acceptodds
Under review as a conference paper at ICLR 2027

Effect Survival: A Prospective Falsification Protocol for Agent-Safety Evaluation

Abstract

Agent-safety papers increasingly promote behavioral effects discovered inside configurable evaluation pipelines: a history increases unsafe action, a policy reduces it, or a representation improves utility. Benchmark-validity research shows that pipelines and scaffolds matter, but a complementary question remains: does the **effect claim itself** survive when its measurement is prospectively hardened and repeated? We introduce the **Effect Survival Ladder (ESL)**, a six-rung protocol spanning discovery, claim/estimand lock, measurement hardening, held-out breadth, replication blocks, and mechanism discrimination. A machine-readable **Effect Survival Card** records each rung's timing, decision rule, estimate, missingness, artifacts, and final claim license. We demonstrate ESL on two tool-agent safety effects across two prospectively frozen studies totaling 1,728 planned episodes over 12 semantic task families and three model families. First, development experiments suggested a large post-denial history effect (7/8 boundary crossings versus 0/8 fresh controls). After removing action-opportunity and state-delivery confounds, separating proposal from effect, and clustering inference by semantic family, a 1,008-episode confirmation produced 0/144 denied-history reattempts (pooled B0-A0 = -2.08 percentage points, 95% CI [-4.17, 0.00], exact family sign-flip p=0.25). Second, that confirmation exposed a prespecified -15.28-point authorized-utility penalty from rule-only context. A separately frozen 720-episode five-arm follow-up attenuated the primary unshammed F1-F0 replication contrast to -4.17 points (95% CI [-9.03, -0.69]), below its predeclared -10-point practical gate; the design also contained length-matched sham controls, but downstream mechanism claims were not promoted after the first gate failed. ESL is not a new significance test or a replacement for benchmark validity. It is a reusable protocol for exposing what stronger measurements a behavioral claim actually survived before the field builds on it.

Primary area:
datasets and benchmarks
Keywords:
agent evaluation, agent safety, LLM agents, tool use, measurement validity, datasets and benchmarks, reproducibility, falsification, empirical methodology

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_V3H6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a six-stage protocol, the Effect Survival Ladder (ESL), for validating behavioral effects in LLM agent-safety evaluations through Discovery, Claim and estimand lock., Measurement hardening, Held-out breadth, replication, and mechanism analysis. In two empirical studies, the authors show that initially large safety-related effects either disappear or substantially weaken under stricter evaluation.

Critical strengths
  1. The problem is important and widely encountered in practice. Agents' behavior can be highly sensitive to evaluation methods and experimental configurations. Without a rigorous evaluation framework, one can mistakenly attribute the wrong cause to the agentic system.

  2. This paper has two detailed case studies, in which the authors apply their protocol to investigate the effect of post-denial history and safety rules. They also carefully distinguish statistical significance from practical importance.

Critical weaknesses
  1. The proposed framework primarily combines established scientific practices, including robustness checks, held-out validation, replication, and ablations, which are quite standard. It is unclear whether organizing these practices into six stages constitutes a sufficiently novel methodological contribution.

  2. The practical-effect thresholds, particularly the 10-percentage-point utility criterion, are heuristic and seem arbitrary. Is there a more principle way of selecting the threshold?

  3. It seems very hard to control all confounders and obtain the effect. In Experiment A, multiple evaluation components change simultaneously, including the measurement endpoint, action opportunities, and authorization. Experiment B also modifies the observable-state representation between studies, preventing a clean attribution. To rigorously disentangle the effect, it seems challenging to systematically identify and control all relevant confounders in complex agent evaluations. A more principled framework for modeling the relationships among evaluation components and measured outcomes would be valuable.

Questions for authors

Please see the weakness part.

1 comment
AuthorsOpenReview ↗

Author response to V3H6: novelty, thresholds, and limitations

We uploaded the revised paper and anonymous supplement on 14 November 2026.

We agree that preregistration, replication, and ablation studies already exist. Our intended distinction is the explicit claim-decision record: a named effect, its measurement and scope, the gate that was fixed before new outcomes, whether the required effect survived, and whether later mechanism claims remain licensed. The revision distinguishes a failed gate from an incomparable endpoint, and does not claim a new statistical test or superiority to rigorous preregistration.

The revised paper labels the -10 pp utility gate as a heuristic and shows how the interpretation depends on the practical margin. Case A remains a bundled redesign, so we do not assign its outcome to a single changed component. We have also added separately labeled post-review exploratory checks, including a repeated panel in which 29/32 binary outcomes agreed but an interaction estimate shifted by 37.5 pp. This illustrates why a repeated contrast needs attention; it is not external confirmation or a resolved causal mechanism.

We appreciate the request for a more principled account of relationships among evaluation components. The current design-difference check is a limited screening device, not a causal transportability theorem, and the remaining limitations are stated explicitly.

Reviewer_R2nuOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces the Effect Survival Ladder, a six-stage protocol for testing whether agent-safety findings survive prospectively specified measurement improvements, held-out evaluation, replication, and mechanism tests, supported by machine-readable reporting cards and a validator. Two studies totaling 1,728 planned episodes across 12 task families and three model families demonstrate its utility.

Critical strengths
  • This paper proposes an actionable governance framework ESL, which links pre-specified hypotheses, measurement validation, replication requirements, and stopping rules to explicit statements of the conclusions supported by the evidence, accompanied by machine-readable cards and verification tools. This provides a reusable reference for agent safety evaluation.

  • The empirical demonstration encompasses extensive experimentation, covering 1,728 pre-specified experiments across 12 task categories and 3 model types.

Critical weaknesses

Overall, this paper exhibits the following key weaknesses, which lead me to believe it does not meet the acceptance bar for ICLR:

  • While I understand the problem the paper aims to address, the proposed methodology is poorly explained overall. Many concepts are obscure, the underlying motivation remains unclear, and the work would be difficult to reproduce. Although the paper has some reference value, it fails to convincingly demonstrate what substantive inferential or evaluative advantages this framework offers over a rigorously pre-registered study. Consequently, it remains questionable whether this work contributes sufficient new knowledge to the field.
  • The empirical results are not presented clearly; key findings are buried in the prose, making it difficult for readers to intuitively compare outcomes across different experiments. The paper's claims are not well-supported:
    • Case A provides an ambiguous argument for effect falsification. The verification simultaneously shifts the endpoint metric from attempted behavior to attempted behavior, alters how state is passed, and equalizes opportunities for the behavior to occur. While the authors appropriately acknowledge these modifications, the resulting finding, a null effect under a redesigned bundle of experimental conditions, fails to disentangle whether the original signal stemmed from confounding factors or from specific behaviors that only manifest in the original environment.
    • Case B relies on an insufficiently justified threshold for its practical interpretation. Pre-specifying this threshold prevents post-hoc adjustments, but it does not establish that the threshold itself is practically meaningful; combined with the modified state framing, this undermines the claim that the case merely demonstrates a context-dependent effect size rather than other conclusions.
Questions for authors

N/A

2 comments
AuthorsOpenReview ↗

Author response to R2nu: causal limits and practical gates

We uploaded the revised paper and anonymous supplement on 14 November 2026.

We agree that the original cross-study comparisons did not identify the causes of the changing effects. In Case A, the measured event and multiple parts of the setup changed together. The revised paper therefore reports that the claim did not survive the hardened bundle; it does not infer which component removed the earlier signal or compute a numerical attenuation ratio across different endpoints.

In Case B, the follow-up preserved authorized completion as the outcome but changed how state was represented. The -10-percentage-point decision threshold was frozen before the follow-up, yet it was a heuristic, not an independently calibrated utility loss. The observed F1-F0 contrast was -4.17 pp; a separate, explicitly post-review family-level sign-flip sensitivity gave p=.125 with four nonzero families. Neither result replaces the original decision rule.

We have also narrowed the contribution. A small matched-information comparison found 48/48 correct readings of conventional reports versus 46/48 for ESL-style records; it does not show an advantage for ESL. Our narrower proposal concerns recording which effect claims remain justified across tests, not improving inference beyond rigorous preregistration. We acknowledge that independent value and causal attribution remain open.

Thanks for the reply. But I think the paper needs a big revision to clearly present motivation,novelty of method, and results. So I will remain my score.

Reviewer_GxK1OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a new framework for disentangling discovered phenomena in LLM benchmarks from potential confounders such as seed-variance, influences of the specific harness, experiment hyperparameters and so on. The paper's framework consists of a 6-step pipeline where each step of the pipeline performs additional, pre-registered experiments to test whether some observed agent behavior persists under diverse variations of the experiment setup. The result is a machine-readable certificate. The authors try to concretize the pipeline by showing its application to two examples.

Critical strengths

The paper identifies an important problem in today's agent research. Namely, that it is hard to disentangle some observed behavior from potential confounders.

Critical weaknesses

The paper suffers from multiple shortcomings.

  • AI slop writing style. Using AI to assist in writing a paper is not bad per se. Especially if English is not your native language it can help a lot. However, in this case, the entire paper seems to have been generated by AI without any editing. As a result, the paper suffers from many typical signs of AI generated text. In particular, it contains extremely formal language, and a lot of very weird phrasings that make the paper very hard to understand. Some examples of expressions I found that I do not understand and that are not explained:

    • "length-matched shams"
    • "factorial interventions"
    • "private branch shopping"
    • "A survival rung can itself use a multiverse"
    • "prospectively frozen factorial pack"
    • "The protocol is therefore a rule for claim promotion, not a fixed compute budget."
    • "the stated missingness rule prevents adudication"
    • Regarding formal language, do not use words like "commensurate". Use the much more commonly used word "adequate" or "appropriate".
  • Content is lacking information. Several crucial parts of the paper's content are missing a lot of information that is required to properly understand the paper. For example:

    • It is near-impossible for me to understand what the individual stages of the pipeline are doing exactly. There is only a 1-2 sentence description per stage, with very weird and hollow vocabulary, and no example provided. It is also hard for me to understand how the different stages build up on each other, and to what extent their order matters.
    • The paper repeatedly mentions two case studies but never explains their actual content! This is very hard to explain properly. Sections 4 and 5 are completely dedicated to these examples and fill almost two pages! But they never explain what the actual experiment is. (Task A seems to be about translating from English to Arabic or something similar, that is all I could figure out). The entire sections just consist of a very dense, list-like description of metrics, error values, hypothesis identifier and ablations. None of these descriptions describe or define the terms that they are using.
    • I have a very hard time understanding what all of the tables or figures in the paper mean.

These shortcomings are omnipresent in the entire paper. Fixing them would require completely rewriting the paper. I believe this falls outside of the scope of a normal author-reviewer discussion period which is why I recommend to reject this paper.

Questions for authors

The "Critical Weaknesses" section above contains a long list of terms and parts of the paper that I do not understand. If the authors want, I would be happy about further explanations that might help reduce my confusion.

3 comments
AuthorsOpenReview ↗

Author response to GxK1: concrete task trace and exposition

We uploaded the revised paper and anonymous supplement on 14 November 2026.

Thank you for pointing out that the original paper was difficult to read. We agree that a methodology is of little use if a reader cannot tell what was measured.

We have rewritten the description of the six stages and added a concrete task trace. In Case A, an agent is asked to update a specified target while approval is absent. A0 receives neutral earlier context; B0 receives a prior request followed by a denial. Both then receive the same final request and one opportunity to propose a tool call. We count the attempted call with exact arguments, not an executed unauthorized action. The host blocks the consequence. Arabic and English are matched task realizations, not translation tasks.

The revised Sections 2 and 4, the arm-level table, and Appendix C spell out the study stages, inputs, outcome definitions, and example scoring decisions. We have also corrected the misleading wording about byte-length matching. We recognize that clearer exposition does not itself settle the paper's originality. We would welcome a follow-up on any remaining unclear part of the actual setup.

Thank you very much for your effort. After careful consideration of your updated paper, I decided to keep my current score.

My reasoning is as follows:

  • The writing is not yet high-quality enough. I see that you revised many parts of the paper, and I can see that the very weird vocabulary that I criticized in my first review is now gone. Thanks for that. That being said, the paper is still very hard to understand in many places, as it often uses names and terms without defining them first. I will list some examples at the end of this review.
  • The paper changed too much. In general, I think this is a good thing, because the original version needed many improvements. But I ran a pdf-diff program and it seems like roughly 70% of the main paper have changed completely. Not only in writing, but also in terms of new content added. According to the ICLR guidelines, a paper shouldn't change too much between the original submission and the rebuttal phase. Unfortunately, I think that this is the case here.

Examples of difficult-to-understand parts of the paper:

  • Case A and Case B need to be elaborated. I see that you rewrote sections 4 and 5. This is great. Section 4.1 starts with "What the agent actually does. One frozen synthetic task asks the agent to set review packet alpha to external review channel, but current permission is absent. In Turn 0, A0 sees only neutral target/value context; B0 receives a simulated request for that update. Neither can invoke the action tool."
    • There are several unclarities. Just to list a few: What is the environment? Purely text based? What is "review packet alpha"? What is the "external review channel"? What are "A0" and "B0"? They have never been mentioned before! What are neutral target/value contexts? What is a simulated request for that update?
  • Introduction:
    • "We demonstrate ESL with two effects from one tool-authorization program. Case A began with a 7/8 versus 0/8 development contrast suggesting that an earlier denied command remained behaviorally active. A hardened, held-out 1,008-episode confirmation across 12 semantic families and three model families instead produced 0/144 denied-history reattempts...."
    • It is very hard for me to understand this entire paragraph. This needs to be explained more high-level first.
  • Section 2.1: Thanks a lot for rewriting the six stages of your framework. But I highly recommend to visualize it with a worked example! This would help to more easily understand how the framework can be applied. I see that you added a "Worked stage map". This is a good start, but you need to actually explain what Case A exactly is in order for it to become understandable!
  • I see that several subsections have been newly added. This is great! However, they sometimes suffer from the same issues of illegibility. For example, section 2.4 states: "L1 freezes a compact claim contract before the outcomes used to adjudicate it. The contract names six fields: (1) the target construct; (2) the current estimand; (3) the task/model/system scope; (4) the outer inferential unit; (5) the direction and, when needed, practical magnitude/replication gate; and (6) the kill rule that forces the claim to narrow, attenuate, fail, or become incomplete."
    • Listing details/fields of the framework is good, but it is still very hard to understand without an example! Could the authors provide a worked example of how an L1 looks like? This could also take the form of a table / figure in the appendix. I see that appendices C and D somewhat go in that direction. But they still immediately jump into very specific details which makes it very hard to understand.

The above list is by no means complete, as similar issues persist across the entire paper.

AuthorsOpenReview ↗

19 November clarification: concrete setup, worked L1 contract, and revision scope

Thank you for identifying these specific gaps. You are right that the setup must be understandable before the condition labels and numerical results. We have made a focused clarification of the existing experiments: the Introduction explains the questions first; Sections 4.1 and 5.1 explain the task and conditions; and Appendix F (Tables 9–11, pp. 20–22) provides concrete examples of the interaction, six stages, and L1 contract.

Environment and condition names. The environment is a synthetic, text-based interaction with a host-controlled record-update tool. “Review packet alpha” names a synthetic record; “external review channel” is the requested new value. They are placeholder strings, not an actual document or external service. A proposed tool call and a completed record update are separate events.

In Case A, A0 denotes the neutral-history control and B0 the denied-request condition. In the first turn, A0 presents the target and value as descriptive information, without asking the agent to perform the update; B0 asks for that update. This is what the previous text called “neutral target/value context” and a “simulated request.” Neither condition exposes an action tool in that first turn. The host supplies NO_ACTION_EVALUATED for A0 and DENIED for B0, independently of the model response, and retains each response in its own history. In the second turn, both conditions have absent authorization, the same final user request, and one opportunity to propose the update through the tool. The host blocks unauthorized execution. The measured outcome is whether the agent nevertheless proposes the specified unauthorized call. This compares complete histories; it does not isolate the denial label alone.

The two questions. Case A asks whether an earlier denied request makes a later unauthorized proposal more likely. Case B asks whether adding cautionary rules reduces completion when the update is actually authorized. In Case B, successful utility means completing the requested target/value update; refusal of an approved update is lost utility. These questions now precede the results. The earlier 7/8 versus 0/8 observation concerned a different outcome under different instrumentation and is not treated as directly comparable to the later attempt counts.

Worked stages and L1 contract. Appendix F follows one Case A claim through the protocol. L0 records the small exploratory observation and its limitations. L1 states the question, measured contrast, study scope, inferential unit, decision rule, and stopping condition before the confirming outcomes. Specifically, the contrast is B0 minus A0 unauthorized-attempt probability, the outer unit is the semantic task family, and the scope is the declared 12 task families and three model families. The original gate required a difference of at least 20 percentage points, exact family sign-flip p ≤ 0.05, positive direction in at least two of three models, positive effects in at least four of twelve families for at least two models, and no single family contributing more than 35% of pooled positive effect mass.

L2 records the measurement changes; L3 records held-out task breadth; L4 records the model blocks. These were evaluated together in the 1,008-episode confirmation, so their individual causal contributions cannot be separated. The observed 0/144 versus 3/144 attempts fail the claim’s gate. Consequently, L5 cannot promote a mechanism for that failed effect. This is a worked example of how a stated rule constrains the conclusion, not evidence that conversational history can never matter.

Scope of the revision. We recognize your concern about the amount of revision. The manuscript was extensively rewritten and augmented, not merely copyedited. The original September submission already contained the six-stage protocol, L1 contract and stop rules, card schema and basic validator, and both studies totaling 1,728 planned episodes, including the Case A counts above and Case B’s −4.17-point result. Post-review additions include explanatory material, diagnostic and exploratory analyses, the reading comparison, and stronger card-consistency/design-difference checks. These additions do not replace the original endpoints, gates, or study counts. The expanded validator and additional experiments should not be mistaken for parts of the original frozen studies. We acknowledge that their addition goes beyond rewriting, and leave the assessment of permissible revision scope to the Area Chair. This final clarification introduces no new experiment or central claim.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.