general-purpose language understanding systems , author=. Advances in neural information processing systems , volume=

    work page

  16. [16]

    Holistic Evaluation of Language Models

    Holistic evaluation of language models , author=. arXiv preprint arXiv:2211.09110 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  17. [17]

    Annals of the New York Academy of Sciences , volume=

    Holistic evaluation of language models , author=. Annals of the New York Academy of Sciences , volume=. 2023 , publisher=

    work page 2023

  18. [18]

    Harvard Data Science Review , number=

    Ai transparency in the age of llms: A human-centered research roadmap , author=. Harvard Data Science Review , number=. 2024 , publisher=

    work page 2024

  19. [19]

    2008 , publisher=

    Statistical theories of mental test scores , author=. 2008 , publisher=

    work page 2008

  20. [20]

    Proceedings of the 58th annual meeting of the association for computational linguistics , pages=

    Adversarial NLI: A new benchmark for natural language understanding , author=. Proceedings of the 58th annual meeting of the association for computational linguistics , pages=

    work page

  21. [21]

    1990 , publisher=

    Governing the commons: The evolution of institutions for collective action , author=. 1990 , publisher=

    work page 1990

  22. [22]

    arXiv preprint arXiv:2111.15366 , year=

    AI and the everything in the whole wide world benchmark , author=. arXiv preprint arXiv:2111.15366 , year=

    work page arXiv

  23. [23]

    Proceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society , pages=

    Outsider oversight: Designing a third party audit ecosystem for ai governance , author=. Proceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society , pages=

    work page 2022

  24. [24]

    Findings of the Association for Computational Linguistics: EMNLP 2023 , pages=

    NLP evaluation in trouble: On the need to measure LLM data contamination for each benchmark , author=. Findings of the Association for Computational Linguistics: EMNLP 2023 , pages=

    work page 2023

  25. [25]

    Proceedings of the 1st Workshop on Data Contamination (CONDA) , pages=

    Data contamination report from the 2024 CONDA shared task , author=. Proceedings of the 1st Workshop on Data Contamination (CONDA) , pages=

    work page 2024

  26. [26]

    The review of economics and statistics , volume=

    The pure theory of public expenditure , author=. The review of economics and statistics , volume=. 1954 , publisher=

    work page 1954

  27. [27]

    Transactions on machine learning research , year=

    Beyond the imitation game: Quantifying and extrapolating the capabilities of language models , author=. Transactions on machine learning research , year=

    work page

  28. [28]

    1988 , publisher=

    The theory of industrial organization , author=. 1988 , publisher=

    work page 1988

  29. [29]

    Proceedings of the 2018 EMNLP workshop BlackboxNLP: Analyzing and interpreting neural networks for NLP , pages=

    GLUE: A multi-task benchmark and analysis platform for natural language understanding , author=. Proceedings of the 2018 EMNLP workshop BlackboxNLP: Analyzing and interpreting neural networks for NLP , pages=

    work page 2018

  30. [30]

    Proceedings of the 57th annual meeting of the association for computational linguistics , pages=

    Hellaswag: Can a machine really finish your sentence? , author=. Proceedings of the 57th annual meeting of the association for computational linguistics , pages=

    work page

  31. [31]

    Journal of economic perspectives , volume=

    Towards a political theory of the firm , author=. Journal of economic perspectives , volume=. 2017 , publisher=

    work page 2017

  32. [32]

    2024 , howpublished =

    AI Safety Institute approach to evaluations , author =. 2024 , howpublished =

    work page 2024