acceptodds
Under review as a conference paper at ICLR 2027

Q-Safe: Compiling Bounded-Loss Contracts for Risk-Controlled Policy Replacement

Abstract

Adaptive imaging can fail before learning begins: a nonuniform schedule may have no equal-cost oracle advantage, its objective may not match the scientific query, or its proxy may not align with a downstream human endpoint. We organize these failures into a diagnose–then–replace protocol and present Q-Safe as a model-agnostic compiler for the remaining decision: whether a proposed action may replace an executable, budget-matched fallback. It compiles declared bounded losses into an auditable replacement contract and emits certificates for schedule headroom, calibration, and deployment risk; the proposal itself may come from a neural model or a fixed rule. An initial frozen binary-harm contract improves ConnectomeBench2 utility over ExtraTrees by 1.16 points (paired 95% CI: 0.81–1.51), although the gain is mouse-driven. We then freeze a severity-aware fallback-regret contract before opening four additional shards. All pass the 2% endpoint; pooled regret is 0.191%, and relative to the frozen incidence contract, coverage and utility rise by 13.99 and 3.30 points. Conversely, on 36 registered EMDiffuse regions from three tissues, an equal-cost tile oracle fails to beat uniform acquisition in all 18 tested conditions for one fixed preview-plus-complete-rescan schedule. The contribution is therefore an auditable replacement interface, not a claim of architectural superiority or universal benefit from learned scanning.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
computational imaging, image restoration, electron microscopy, adaptive acquisition

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Z2tqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies when an adaptive scientific imaging policy should replace a conservative fallback under a fixed resource budget. The authors propose Q-Safe, a framework that combines feasibility checks, objective and proxy validation, and bounded-loss calibration to control the risk of policy replacement. Its core replacement mechanism uses a counterfactual representation of the proposed and fallback actions to predict replacement risk, followed by threshold calibration over spatial blocks.

The main experiments on ConnectomeBench2 evaluate two loss contracts, binary harm and fallback regret, using prospectively frozen test sets. Additional experiments on other imaging datasets examine schedule feasibility, objective alignment, and proxy validity. The results demonstrate that risk-controlled replacement can improve task utility and that different loss definitions can lead to substantially different deployment decisions.

Critical strengths
  • The paper addresses a meaningful problem: when should an adaptive policy be allowed to replace an existing policy under a fixed budget. Rather than assuming that learned adaptation is always beneficial, the authors explicitly consider the potential harm of replacement and the availability of a conservative fallback. This is a useful perspective for scientific imaging, where incorrect decisions can affect downstream scientific analyses.

  • The authors make a substantial effort to evaluate their approach through frozen protocols, independent test shards, and comparisons across different species and loss definitions. The second confirmation experiment is particularly informative in demonstrating how changing the loss contract affects coverage, utility, and harmful actions. The external diagnostic experiments also provide useful negative results and help clarify the limitations of adaptive acquisition.

Critical weaknesses
  • The methodological novelty of Q-Safe is not sufficiently established. The core approach builds on existing work in risk-controlled policy switching and bounded-loss calibration (e.g., Joshi et al., 2026; Bi et al., 2026; Angelopoulos et al., 2024). While the application setting and decision granularity differ, the paper does not clearly establish what methodological advances Q-Safe introduces beyond these approaches. The limited utility changes in the representation ablations and the competitive performance of simpler encoders further leave the source of its contribution unclear.

  • The experimental evidence does not fully support the proposed integrated framework. The paper presents a unified diagnose–then–replace framework, but its components are largely evaluated independently. The main experiments support the effectiveness of risk-controlled policy replacement, while the external studies separately examine schedule feasibility, objective alignment, and proxy validity. How these components interact and what additional value their integration provides remain insufficiently demonstrated.

Questions for authors
  1. What are the specific methodological capabilities of Q-Safe that distinguish it from existing risk-controlled policy switching and adaptive-policy authorization methods, particularly Joshi et al. (2026) and Bi et al. (2026), beyond differences in application setting and decision granularity?

  2. Can the authors demonstrate how the schedule-feasibility, objective-alignment, and proxy-validation modules interact with the policy-replacement mechanism within a single end-to-end task? In particular, how do the outcomes of these checks affect the subsequent calibration and final deployment decisions?

  3. What role does the predicted utility difference (\hat{\Delta}\theta) play in the final replacement decision, given that the calibration procedure appears to rely on the harm score (s\theta)? Would removing the utility-prediction head affect the final risk–coverage–utility trade-off?

  4. How much of the observed improvement can be attributed specifically to Q-Safe rather than applying an existing bounded-loss risk-control procedure to the same proposal–fallback pairs and learned scores? Could a matched comparison help isolate this contribution?

References

[1] Joshi, S., Wang, T., Hassani, H., & Dobriban, E. (2026). Risk-Controlled Post-Processing of Decision Policies. arXiv:2605.06479.

[2] Bi, J., Pinilla, S., & Zhu, C. (2026). Certifying When Decision-Time Information Justifies Adaptive Experimentation. arXiv:2607.27651.

3 comments

We thank Reviewer Z2tq for the specific questions.

Q1 & W1. Relationship to prior risk-controlled switching

Joshi et al. (2026) preserve agreement with an existing policy subject to a constraint on loss violations. Bi et al.'s OPAL (2026) uses held-out evidence on activation risk, coverage, and net value to authorize an adaptive acquisition rule against a fixed plan. Q-Safe combines the following capabilities for budget-matched acquisition.

  1. Pre-calibration checks. We test whether the allowed schedule has equal-cost oracle headroom after accounting for preview and rescan costs (Eq. 9), and whether the objective and observable score align with the endpoint. These checks determine whether a proposal advances to calibration.
  2. Paired action representation. We pool the complete proposal and fallback sets separately and provide hph_p, h0h_0, and hp−h0h_p-h_0 alongside full-support context (Eq. 5). For additive utility, shared candidates cancel in the realized utility difference (Eq. 7), motivating the signed comparison.
  3. Fallback-relative loss. We apply the add-one rule of Angelopoulos et al. (2024) to spatial-block means of binary harm or fallback regret Lr=max⁡{−Δ,0}L_r=\max\{-\Delta,0\}. We intersect thresholds across observed calibration domains within each species. Species with insufficient calibration support, such as zebrafish, use the fallback.

Q2 & W2. How the checks affect deployment

Our registered multi-dwell EM replays (Global Response §2) show how validation determines whether to calibrate a gate or retain uniform acquisition.

  • 2D Liver EM. The mitochondrial BCE oracle had +0.00734 mean headroom. The objective and proxy checks passed on validation, with complete-action gain correlation ρ = 0.249. On held-out fields, Always Proposal had 52.50% harm and utility −1.181. The pooled gate reduced harm from 23.93% at the fixed threshold to 2.143% at 7.86% coverage, with regret 0.0010% and utility +0.0024 (×10² BCE). One validation field had no admissible threshold, so the per-field rule selected uniform fallback.
  • 3D PFIB mitochondria. The processor improved validation Dice (0.8268 vs. 0.7866) and VOI (0.3040 vs. 0.3420) over uniform, and the selector passed with ρ = 0.371. This triggered calibration and the one-shot test. The gate achieved 62.50% coverage and 1.089% regret, improving Dice (0.3804 → 0.6208) and VOI (0.3646 → 0.3591) over uniform fallback. Always Proposal had 34.38% harm and 0.3734 VOI.
  • 3D PFIB synapses. Selective Dice was 0.4301 versus 0.5420 for uniform, and the selector had ρ = −0.019. The processor and selector checks stopped the workflow at validation and selected uniform fallback.

Q3. Role of the utility-prediction head

For supported groups, acceptance is determined by sθ(q,k)≤tcs_\theta(q,k)\le t_c. The utility head contributes through the auxiliary Smooth L1 training loss (weight 0.15, Eq. 6). In a single-seed post-confirmation ablation, we set that weight to zero, retrained, and reapplied the same calibration procedure.

  • Validation harm AP: 0.2447 → 0.2170.
  • test25–28 (binary harm): Coverage 59.79% → 57.55%, harm 0.190% → 0.209%, and utility 12.816 → 12.919 pp.
  • test29–32 (regret): Coverage 74.72% → 72.39%, regret 0.191% → 0.181%, and utility 14.753 → 14.947 pp.

Auxiliary utility supervision improved validation harm AP and increased coverage in this ablation. Deployed utility was slightly higher with its weight set to zero. We will clarify the head's training role in §4.3.

Q4. What is attributable to Q-Safe beyond CRC?

With the same action pairs, scorer, loss, spatial blocks, and threshold rules, applying standard add-one CRC within each calibration group and intersecting the resulting thresholds yields the Q-Safe gate. On test25–28, holding the scorer fixed, calibration changed observed harm from 1.881% to 0.190% and utility from 21.847 to 12.816 pp. Holding calibration fixed, the counterfactual scorer gained 1.157 pp utility over ExtraTrees.

With the scorer fixed on test29–32, changing the replacement loss from incidence to regret increased coverage by 13.99 pp and utility by 3.30 pp, with 0.191% pooled regret and higher harm incidence (1.018% vs. 0.539%). This comparison isolates the effect of the declared loss on deployment under the same calibration procedure.

Thank you for the detailed response and additional experiments. I consider Q2 and Q3 adequately addressed.

My remaining concerns are Q1 and Q4. Could you clarify what methodological capabilities distinguish Q-Safe from Joshi et al. (2026) and Bi et al. (2026), beyond differences in application setting and decision granularity?

Relatedly, given that standard CRC produces an identical gate under matched conditions, what additional contributions or measurable benefits arise from Q-Safe's proposed design and framework integration?

This would help me better assess the novelty of the proposed method.

Thank you for the follow-up.

Q1. Methodological distinction

Joshi et al. choose when to depart from an incumbent policy using the estimated change in the probability of exceeding a specified loss cutoff. Q-Safe defines loss relative to an executable, budget-matched fallback. It calibrates either the incidence of harmful replacements or their mean utility shortfall, with fallback executions contributing zero loss. On ConnectomeBench2, the harm scorer pools the two complete top-kk sets separately and uses both embeddings and their signed difference (hp−h0h_p-h_0). Shared items cancel in the realized utility difference under additive utility (Eq. 7), motivating this paired representation.

OPAL uses held-out evidence to authorize adaptive measurement and distinguishes oracle opportunity from decision-time information. Q-Safe applies this distinction to the complete acquisition and processing pipeline. For additive tile losses, the equal-cost oracle test upper-bounds the gain of any rescan ranking after accounting for preview cost. On validation data, we check the acquisition objective against the target endpoint, compare the processed proposal with the processed uniform reference, and test whether the preview-derived score ranks their realized utility differences. The processor comparison measures how mixing preview and rescan inputs affects the task outcome. These checks determine which configurations advance to calibration.

Q4. What changes the replacement decision

With the action pairs, scores, loss, spatial blocks, and domain rule fixed, applying add-one CRC within each calibration group and intersecting the resulting thresholds yields the same gate. The following comparisons show how the representation, loss, validation criteria, and calibration grouping affect the resulting decisions.

  • Paired representation (Tables 1 and 2). Removing the explicit embedding difference lowers harm AP from 0.249 to 0.224 and raises worst-shard harm on reused development shards from 1.579% to 1.835%. Utility decreases by 0.015 pp. Under the same calibration procedure, the full counterfactual scorer adds 1.157 pp utility over ExtraTrees on test25–28 (paired 95% CI [0.812, 1.512], primarily in mouse). On test29–32, under the same regret calibration procedure, it achieves 2.61 pp higher coverage, 1.01 pp higher utility, and 0.048 pp lower regret.
  • Replacement loss (Table 3). With the scorer and calibration procedure fixed, switching the 2% target from harm incidence to fallback regret on test29–32 increases coverage by 13.99 pp and utility by 3.30 pp, primarily in fly. Pooled regret is 0.191%, and harm incidence rises from 0.539% to 1.018%. The loss definition determines which frequency–severity trade-off the gate implements.
  • Validation before calibration (3D PFIB replay). At matched nominal dwell, the earlier mitochondrial processor with a gap veto reached 0.7943 Dice and 0.3723 VOI on validation. With component consolidation, the selective configuration reached 0.8268 Dice and 0.3040 VOI, versus 0.7997 and 0.3112 for uniform with the same consolidation. The context-level selector passed validation (ρ = 0.371), and the pipeline advanced to calibration. On the one-shot spatial holdout, the gate achieved 62.50% coverage and 1.089% regret. Relative to uniform fallback, mean Dice rose from 0.3804 to 0.6208 and mean VOI fell from 0.3646 to 0.3591. For synapses, selective processing trailed uniform on validation and the selector had ρ = −0.019. The workflow retained uniform acquisition before calibration.
  • Calibration across fields (2D Liver replay). With the same scorer and 2% binary-harm target, pooled calibration yielded 7.86% coverage and 2.143% observed harm. One validation field had no admissible threshold, so per-field intersection selected uniform fallback throughout, with 0% coverage and 0% harm.
Reviewer_XFL1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces Q-Safe, a model-agnostic policy-replacement interface that decides whether a proposed adaptive action may replace a budget-matched fallback. Before making the replacement decision, the framework checks schedule feasibility using an oracle headroom test, followed by objective alignment and proxy alignment. The method is evaluated through two frozen confirmation experiments on the ConnectomeBench2 proofreading task, along with diagnostic tests on four imaging datasets that yield mostly negative results.

Critical strengths
  1. The paper studies the important problem of deciding when a learned adaptive policy should replace a known fallback under the same budget. It treats rejection as executing a concrete fallback action. The framing is likely to be of interest to practitioners.

  2. The paper tests whether adaptation has any oracle headroom over an equal-cost uniform schedule before training a ranker, and treats “do not adapt” as a legitimate outcome. Non-adaptive baselines are often under-evaluated, so a simple pre-training check against them could be useful.

Critical weaknesses
  1. Figure 1 presents three diagnostic checks: (1) whether the schedule is feasible, (2) whether the objective is aligned, and (3) whether the proxy is aligned. The paper presents this staged diagnostic procedure as one of its contributions. However, only the first check has a clearly defined procedure. The other two are mostly described conceptually and evaluated through one experiment each, without a clear criterion for what counts as passing or failing the check. As a result, the three-stage hierarchy feels more like a summary of the experiments than a well-defined diagnostic protocol.

  2. The organization of the method section does not clearly follow the diagnostic stages presented in Figure 1. This makes it difficult to connect the individual experiments and analyses back to the proposed framework and makes the paper harder to follow.

  3. The oracle headroom test (Eq. 9) seems weaker than the paper suggests. Passing the test only shows that there is room for improvement with a perfect ranker, but does not show that such a policy can actually be learned. On the other hand, failing the test only rules out the particular schedule being tested, rather than adaptation more generally. Also, since the result of the test depends on the choice of loss, it is unclear why “schedule infeasibility” should be considered a separate failure mode from “objective mismatch.”

  4. The paper frames its contribution as risk-controlled replacement of an adaptive acquisition policy, but it evaluates only a one-shot choice between two fixed top-k sets, with both outcomes known for every query. There is no sequential decision-making or learned policy, so the problem seems to reduce to cost-sensitive classification with a calibrated threshold.

Questions for authors
  1. Where are checks 2 (Objective aligned) and 3 (Proxy aligned) in Figure 1 defined as procedures, and what are their pass criteria?

  2. Since the result of Eq. 9 depends on the choice of loss, could changing the objective turn a schedule failure into a pass?

  3. Which parts of the guarantee survive in a sequential setting with a state-dependent fallback?

1 comment

We thank Reviewer XFL1 for the detailed questions.

Q1, W1 & W2. Diagnostic procedures and pass criteria

Section 3.1 outlines the diagnostic sequence, §4.6 defines the schedule test, and §§5.5–5.6 report the alignment audits. We will organize §4 around Figure 1 and state the validation criteria and stop decisions for each task.

  • Check 1 (Schedule feasibility). Fix the target loss LL, allowed actions, and cost model. For the additive tile loss in Eq. (9), compare the best allowed rescan selection with equal-cost uniform acquisition. The Liver replay required positive mean oracle headroom and positive headroom in at least two of three validation fields. Both criteria were met.
  • Check 2 (Objective alignment). Specify the target structure and test the acquisition objective against a target-specific endpoint. In the Liver replay, the criterion was lower mean BCE against manual mitochondrial masks at 3 µs than at 1 µs. The reduction was 0.01560. In the SBEM audit, the membrane-aligned score captured 40.9% of membrane disagreement at a 25% rescan budget, versus 28.4% for the pixel-risk score (Table 10).
  • Check 3 (Proxy alignment). Compare the preview-only score with the measured proposal–fallback outcome. In the Liver replay, we required harm AP above prevalence and positive Spearman correlation between the negative harm score and complete-action gain. The validation results were AP 0.666 versus 0.536 prevalence and ρ = 0.249.

In the 3D PFIB replay (Global Response §2B), the mitochondrial processor improved both Dice (0.8268 vs. 0.7866) and VOI (0.3040 vs. 0.3420) over uniform on validation. The selector passed its validation criteria (context-level ρ ≥ 0.30, a 95% CI for top-quartile gain above zero, and a train–validation ρ gap < 0.10), with ρ = 0.371. This triggered calibration and the one-shot test (62.50% coverage, 1.089% regret). For synapses, the processor and selector checks stopped the workflow on validation (Dice 0.4301 vs. 0.5420 and ρ = −0.019).

Q2 & W3. Can changing the loss reverse Eq. (9)?

Yes. Changing LL changes the gains from both rescanning and uniform acquisition, so it can reverse the oracle comparison at fixed cost. Table 11 reports negative headroom for pixel, edge, and dark-structure losses. The Liver replay found +0.00734 BCE mean headroom for mitochondrial segmentation on its labeled validation subset, using three-tile action units (Global Response §2A).

Schedule feasibility asks whether any allowed rescan selection can improve the declared endpoint at the fixed budget. When the oracle gap is nonpositive, no ranker in that action class can beat uniform in that comparison. Objective alignment asks whether the acquisition objective represents the scientific target. A mismatch calls for changing the objective and rerunning the oracle at the same budget. Positive headroom quantifies what a perfect ranker could achieve. Check 3 tests whether the available preview predicts useful replacements on held-out data.

Q3 & W4. Sequential decisions and a state-dependent fallback

ConnectomeBench2 learns a harm score for a one-shot choice between two fixed top-kk actions. The PFIB replay likewise makes one replacement decision per spatial context.

  • One choice between complete policies. From the same observed state, choose between two fixed policies with the same remaining budget. Each policy can adapt internally, including the state-dependent fallback. If ℓb∈[0,M]\ell_b \in [0,M] is the proposal's bounded loss relative to fallback and sbs_b is its score, then Lb(t)=1{sb≤t}ℓbL_b(t)=\mathbf{1}\{s_b\le t\}\ell_b is nested in the threshold. With exchangeable calibration and deployment loss curves, Eq. (8) controls the expected block-level loss.
  • Repeated switching within a trajectory. Changing the threshold can change the states reached later and the resulting loss. Applying Eq. (8) in this setting requires valid trajectory-level evaluation of the candidate controllers and exchangeable loss curves that are monotone in the calibration threshold. We will state this distinction in §§3 and 7.
Reviewer_AuhNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes Q-Safe, a gate that decides for each query whether to use a new learned policy or fall back to an existing equal-budget policy. A scorer predicts whether the new policy would do worse, and a threshold is calibrated with conformal risk control so that harmful replacements stay below 2%. It is evaluated on connectome proofreading across four species, where the risk target is met with small utility gains over a simple tree-based baseline.

Critical strengths
  1. The setting is reasonable. In microscopy or proofreading the system cannot just abstain, it still has to do something with the budget. So treating “reject” as running the old equal-budget policy, and measuring risk on the difference between the two policies, makes sense to me. This is different from standard selective prediction.

  2. The procedure is easy to reuse. The calibration works with any scorer, including simple tree models or hand-written rules, and domains without enough calibration data simply fall back to the old policy. It could be added on top of an existing pipeline without much change.

Critical weaknesses
  1. The technical contribution is limited. The risk guarantee comes directly from the plus-one bound in conformal risk control, and taking the strictest threshold over domains and calibration splits is a straightforward extension. The authors also say the contribution is not architectural: the ablations change utility by at most 0.028 points, and simpler encoders (DeepSets, generic attention) get higher utility in the appendix. The frozen ExtraTrees gate also stays well below the 2% target (0.248% pooled harm), and the paired differences in harm and coverage are not statistically significant; only the utility gain (about 1.16 points) is, and it mostly comes from mouse. So the learned scorer adds little, the calibration is existing work, and it is unclear what Q-Safe itself adds.

  2. The claims are not well supported by the experiments. The paper is framed around adaptive imaging and acquisition budgets, but the full Q-Safe pipeline is only run on connectome proofreading, which is selecting k candidates to check and does not involve acquisition. On the imaging datasets (EMDiffuse, SBEM, PFIB, HeLa), only separate diagnostic steps are tested. The paper also does not report the “always use the proposal” or uncalibrated gate baselines, so it is hard to see how much harm the gate actually prevents and how much utility it costs.

  3. The paper is hard to follow. The core method is simple, but it is buried under many new terms such as “compiler”, “contract”, “certificate”, and “ledger”, which mostly rename standard things. The main text also switches between many datasets. In addition, Appendix B (Compiled Token Specification) appears to be empty.

Questions for authors
  1. What are the harm rate and utility on the blind shards for (a) always using the proposal and (b) a gate with a fixed threshold and no calibration? Without these, it is hard to judge what the gate actually buys.

  2. If the Q-Safe scorer is replaced by ExtraTrees and everything else is kept the same, what remains different? Should the improvement be credited to the scorer or to the calibration?

1 comment

We thank Reviewer AuhN for the clear feedback.

Q1. Harm and utility of the uncalibrated baselines

On the original blind ConnectomeBench2 shards (test25–28, Global Response §1), the post-confirmation audit found:

  • Always Proposal: 100.00% coverage, 3.408% block harm, and 22.531 pp utility. All four shards exceeded 2% harm.
  • Fixed threshold (harm probability ≤0.5): 96.23% coverage, 1.881% block harm, and 21.847 pp utility. Two shards exceeded 2% harm (test26–27).
  • Calibrated Q-Safe: 59.79% coverage, 0.190% block harm, and 12.816 pp utility. All four shards were below 2% harm.

Relative to the fixed threshold, calibration reduced pooled harm by 1.691 pp at a utility cost of 9.031 pp.

Q2 & W1. Scorer versus calibration

In the matched audit, we used the same spatial-block calibration procedure for ExtraTrees and the counterfactual set scorer, fitting thresholds separately for each. This reproduced the operating points in Table 2.

  • Calibration. Both calibrated gates had observed harm below 2% on all four test25–28 shards, with pooled harm of 0.248% for ExtraTrees and 0.190% for Q-Safe. Zebrafish queries used the budget-matched fallback because calibration support was insufficient.
  • Scorer. Holding calibration fixed, the counterfactual scorer added 1.157 pp utility over ExtraTrees on test25–28 (paired 95% CI [+0.812, +1.512]), primarily in mouse. Under the regret loss on test29–32, it added 2.61 pp coverage and 1.01 pp utility, with 0.048 pp lower regret.

The replacement workflow supports either scorer and lets users choose which loss to control. With the scorer fixed on test29–32, calibrating for regret increased coverage by 13.99 pp and utility by 3.30 pp relative to incidence calibration, with 0.191% pooled regret and higher harm incidence (1.018% versus 0.539%). We will add the baseline table from Global Response §1 to §5.2.

W2. End-to-end imaging evaluation

We evaluated preview-plus-rescan, segmentation, and calibrated replacement together in replays on registered 2D Liver EM and 3D PFIB EM acquisitions (Global Response §2).

  • 2D Liver. Always Proposal yielded 52.50% harm (utility −1.181), and the fixed threshold yielded 23.93% harm (utility −0.082). The pooled gate reduced observed harm to 2.143% at 7.86% coverage, with utility +0.0024 (all utility values in ×10² BCE). The per-field rule selected uniform fallback throughout.
  • 3D PFIB mitochondria. After the validation checks passed, the frozen gate achieved 62.50% coverage and 1.089% regret in the one-shot held-out test. Compared with uniform fallback, mean Dice rose from 0.3804 to 0.6208 and mean VOI fell from 0.3646 to 0.3591.
  • 3D PFIB synapses. Selective scanning trailed uniform on validation Dice (0.4301 versus 0.5420), and the selector had ρ = −0.019. The workflow stopped at validation and selected uniform fallback.

W3. Terminology, dataset flow, and Appendix B

We will use plain terms such as bounded loss, calibrated threshold, and evaluation table. We will present the ConnectomeBench2 results together and group the imaging audits by diagnostic step. We will place Table 5 directly under the Appendix B heading.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.