Pith. sign in

REVIEW 2 major objections 6 minor 41 references

Interventional interpretability scores can be rewritten as bounded causal means and certified with anytime-valid sequences that stay valid under peeking and adaptive sampling.

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 · grok-4.5

2026-07-10 08:56 UTC pith:3XL7J5H4

load-bearing objection Solid verification layer: standard anytime-valid CS tools packaged as explicit causal estimands for mech-interp interventions, with real sample savings and honest negatives. the 2 major comments →

arxiv 2607.08349 v1 pith:3XL7J5H4 submitted 2026-07-09 cs.LG

Certified Interventional Fidelity: Anytime-Valid, Adaptive Evaluation of Causal Claims in Mechanistic Interpretability

Amir Asiaee This is my paper
classification cs.LG
keywords mechanistic interpretabilityinterventional fidelityconfidence sequencesanytime-valid inferenceactivation patchingcausal abstractionimportance samplingbetting sequences
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.

When researchers intervene on neural networks—patching activations, ablating components, or testing a compressed model—they usually report a single point estimate, even though they often monitor results and steer later tests toward suspected failures. That practice makes it hard to know whether a fidelity or recovery number is a stable causal claim or an artifact of finite samples and adaptive choices. This paper introduces Certified Interventional Fidelity (CIF): rewrite the reported quantity as an expectation of a bounded score over a stated input distribution and a stated intervention distribution, then attach confidence intervals and anytime-valid confidence sequences to that mean. The sequences remain valid under repeated monitoring and data-dependent stopping, and bounded mixture importance weighting lets the sampler chase failures without changing the target claim. Variance-adaptive betting sequences cut the forward-pass cost of certification by roughly 10–30× versus simple range-based bounds on MNIST abstractions and GPT-2 Small IOI circuits, certify high-fidelity claims, and show when apparent method differences are not statistically supported.

Core claim

Abstraction/reduction fidelity and component-effect recovery both reduce to the same statistical object: a bounded interventional score whose population mean is taken over a declared input distribution and intervention distribution. CIF certifies that mean with finite-sample confidence intervals and anytime-valid confidence sequences, including under adaptive intervention sampling via bounded mixture importance weighting, so the certified claim survives monitoring, early stopping, and failure-directed reallocation of tests.

What carries the argument

Certified Interventional Fidelity (CIF): the reported metric is the mean θ = E[Y] of a score Y in [0,1] under D × Π; Hoeffding-style and Online-Newton-Step betting confidence sequences bound θ at every sample size, and adaptive proposals are mixed with Π so importance weights stay bounded and the estimand does not change.

Load-bearing premise

The guarantees apply only after scores are forced into a fixed bounded range; if clipping changes what the score means, the certificate is for the clipped version, not the raw unbounded effect.

What would settle it

Repeat the coverage check with known ground-truth mean under aggressive peeking and adaptive proposals: if betting-sequence coverage falls well below the nominal level, or if the 10–30× sample-cost reduction vanishes when score variance is higher, the practical certification claims fail.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • A fidelity or patching threshold can be certified with a data-dependent stopping time that remains valid under peeking.
  • Point-estimate rankings of abstractions or circuits can be overturned once intervals show differences are unsupported.
  • Sensitivity of conclusions to the intervention distribution (for example swap probability) becomes a required, visible part of the claim.
  • Betting sequences make certification practical for full-model forward passes on models the size of GPT-2 Small.
  • Failure-directed adaptive sampling can hunt counterexamples without biasing the certified population mean, provided mixture importance weights are used.

Where Pith is reading between the lines

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

  • Automated circuit-discovery loops could use one-sided CIF stopping rules as early-pruning controllers, moving budget away from low-fidelity candidates mid-search.
  • The same bounded-mean sequence layer could wrap other sequential interpretability scores (scrubbing degradation, multi-layer abstraction fidelity) once they are written as expectations under an explicit intervention law.
  • Community reporting may need fixed default clip ranges so certificates remain comparable when different papers bound the same underlying effect differently.
  • Replacing Bonferroni with e-value FDR control inside CIF would make large per-head or per-edge scans less wasteful while keeping selection honest.

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

2 major / 6 minor

Summary. The paper introduces Certified Interventional Fidelity (CIF), a statistical layer that reformulates common interventional mechanistic-interpretability evaluations (abstraction/reduction fidelity and component-effect recovery) as bounded causal estimands θ = E_{(X,I)~D×Π}[Y] over declared input and intervention distributions. It supplies fixed-budget intervals and anytime-valid confidence sequences for θ under i.i.d. sampling and under adaptive mixture importance sampling with bounded weights, instantiated with Hoeffding-style sequences and variance-adaptive ONS betting sequences. Algorithms cover two-sided tracking, one-sided certification, paired comparison, and a reporting checklist. Experiments on MNIST neural abstractions under interchange interventions and GPT-2 Small IOI circuit patching show that betting sequences cut certification cost by roughly 10–30× versus Hoeffding, that some apparent method differences are not supported once uncertainty is reported, and that conclusions can depend on the intervention distribution Π and the discrepancy metric.

Significance. If the results hold as stated, CIF is a practically useful verification layer for a community that currently reports point estimates under sequential and adaptive workflows. The contribution is less a new concentration inequality than a clean reformulation that makes estimands, monitoring, and adaptivity explicit, with standard tools (Hoeffding/Azuma spending, Ville betting capital processes, Horvitz–Thompson mixture weights) correctly applied. Strengths include short matching proofs in Appendix C, a reproducible code release, honest negative results (no high-fidelity certification under aggressive swaps; adaptive sampling often slows high-fidelity certification), and explicit sensitivity analysis over Π and metrics. That combination is valuable even though the statistical machinery is largely off-the-shelf.

major comments (2)
  1. Section 3 frames four tasks (P1)–(P4), including selection and multiple comparisons across many components or abstractions. Section 3.6 only sketches Bonferroni, peeling, and e-value FDR, and the main experiments (E1–E2, Tables 1–3) do not demonstrate corrected multi-hypothesis certification on a realistic scan. The single-estimand CS results remain sound, but the paper’s claim to address (P4) as part of the CIF layer is currently under-supported relative to (P1)–(P3). Either add a small multi-candidate experiment with corrected CSs, or narrow the framing so (P4) is clearly future work rather than a delivered capability.
  2. Sections 2.4 and 3.1–3.4 (and the Discussion) state that certificates apply to bounded or clipped scores Y ∈ [0,1]. For IOI, Δ is defined via clip[0,1] of a normalized logit-gap recovery. That is methodologically fine, but the manuscript should state more sharply in the main text (not only Discussion) that the certified object is the clipped estimand, and that clipping can change ranking or effect size relative to raw logit differences. A short numerical check—e.g., how often clipping binds and whether μ rankings change without clip—would make the scope of the GPT-2 claims precise without altering the theorems.
minor comments (6)
  1. Abstract and Introduction attribute a 10–30× cost reduction to betting sequences; adaptive sampling is a separate feature and Appendix F shows it often increases n for high-fidelity certification. A single clarifying sentence would prevent readers from attributing the speedup to adaptivity.
  2. Figure 1 caption and §4.1: the adaptive proposal is described as weight-norm based with α=0.3; briefly state whether the same α and proposal are used in all adaptive panels of Figure 1 for reproducibility of the plotted curves.
  3. Table 1 reports estimated fidelity at p=0.5 with Hoeffding half-width 0.046 at n=5000; consider also reporting betting intervals in that table for consistency with Table 2 and Figure 2.
  4. Notation: Z is used both for the fidelity discrepancy (Eq. 1) and later as a generic score in places; a short notation table or consistent use of Y for the generic bounded score would help.
  5. Appendix A related work is thorough; a one-sentence pointer in the main text to Méloux et al. (2025) and Karampatziakis et al. (2021) as the closest statistical predecessors would help readers who skip the appendix.
  6. Algorithm 1 implements the Hoeffding adaptive radius; a parallel note or algorithm line for the betting tracker (as promised in §3.4) would make the “interchangeable evaluation layers” claim operational.

Circularity Check

0 steps flagged

No significant circularity: CIF declares estimands as expectations and applies standard external concentration tools; self-citation is only experimental scaffolding.

full rationale

The paper’s load-bearing chain is: (i) write interventional scores as θ = E_{(X,I)~D×Π}[Y] with Y∈[0,1]; (ii) apply Hoeffding/Azuma spending and ONS betting confidence sequences (Howard et al. 2021; Waudby-Smith & Ramdas 2024); (iii) preserve the target under adaptive sampling via mixture Horvitz–Thompson weights with W_max = 1/(1−α) (Lemma 1, Theorem 2). None of these steps defines the estimand in terms of a fitted parameter that is later “predicted,” nor do they import a uniqueness theorem or ansatz from the author’s prior work as a forced mathematical fact. Asiaee (2026) is cited for the MNIST abstraction construction and mechanism-reduction setting, not for the statistical guarantees; those rest on external, independently published concentration and off-policy CS results. Experiments report empirical certification costs and coverage under declared (D, Π, score) triples; they do not close a definitional loop. Bounded/clipped scores are an explicit scope limitation, not a circular reduction. Score 0 is therefore appropriate.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

Statistical validity rests on classical concentration and martingale tools plus the modeling choice that the scientific claim is a mean of a bounded interventional score under declared D and Π. Free parameters are design knobs (mixture rate, betting step sizes, certification thresholds), not fitted physical constants. No new physical entities are postulated; CIF is a method layer.

free parameters (4)
  • mixture rate α = 0.3 (main runs); swept in {0,0.1,...,0.7}
    Controls adaptive vs target mass in qt=(1-α)Π+α q̃t; experiments use α=0.3 and sweep α. Affects efficiency and Wmax, not the estimand if weights are used.
  • ONS betting hyperparameters η, c, λ clip = η=2/(2-log 3)≈2.22, c=1/2
    Online Newton Step betting uses η≈2.22, c=1/2, λ1=0, A0=1 as in Waudby-Smith & Ramdas; these are fixed algorithmic constants that affect tightness, not the target θ.
  • error-spending schedule δn=6δ/(π² n²) = 6δ/(π² n²)
    Standard summable allocation for union-bound Hoeffding CS; choice affects finite-n width, not asymptotic validity.
  • certification thresholds F0 / μ0 and clip range for Δ = F0, μ0 ∈ {0.90,0.95,0.99}; Δ clipped to [0,1]
    User-chosen scientific thresholds (0.90, 0.95, 0.99) and [0,1] clipping of patching recovery; certificates apply to the clipped score.
axioms (5)
  • standard math Hoeffding / Hoeffding–Azuma concentration for bounded i.i.d. or martingale-difference variables in [0,1] or [0,Wmax].
    Proposition 1, Theorems 1–2; Appendix C.
  • standard math Ville’s inequality for nonnegative supermartingales / capital processes yields anytime-valid CS for betting methods.
    Section 3.4; capital process Kn(m).
  • domain assumption Interventional scores Y are bounded in [0,1] (or rescaled after clipping), so range-based and betting trackers apply as stated.
    Sections 2.4, 3, Discussion; load-bearing for stated finite-sample guarantees without sub-Gaussian extensions.
  • domain assumption The scientific claim of interest is the population mean under a declared input distribution D and intervention distribution Π (estimation, not identification of the 'true' mechanism).
    Introduction and Section 5; CIF certifies the estimand, not that the metric equals mechanistic truth.
  • domain assumption Adaptive proposals qt are Ht−1-measurable and mixture-supported so importance weights wt=Π(It)/qt(It) ≤ 1/(1−α).
    Section 3.3; required for unbiasedness (Lemma 1) and bounded-weight CS (Theorem 2).
invented entities (1)
  • Certified Interventional Fidelity (CIF) framework independent evidence
    purpose: Name the statistical layer that maps interventional interpretability scores to causal estimands plus anytime-valid CS under optional adaptive sampling.
    Methodological packaging, not a latent physical object; independent evidence is the theorems plus experiments, not an external detection claim.

pith-pipeline@v1.1.0-grok45 · 24766 in / 3614 out tokens · 107382 ms · 2026-07-10T08:56:01.269857+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Certified Interventional Fidelity: Anytime-Valid, Adaptive Evaluation of Causal Claims in Mechanistic Interpretability." pith.science (2026). https://pith.science/paper/3XL7J5H4

@misc{pith2026260708349,
  author       = {Pith},
  title        = {Pith review of: Certified Interventional Fidelity: Anytime-Valid, Adaptive Evaluation of Causal Claims in Mechanistic Interpretability},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3XL7J5H4}},
  note         = {Machine review of arXiv:2607.08349}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Mechanistic interpretability often evaluates explanations by intervening on a model: swapping hidden states, patching activations, ablating components, or comparing a compressed model to the original one. These experiments are usually summarized by a point estimate, even though the evaluation may be monitored while it runs or adapted toward suspected failures. This makes it hard to tell whether a reported fidelity or patching effect is a stable causal claim or a consequence of finite sampling and evaluation choices. We introduce Certified Interventional Fidelity (CIF), a statistical layer for interventional interpretability evaluations. CIF first writes the quantity being reported as a causal estimand: an expectation of a bounded score over a stated input distribution and a stated intervention distribution. It then provides confidence intervals and anytime-valid confidence sequences for this estimand, including under adaptive intervention sampling via bounded mixture importance weighting. We instantiate CIF with Hoeffding-style sequences and variance-adaptive betting sequences, the latter reducing certification cost by 10-30x in our experiments. On MNIST abstractions and GPT-2 Small IOI circuits, CIF certifies high-fidelity claims, shows when apparent method differences are not statistically supported, and makes sensitivity to the intervention distribution explicit.

Figures

Figures reproduced from arXiv: 2607.08349 by Amir Asiaee.

Figure 1
Figure 1. Figure 1: Confidence-sequence width versus forward passes [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Certification-cost ratio nHoeffding/nBetting for each method at k= 256, p= 0.1, under i.i.d. sampling. Each run stops when the lower confidence bound on F exceeds the target F0. Hard pruning is omitted because it never certifies; dotted reference lines mark 10× and 30×. At the aggressive swap probability p = 0.5, no non-identity abstraction reaches the certification threshold F ≥ 0.90 ( [PITH_FULL_IMAGE:f… view at source ↗
Figure 3
Figure 3. Figure 3: Confidence sequences for patching recovery [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Sensitivity of certified fidelity to swap probability [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Stopping times across five matched seeds per sampler (boxes: IQR; dots: individual seeds). Left: low-fidelity [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Median certification cost versus mixture rate [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗

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

41 extracted references · 41 canonical work pages · 3 internal anchors

  1. [1]

    2009 , publisher =

    Causality: Models, Reasoning, and Inference , author =. 2009 , publisher =

    work page 2009

  2. [2]

    Distill , year =

    Olah, Chris and Cammarata, Nick and Schubert, Ludwig and Goh, Gabriel and Petrov, Michael and Carter, Shan , title =. Distill , year =

    work page

  3. [3]

    Advances in Neural Information Processing Systems , year =

    Geiger, Atticus and Lu, Hanson and Icard, Thomas and Potts, Christopher , title =. Advances in Neural Information Processing Systems , year =

    work page

  4. [4]

    Proceedings of the 39th International Conference on Machine Learning , pages =

    Inducing Causal Structure for Interpretable Neural Networks , author =. Proceedings of the 39th International Conference on Machine Learning , pages =. 2022 , editor =

    work page 2022

  5. [5]

    Proceedings of the Third Conference on Causal Learning and Reasoning , pages =

    Finding Alignments Between Interpretable Causal Variables and Distributed Neural Representations , author =. Proceedings of the Third Conference on Causal Learning and Reasoning , pages =. 2024 , editor =

    work page 2024

  6. [6]

    Journal of Machine Learning Research , year =

    Causal Abstraction: A Theoretical Foundation for Mechanistic Interpretability , author =. Journal of Machine Learning Research , year =

    work page

  7. [7]

    Proceedings of the 33rd Conference on Uncertainty in Artificial Intelligence (UAI) , year =

    Causal Consistency of Structural Equation Models , author =. Proceedings of the 33rd Conference on Uncertainty in Artificial Intelligence (UAI) , year =

    work page

  8. [8]

    Proceedings of the 33rd AAAI Conference on Artificial Intelligence (AAAI-19) , year =

    Abstracting Causal Models , author =. Proceedings of the 33rd AAAI Conference on Artificial Intelligence (AAAI-19) , year =

    work page

  9. [9]

    Proceedings of The 35th Uncertainty in Artificial Intelligence Conference , pages =

    Approximate Causal Abstractions , author =. Proceedings of The 35th Uncertainty in Artificial Intelligence Conference , pages =. 2020 , editor =

    work page 2020

  10. [10]

    Proceedings of the Second Conference on Causal Learning and Reasoning , pages =

    Causal Abstraction with Soft Interventions , author =. Proceedings of the Second Conference on Causal Learning and Reasoning , pages =. 2023 , editor =

    work page 2023

  11. [11]

    2022 , note =

    Causal Scrubbing: a method for rigorously testing interpretability hypotheses , author =. 2022 , note =

    work page 2022

  12. [12]

    Interpretability in the Wild: a Circuit for Indirect Object Identification in

    Wang, Kevin Ro and Variengien, Alexandre and Conmy, Arthur and Shlegeris, Buck and Steinhardt, Jacob , booktitle =. Interpretability in the Wild: a Circuit for Indirect Object Identification in

    work page

  13. [13]

    Advances in Neural Information Processing Systems , year =

    Towards Automated Circuit Discovery for Mechanistic Interpretability , author =. Advances in Neural Information Processing Systems , year =

    work page

  14. [14]

    Localizing Model Behavior with Path Patching

    Localizing Model Behavior with Path Patching , author =. arXiv preprint arXiv:2304.05969 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  15. [15]

    International Conference on Learning Representations , year =

    Towards Best Practices of Activation Patching in Language Models: Metrics and Methods , author =. International Conference on Learning Representations , year =

    work page

  16. [16]

    AtP*: An efficient and scalable method for localizing LLM behaviour to components

    Kram. arXiv preprint arXiv:2403.00745 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  17. [17]

    Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP , year =

    Attribution Patching Outperforms Automated Circuit Discovery , author =. Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP , year =

    work page

  18. [18]

    Locating and Editing Factual Associations in

    Meng, Kevin and Bau, David and Andonian, Alex and Belinkov, Yonatan , booktitle =. Locating and Editing Factual Associations in. 2022 , volume =

    work page 2022

  19. [19]

    The Annals of Statistics , year =

    Time-uniform, nonparametric, nonasymptotic confidence sequences , author =. The Annals of Statistics , year =

    work page

  20. [20]

    The Annals of Statistics , volume =

    Time-uniform central limit theory and asymptotic confidence sequences , author =. The Annals of Statistics , volume =. 2024 , doi =

    work page 2024

  21. [21]

    2026 , eprint=

    Causal Mechanism Reduction: Mechanism Replacement for Neural Network Pruning and Abstraction , author=. 2026 , eprint=

    work page 2026

  22. [22]

    Horvitz, D. G. and Thompson, D. J. , title =. Journal of the American Statistical Association , volume =. 1952 , doi =

    work page 1952

  23. [23]

    First Conference on Language Modeling (COLM) , year =

    Have Faith in Faithfulness: Going Beyond Circuit Overlap When Finding Model Mechanisms , author =. First Conference on Language Modeling (COLM) , year =

    work page

  24. [24]

    Ville, Jean , year =. \'

    work page

  25. [25]

    Game-Theoretic Statistics and Safe Anytime-Valid Inference , journal =

    Ramdas, Aaditya and Gr. Game-Theoretic Statistics and Safe Anytime-Valid Inference , journal =. 2023 , doi =

    work page 2023

  26. [26]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , volume =

    Waudby-Smith, Ian and Ramdas, Aaditya , title =. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume =. 2024 , doi =

    work page 2024

  27. [27]

    Safe Testing , journal =

    Gr. Safe Testing , journal =. 2024 , doi =

    work page 2024

  28. [28]

    The Annals of Statistics , volume =

    Vovk, Vladimir and Wang, Ruodu , title =. The Annals of Statistics , volume =. 2021 , doi =

    work page 2021

  29. [29]

    Journal of the Royal Statistical Society Series A: Statistics in Society , volume =

    Shafer, Glenn , title =. Journal of the Royal Statistical Society Series A: Statistics in Society , volume =. 2021 , doi =

    work page 2021

  30. [30]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , volume =

    Wang, Ruodu and Ramdas, Aaditya , title =. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume =. 2022 , doi =

    work page 2022

  31. [31]

    Advances in Neural Information Processing Systems , volume =

    Vig, Jesse and Gehrmann, Sebastian and Belinkov, Yonatan and Qian, Sharon and Nevo, Daniel and Singer, Yaron and Shieber, Stuart , title =. Advances in Neural Information Processing Systems , volume =

    work page

  32. [32]

    , title =

    Wu, Zhengxuan and Geiger, Atticus and Icard, Thomas and Potts, Christopher and Goodman, Noah D. , title =. Advances in Neural Information Processing Systems , volume =

    work page

  33. [33]

    and Rubin, Donald B

    Imbens, Guido W. and Rubin, Donald B. , title =. 2015 , publisher =

    work page 2015

  34. [34]

    Proceedings of the 38th International Conference on Machine Learning , pages =

    Karampatziakis, Nikos and Mineiro, Paul and Ramdas, Aaditya , title =. Proceedings of the 38th International Conference on Machine Learning , pages =. 2021 , volume =

    work page 2021

  35. [35]

    Mechanistic Interpretability as Statistical Estimation: A Variance Analysis

    Mechanistic Interpretability as Statistical Estimation: A Variance Analysis , author =. arXiv preprint arXiv:2510.00845 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  36. [36]

    Is This the Subspace You Are Looking for?

    Makelov, Aleksandar and Lange, Georg and Nanda, Neel , booktitle =. Is This the Subspace You Are Looking for?

    work page

  37. [37]

    Hypothesis Testing the Circuit Hypothesis in

    Shi, Claudia and Beltran-Velez, Nicolas and Nazaret, Achille and Zheng, Carolina and Garriga-Alonso, Adri. Hypothesis Testing the Circuit Hypothesis in. Advances in Neural Information Processing Systems , volume =

    work page

  38. [38]

    First Conference on Language Modeling , year =

    Transformer Circuit Faithfulness Metrics are not Robust , author =. First Conference on Language Modeling , year =

    work page

  39. [39]

    Transcoders Find Interpretable

    Dunefsky, Jacob and Chlenski, Philippe and Nanda, Neel , booktitle =. Transcoders Find Interpretable

    work page

  40. [40]

    Advances in Neural Information Processing Systems , year =

    Compact Proofs of Model Performance via Mechanistic Interpretability , author =. Advances in Neural Information Processing Systems , year =

    work page

  41. [41]

    Journal of the American Statistical Association , year =

    Probability Inequalities for Sums of Bounded Random Variables , author =. Journal of the American Statistical Association , year =

    work page