acceptodds
Under review as a conference paper at ICLR 2027

Acting for the Right Reasons: Decision-Cause Alignment in Trajectory Planners

Abstract

Vision–language models (VLMs) are increasingly adopted as end-to-end trajectory planners trained by behavior cloning, which aligns actions but not necessarily the reasons behind them and can therefore mask safety-critical failures. Existing methods infer a planner's reasons indirectly from representations, interventions, or behavioral associations. However, such evidence is not the decision-driving event itself; if the inferred reason changes with the diagnostic, it reflects the diagnostic–planner pair rather than a property of the planner. We instead define the *decision-driving event*, the world event that actually drives the planner's response, and call a planner **decision-cause aligned** when this event is the human-relevant event rather than a co-occurring cue. When the human-relevant event and the cue share an onset and response form, static behavior cannot separate their contributions. When the cue appears first, ordinary driving logs can identify its contribution under explicit response-latency and disturbance-isolation assumptions. Building on this result, we introduce **TREAT** (loca**T**e–cu**R**e–c**E**rtify for decision-c**A**use alignmen**T**): Locate measures temporal locking to candidate cues, Cure removes diagnosed dependence with audit-targeted decorrelation records, and Certify calibrates the audit by defect injection. Across nuScenes and Argoverse 2, four VLM families, the non-VLM planner UniAD, and 100+ trained instances, replica fleets show cue-ward dependence that is graded and population-structured. Although the confirmed case occurs in the most cue-ward fleet observed, it remains rare at the instance level; certificates therefore bind individual trained weights. **TREAT** isolates this naturally occurring defect where four standard diagnostics do not, repairs it with less task degradation than compute-matched continued training (ADE 2.72 vs. 3.22 m; 1.62 m before repair), and calibrates audit resolution with per-planner false-positive rates and detection limits at 80% power. Code: https://anonymous.4open.science/r/TREAT-5C14/README.md

Primary area:
applications to robotics, autonomy, planning
Keywords:
trajectory planning, VLMs, decision-cause alignment, Acting for the Right Reasons

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_VtFxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies whether trajectory planners act for the intended event or a correlated visual cue, even when their predicted trajectories appear reasonable. TREAT combines three stages: LOCATE compares braking aligned to cue onset and hazard-event onset; CURE fine-tunes on logged records covering all four cue/event configurations; and CERTIFY measures audit sensitivity using injected cue-locked output perturbations. Temporal identification is formulated under explicit latency and disturbance-isolation assumptions, and certificates are restricted to particular trained weights and audit conditions.

Critical strengths
  1. The paper addresses a meaningful limitation of trajectory accuracy: similar actions can arise from different dependencies. Its black-box timing audit provides evidence that complements representation probing and image interventions, and the trained positive control helps test sensitivity to a known cue-driven behavior.
  2. The evaluation includes useful controls and transparent limitations. It reports replica fleets, a disjoint-split Argoverse 2 evaluation, cue/window sensitivity, human references, and continued-training controls. Audited scenes are excluded from repair data, and failed repairs, train–audit overlap, and ADE degradation are explicitly disclosed.
  3. The certification guarantees are insufficiently validated. Calibration relies on bootstrapping already-cleared responses and transferring results between planners and cues. These procedures provide useful sensitivity estimates, but do not yet establish the claimed false-positive and missed-defect bounds.
Critical weaknesses
  1. The timing audit may confuse reasonable anticipation with incorrect cue reliance. A planner can react to an agent before it crosses the 20 m threshold. The planner is also measured using its next-three-second planned speed, while the human reference uses current logged speed. These differences weaken the causal interpretation, despite the reported geometry and window checks.
  2. Independent confirmation and repair evidence are limited. The sole confirmed natural case uses 11 scenes, eight overlapping original training scenes, without a new held-out confirmation set. The CURE–control difference has a paired p=0.16, and repair substantially worsens ADE. These results leave the repair benefit and its accuracy trade-off uncertain.
Questions for authors
  1. Could you measure the human reference using the same three-second future-speed quantity as the planner? How does LOCATE distinguish cue reliance from appropriate anticipation of an agent visible before crossing the 20 m threshold?
  2. What justifies treating bootstrap flag rates from cleared responses as false-positive bounds for causal dependence? Please clarify the assumptions behind transferring calibration between planners and cues, and validate them with known aligned negative controls.
  3. Can the confirmed case and repair effect be reproduced on independent unseen scenes? Please report uncertainty for the paired CURE–control difference and explain how the ADE increase should be weighed against the improved locking statistic.
  4. What fixes the scale of the response kernel in Proposition 2? With an unknown, unnormalized kernel, the observable mean appears to identify the contribution times the kernel's average height, rather than the coefficient alone. Please clarify the identification claim and its connection to LOCATE.
4 comments

We thank the reviewer for the careful reading and the four questions. We answer the questions in order, in three comments, and each answer covers the weakness or the remark it belongs to.

Q1 & W1.The human reference and anticipation. We recomputed the human reference with the planner's quantity, the mean speed over the next three seconds of the logged future. In this quantity the drivers show no locking in either direction on both datasets.

Human referencenuScenes (11 scenes)Argoverse 2 (17 scenes)
Current logged speed (submission)0/11, −2.67 m/s, p < 0.0015/17, −0.05 m/s, p = 0.90
Mean logged speed over the next 3 s6/10, +0.19 m/s, p = 0.586/17, −0.31 m/s, p = 0.62

Looking three seconds ahead moves the drivers' braking about 1.5 s earlier, so part of it falls into the cue-first window. The strong agent-ward contrast of the submission on nuScenes (−2.67 m/s) therefore came partly from the mismatch that the reviewer points out, and we no longer state that the drivers lock to the agent there. In neither quantity are the drivers locked to the crosswalk. One nuScenes scene is left out of the second row, because its log ends less than three seconds after the windows. Table 1 of the revision gives both references, so planners and drivers can be compared in the same quantity.

On anticipation, LOCATE separates it from cue reliance in one case and not in another. A planner that slows progressively while the agent approaches brakes more in the event window than in the cue-first window. Rule-based planners of this kind are flagged in at most 0.5% of 10,000 draws (table under Q2, in our second comment). A planner that has finished braking by the time the agent crosses 20 m gives a cue-ward contrast although its cause is the agent. Such planners are flagged in up to 6–8% when they finish early. Timing cannot decide this second case, so we added unseen scenes in which only one of the two causes is present. In 16 nuScenes scenes a crosswalk comes into view with no agent on the ego path. In 30 scenes an agent enters the danger zone with no crosswalk within 60 m. None of these scenes is in the audit set, the training manifest or the repair mixtures. The table reports the mean braking after each onset (Eq. 12), and the drivers are measured with the planner's quantity.

PlannerCrosswalk only (16 scenes), braking after the crosswalk onset (m/s)Agent only (30 scenes), braking after the agent onset (m/s)
Logged human drivers1.221.22
Flagged instance0.810.96
+ CURE0.921.40
+ compute-matched control0.920.89
Seven cleared planners (range of means)0.68–1.530.94–1.19

At a crosswalk without an agent, the flagged instance slows by 0.81 m/s and the drivers by 1.22 m/s. This slowing is ordinary caution and not a defect, and we changed Definition 2 to say so. Alignment now requires that the agent drives the braking with or without a cue. Misalignment means that the cue takes the agent's place. The agent-only scenes test this. There the drivers brake by 1.22 m/s and the cleared planners by 0.94–1.19 m/s. The flagged instance brakes by 0.96 m/s on the same scenes and in the same window, and the paired difference to the drivers has a 95% interval of [−0.69, +0.31]. The control therefore does not show that the crosswalk takes the agent's place. Section 5.3 and Appendix D.2 (Table 8) of the revision report both controls. A flag is confirmed only if the instance fails the agent-only one, so the revision reports this instance as flagged and no longer as confirmed.

Q2.The false-positive rate. A bootstrap of cleared responses does not justify a bound on false positives for causal dependence, and the revision no longer reads it as one. Such a bootstrap resamples the reading of one planner, and whether that planner is free of cue dependence is what the audit was meant to establish. The reviewer describes these numbers as sensitivity estimates. That is the right description, and we adopt it.

The false-positive rate is now obtained from planners whose alignment is known, as the reviewer asks. Each control planner follows a written rule that never reacts to the crosswalk. There are four rules. The planner ignores both events, brakes when the agent enters the danger zone, slows progressively while the agent approaches, or finishes braking before the agent reaches 20 m. Its noise is resampled from the pre-onset residuals of the audited planners. For each rule, each of three braking magnitudes and each of two noise models, we audit 10,000 planners on the same cue-first scenes with the unchanged flag rule. A second check needs no model of the planner. We flip the signs of each audited planner's own scene-level contrasts, which imposes the null, and count the flags over all sign assignments.

Planner under testnuScenes (11 scenes)Argoverse 2 (17 scenes)
Nominal level of the flag rule2.5%2.5%
Audited planners, null imposed on their own contrasts (range)2.1–2.9%2.5–2.6%
Flagged LFM2-VL-1.6B instance, null imposed on its own contrasts2.7%–
Rule-based, no response to either event2.3–2.6%2.3–2.5%
Rule-based, brakes when the agent enters the danger zone0.0–0.5%0.0–0.3%
Rule-based, slows progressively while the agent approaches0.0–0.5%0.0–0.3%
Rule-based, completes its braking before the agent reaches 20 m (lead varied)2.0–6.2%3.1–8.2%

The rule stays at or near its nominal level except for the planner that finishes braking early, which is the case treated under Q1 in our first comment. We also removed the transfers between planners and between cues. Every instance is calibrated on its own contrasts. This includes the flagged one, which no longer takes a value from another planner. Every cue is calibrated on its own frames, and the rescaling by scene count is gone. The certificate is limited to the audited weights, cue set and scene population, and its detection limit refers only to the injected family of defects. This is in Section 4.3 and Appendix C.3 (Table 7) of the revision.

Q3 & W2.Unseen scenes, the repair arms. We tested the natural case on unseen scenes in two ways, and neither confirms it. The revision therefore reports the case as flagged on the pre-specified audit set and no longer as confirmed. First, we formed a second audit set with the same rule on the Boston part of the nuScenes test split, which is in no training manifest. Its agent onsets come from a camera detector, because the split has no public object annotations. On its twelve usable scenes the instance leans the same way but is not flagged (8/12, +0.78 m/s, p = 0.21). Second, the single-cause scenes introduced under Q1 in our first comment are unseen as well, and on the agent-only scenes the instance is not separated from the drivers (0.96 against 1.22 m/s). The repair comparison that follows is made on the eleven audit scenes, and we do not claim it beyond them.

On the eleven audit scenes, a single run per arm does not separate CURE from the control. The paired difference is +0.56 m/s, with a 95% interval of [−0.26, +1.38] (p = 0.16), and the revision says so. We therefore trained each arm four times and added a control that takes as many optimizer steps as CURE. CURE ends 0.88 m/s below this step-matched control (p = 0.003) and 0.52 m/s below the original control (p = 0.084). On the overlap, the eight scenes that are also in the training manifest were trained with the drivers' own futures as targets. A planner that memorized them would read like the drivers, who are not locked to the crosswalk on the audit scenes (+0.19 m/s, p = 0.58), so the overlap does not produce the flag. The three unseen scenes all lean toward the crosswalk (3/3, +1.98 m/s).

The ADE increase comes from the training recipe of the repair arms and not from the decorrelation records. Over four runs the ADE is 2.51 m for CURE, 2.83 m for the original control and 2.64 m for the step-matched control, against 1.62 m before repair. We do not claim that the repaired planner is the better planner overall. The claim is that the arm with decorrelation records lowers the cue-ward contrast on the audit scenes more than the arms without them. Whether 0.9 m of ADE is an acceptable price depends on the deployment, and an open-loop audit cannot settle it. The revision therefore reports both quantities for every arm (Table 2).

Q4.Proposition 2. The reviewer is right. With an unknown and unnormalized kernel, the window mean identifies the product βZgˉZ\beta_Z\bar g_Z, and the coefficient alone is not identified. The revision fixes the scale by the convention gˉZ=1\bar g_Z=1 on the cue-first window. With this convention, βZ\beta_Z is the mean cue-driven change of the response in that window, in m/s, and Proposition 2 is restated for this quantity. Two statements hold for any scale, provided the kernel has one sign on the window, which we now assume for a braking response. The window mean is zero only if βZ\beta_Z is zero, and its sign is the sign of βZ\beta_Z. The screening step needs only these two statements. LOCATE estimates the same window mean as the braking after the cue onset (Eq. 12) and compares it with the braking after the event onset, so it never used the scale of the kernel. The projection of Eq. 7 requires the shape of the kernel, and the revision now gives it as the estimator for a known kernel only. These changes are in Section 3.2 of the revision.

We hope these answers and the new controls resolve the concerns, and we thank the reviewer again for the questions, which made the claims of the paper more precise.

Thank you again for your careful review. As the discussion period is about to close, we would like to check whether our response and the revised paper have addressed your concerns.

If any point remains open, we would be glad to answer it in the time that is left. If your concerns are resolved, we would be grateful if you could consider updating your rating.

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

This paper presents an interesting methodology, Locate–Cure–Certify (TREAT), for detecting when a driving planner imitates human behavior for the wrong reasons, i.e., responds to a co-occurring cue rather than the human-relevant event. Despite limitations in the mathematical formulation, the proposed alignment measures offer complementary insights beyond standard accuracy metrics and could serve as a basis for instance-level certification.

Critical strengths

Rigorous statistical formulation and solid analysis results; Novel and safety-relevant formulation

Critical weaknesses

cue list needs to be defined in advance - no automatic cue discovery yet; Mathematic formulation may be oversimplified for actual practicality

Questions for authors

How do you assess and address the limitation on the mathematical formulation currently?

3 comments

We thank the reviewer for the positive assessment and for the two points on the scope of the method. We answer them in turn.

W1. The cue list is defined in advance. This is correct. The audit tests the cues it is given and does not discover them. Two properties limit the cost. First, the list is not chosen by hand for each planner. It is enumerated from the map and track annotations, and any annotated object with a computable onset can be added without changing the method or replaying the planner, because every cue is tested on the same recorded response series. The revision extends the scan by four nuScenes map layers, walkways, car-park areas, road dividers and lane dividers. Three of them have enough cue-first scenes to be tested, and none of their 24 planner–cue cells is flagged. Car-park areas yield four such scenes, below the minimum of five, and are recorded as untested. Second, the certificate states what the audit covered. It lists the audited cues and the untested ones, and it does not imply that an unlisted cue is clear. Automatic discovery would be a proposal step in front of the audit. For example, an open-vocabulary detector can name candidate objects, and LOCATE then tests each of them as it tests the map layers now. Section 7 of the revision describes this as future work.

W2 & Q1. The mathematical formulation may be too simple. The onset-aligned model of Eq. 2 is a first-order description, and the audit depends on it less than the paper suggested. The model defines the quantity of interest, the contribution of the cue to the response. The test itself needs the two assumptions of Section 3.1 and no functional form. Under them the human-relevant event has not yet acted in the cue-first window. A response that is locked to the cue in that window therefore cannot come from the event, whether the planner combines the two linearly, nonlinearly or through an interaction. The assumption that can fail is that the agent has no effect before it enters the danger zone, and the controls in the next paragraph address it. For the same reason LOCATE works with window means and never fits a kernel or the linear form.

We assess the formulation on cases whose answer is known. A planner trained to brake at crosswalks is flagged on 11 of 11 scenes, and injected cue-locked braking is detected at a rate that rises with its strength. The revision adds the opposite case. Rule-based planners that respond only to the agent are flagged in at most 0.5% of 10,000 runs, whether they brake when the agent enters or slow progressively as it approaches. A planner that responds to neither event is flagged in 2.3–2.6%, close to the nominal 2.5%. The same controls show where the timing argument stops. A planner that finishes braking well before the agent enters the danger zone is flagged in up to 6–8% of the runs. For this case the revision adds control scenes in which only one of the two causes is present (Section 5.3).

The revision also states what the formulation leaves out. Definition 2 now permits slowing at a cue, which human drivers show as well, and a planner is misaligned only when the cue takes the place of the human-relevant event. A new remark notes that an interaction term between cue and event vanishes in the cue-first window, so it does not affect the result. Three limits remain and are listed in Section 7. First, cues that share their onset with the event cannot be separated (Proposition 1). Second, the audit reads only one response, the planned speed. Third, the audit is open-loop. These changes are in Sections 2, 3.2 and 7 of the revision.

We hope this shows how far the formulation is relied on, and we are glad to provide further details.

Thanks for your detailed response and revising the paper draft.

After further review, I can confirm that my questions were properly addressed by the author.

We are glad that our response addressed your concerns. Thank you for your constructive feedback and for your positive rating.

Comments on the paper

5 comments

We thank the reviewer for the precise review. The two weaknesses and the three questions raise four points, on the certificate, on anticipation, on the confirmation of the natural case and on the two repair arms. We answer them in this order, in three comments.

W1 & Q1.The certificate’s bounds. We agree with both objections. The revision no longer presents the false-positive rate and the detection limit as bounds. The bootstrap of Eq. 17 resamples one unmodified series. It therefore measures how stable that planner's reading is, and not how often the rule flags a planner without cue dependence. The value that the flagged instance borrowed from a cleared planner was never tested on that instance, so Eq. 22 bounded the other planner and not the target. We replaced both quantities, and we restricted the certificate to the measured injection distribution, as the reviewer proposes.

The false-positive rate is now measured in two ways, and neither borrows a value from another planner. The first uses the planner's own data. We flip the signs of its scene-level contrasts, which imposes the null of no cue dependence, and apply the unchanged flag rule to all 2N2^N sign assignments. This works for the flagged instance as well. The second uses planners that cannot depend on the cue. Each of them follows a written rule that never reacts to the crosswalk, in the four variants listed in the table, and its noise is resampled from the pre-onset residuals of the audited planners. For every combination of variant, braking magnitude (three) and noise model (two), we audit 10,000 such planners on the same cue-first scenes. The table gives the share of planners that the rule flags. A row that covers several planners or settings gives a range.

Planner under testnuScenes (11 scenes)Argoverse 2 (17 scenes)
Nominal level of the flag rule2.5%2.5%
Audited planners, null imposed on their own contrasts (range)2.1–2.9%2.5–2.6%
Flagged LFM2-VL-1.6B instance, null imposed on its own contrasts2.7%–
Rule-based, no response to either event2.3–2.6%2.3–2.5%
Rule-based, brakes when the agent enters the danger zone0.0–0.5%0.0–0.3%
Rule-based, slows progressively while the agent approaches0.0–0.5%0.0–0.3%
Rule-based, completes its braking before the agent reaches 20 m (lead varied)2.0–6.2%3.1–8.2%

Apart from the last row, the rule flags at a rate close to its nominal level or below it. The last row is the case that timing cannot resolve, and we return to it under W2 & Q2 in our second comment.

The detection limit keeps its definition, but it is no longer transferred between planners or between cues. It is computed on each instance's own series and, for each cue, on that cue's own frames. It is reported for cleared cues only, so the flagged instance no longer carries a limit taken from another planner. We removed the rescaling by scene count of Eq. 21, because Eq. 8 also depends on the noise scale and on the kernel's mean response, as the reviewer notes. The flagged instance is cleared on three cues, intersection, stop line and traffic light. The limits measured directly on them are 40%, 35% and 40%, where the rescaling gave 35%, 34% and 47%.

On the assumptions, another checkpoint and another cue need none. Each is calibrated on its own series, and the certificate is bound to the audited weights and cue set. New scenes must be exchangeable with the audited ones, which means the same inclusion rule applied to the same scene population. Where this does not hold, the instance has to be audited again. The detection limit refers only to the injected family of defects, cue-locked braking of relative strength ss. The certificate now states this, and it no longer gives a miss risk for structural dependence. These changes are in Section 4.3 and Appendix C.3 (Table 7) of the revision.

W2 & Q2.Anticipation. The locking contrast alone cannot separate crosswalk dependence from early braking for a visible agent, so we ran the control that the reviewer proposes. Its result also led us to restate what counts as misalignment. The rule-based planners in the table of our first comment show the problem. A planner that slows progressively while the agent approaches is flagged in at most 0.5% of the draws, because it slows more in the event window. A planner that finishes braking before the agent reaches 20 m is flagged more often the earlier it finishes, in up to 6.2–8.2% of the draws. As the reviewer says, separate baselines do not help in this case. Timing therefore cannot settle it, and we use scenes in which the cause is not in doubt.

We collected two sets of unseen nuScenes scenes with a single cause. In 16 scenes a crosswalk comes into view and no agent is on the ego path within 40 m around that moment. In 30 scenes an agent enters the danger zone and no crosswalk lies within 60 m. None of these scenes is an audit scene, and none is in the training manifest or in the repair mixtures. The table gives the mean braking after the onset (Eq. 12). The human drivers are measured with the planner's quantity, the mean speed over the next three seconds.

PlannerCrosswalk only (16 scenes), braking after the crosswalk onset (m/s)Agent only (30 scenes), braking after the agent onset (m/s)
Logged human drivers1.221.22
Flagged instance0.810.96
+ CURE0.921.40
+ compute-matched control0.920.89
Seven cleared planners (range of means)0.68–1.530.94–1.19

At a crosswalk with no agent near, the flagged instance slows by 0.81 m/s. The drivers slow by 1.22 m/s on the same scenes and the cleared planners by 0.68–1.53 m/s. Slowing at a crosswalk is therefore ordinary driving, and we no longer count it as misalignment. The revision restates Definition 2 accordingly. A planner is aligned when the agent drives its braking whether or not a cue is present. A cue may add to the braking, and the planner is misaligned only when the cue takes the place of the agent. The second column tests this. With an agent in the danger zone and no crosswalk, the drivers brake by 1.22 m/s and the cleared planners by 0.94–1.19 m/s. The flagged instance brakes by 0.96 m/s. The paired difference to the drivers is −0.19 m/s, with a 95% interval of [−0.69, +0.31], so the control does not separate the instance from the drivers or from the cleared planners. It gives no evidence that the crosswalk takes the place of the agent. The revision reports both controls in Section 5.3 and Appendix D.2 (Table 8). A flag is confirmed only if the instance fails the agent-only control. This instance does not fail it, and the revision therefore reports it as flagged and no longer as confirmed.

W2.Confirmation on new scenes. We agree, and the revision no longer labels the case as confirmed. It is reported as flagged on the pre-specified audit set and not confirmed on held-out scenes. Replaying the same weights on the same scenes shows reproducibility and not confirmation. Two observations concern the overlap with training. First, the eight overlapping scenes entered training with the logged human futures as targets. A planner that memorized them would read like the drivers. Measured with the planner's quantity, the drivers are not locked to the crosswalk on the audit scenes (6/10, +0.19 m/s, p = 0.58), so the overlap does not produce the flag. Second, the three unseen scenes all lean toward the crosswalk (3/3, mean contrast +1.98 m/s, against 6/8 and +0.93 m/s on the seen ones). A second audit set exists. We formed it with the same rule on the Boston part of the nuScenes test split, which is in no training manifest. The split has no public object annotations, so the agent onsets come from a camera detector whose depth is calibrated on the annotated scenes. On its twelve usable scenes the instance leans the same way but is not flagged (8/12, +0.78 m/s, p = 0.21). The drivers read 3/12 and −0.07 m/s on these scenes, measured with the planner's quantity. Over the fifteen unseen scenes, the instance leans toward the crosswalk in eleven, with a mean contrast of +1.02 m/s. The agent-only scenes of our second comment share no scene with discovery or with training, and they do not confirm the case either. Section 5.3 and Appendix D.4 of the revision report these readings.

W2 & Q3. Repair arms. The reviewer's reading of the budgets is correct. The budget was fixed by records and not by steps. The 20,003 main-task records do not include the 2,700 decorrelation records, and both arms ran two full epochs at an effective batch size of 8. CURE therefore took 5,676 optimizer steps and the control 5,002, which is 13.5% more. We added a step-matched control, in which 2,700 further main-task records replace the decorrelation records, and we trained every arm four times. The contrast is +0.07 m/s for CURE, +0.58 m/s for the original control and +0.94 m/s for the step-matched control. The additional steps therefore do not explain the difference between the arms.

With a single run per arm, the difference on the eleven audit scenes is not significant, as the reviewer says. The paired gap in contrast is +0.56 m/s, with a 95% interval of [−0.26, +1.38] (p = 0.16). Over the four runs, CURE is 0.88 m/s below the step-matched control (p = 0.003) and 0.52 m/s below the original control (p = 0.084). Table 2 and Appendix C.2 of the revision report the four runs and the step-matched control.

We hope the new null calibration and the single-cause controls address the concerns about soundness, and we would be glad to clarify any remaining point.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.