PaperlayerView exact revision on arXiv

Source-linked AI summary

DepthBenchCAD: When Does Deeper Auditing Yield More Reliable Conclusions?

Hongye Yang, Zhihao Xie, Shengjun Xiong, Boxiao Huang

arXiv:2609.15122v1cs.SEcs.AI

TL;DR

Generative CAD programs are expected to remain behaviorally correct after edits, but under a fixed budget deeper audits can reduce task coverage and independent generations. The paper models evidence at the template, generation, and edit levels and combines variance with execution costs to study allocation. Its results show that the value of deeper auditing depends on the dominant uncertainty source, and no audit depth is universally best.

  • Problem

    Behavioral evaluation must capture failures after CAD parameter edits while balancing edit checks against task and independent-generation coverage under a fixed budget.

    Source: §1 · pp. 1–2

  • Method

    The paper models evaluation as sampling templates, generations, and edit states, then uses variance and measured execution costs to analyze evidence allocation.

    Source: §1 · p. 2, §3.2 · p. 3

  • Results

    No audit depth is universally best, and joint allocation does not consistently outperform strong fixed-depth baselines; its value is diagnosing uncertainty sources and guiding budget adjustment.

    Source: §7 · p. 9, §6 · p. 9

  • Takeaways & Limitations

    Template heterogeneity favors more tasks, generation variability favors more independent programs, and substantial edit-state variation with expensive generation favors deeper auditing.

    Source: §6 · p. 9

    The estimand covers 16 frozen edit states rather than continuous parameter spaces or open-ended interaction, and average failure risk does not establish engineering safety or worst-case correctness.

    Source: §6 · p. 9

Abstract

from arXiv · show

Generative CAD models are expected to remain behaviorally correct after parameter edits, so increasing the number of edit checks is often treated as a direct route to more reliable evaluation. Under a fixed budget, however, auditing each program more thoroughly reduces the number of tasks and independent generations that can be evaluated, which can ultimately make model-level estimates less accurate. We study this phenomenon and the conditions under which it arises. We decompose behavioral evaluation into three evidence levels: task templates, stochastic generations, and within-program edits. We define an average failure risk that is invariant to audit depth, and combine three-level variance with measured execution costs to analyze the tradeoff between deeper edit auditing and broader independent coverage. Experiments across two CAD environments and five generation systems show that the value of deeper auditing depends on where evaluation uncertainty originates. When template heterogeneity or generation stochasticity dominates, additional edit checks can increase total estimation error; when within-program state variation is large and generation is expensive, deeper auditing is more valuable. Variance and cost estimates from calibration predict the direction of this change and provide a diagnostic basis for allocating evidence on held-out tasks. These results show that the thoroughness of program inspection can diverge from the reliability of model evaluation, and they help determine whether the next unit of budget should be spent on a new task, a new generation, or additional edit checks.

1 INTRODUCTION

Generative CAD programs must remain executable and preserve design intent after parameter edits, but evaluation must balance edit checks against task and generation coverage. The paper defines an audit-depth-invariant failure-risk target and studies how variance and cost can guide that balance.

1 INTRODUCTION

  • Parametric CAD programs can fail after edits even when they appear correct initially, so static geometric metrics do not fully capture editability or engineering utility.Failures can include build errors, self-intersections, invalid parameters, and constraint violations.Source: §1 · p. 1
  • Behavioral evaluation spans task templates, independent generations, and within-program edit audits, whose evidence and costs differ.Under a fixed budget, deeper auditing can reduce task coverage, while shallow auditing can miss conditional failures.Source: §1 · p. 2
  • The model-level target is average behavioral-failure probability across a randomly sampled template, generation, and valid edit state, unchanged by audit depth.Freezing valid edit states makes budget allocations comparable against the same capability.Source: §1 · p. 2
  • The paper characterizes how deeper auditing can increase estimation error by reducing task coverage or independent generations, and tests whether variance and execution cost predict its benefit.It evaluates reliability through risk estimation, interval coverage, and judgment error, while reporting calibration cost.Source: §1 · p. 2

2 RELATED WORK

The related work situates generative CAD evaluation within CAD representations and benchmarks, then frames evidence allocation through multistage sampling and stochastic evaluation research.

2 RELATED WORK

  • Prior CAD systems preserve design histories and construction logic through operation sequences, programmatic data, and recovered code.BenchCAD and Text2CAD-Bench extend evaluation toward execution, parameter reasoning, and editing.Source: §2 · p. 2
  • This paper focuses on allocating evaluation evidence under a fixed budget, building on multistage sampling and clustered-observation methods.The cited foundations include Neyman (1934), Horvitz & Thompson (1952), and Liang & Zeger (1986).Source: §2 · p. 2
  • The work also draws on research about stochastic evaluation and selection bias.The passage cites Henderson et al. (2018), Reimers & Gurevych (2017), Dietterich (1998), and Varma & Simon (2006).Source: §2 · p. 2

3 METHOD

The method models generative-CAD behavioral evaluation as nested sampling over templates, generations, and edits. It defines an invariant failure-risk target, decomposes estimation variance, and uses execution costs to select and calibrate evidence allocations.

3 METHOD

  • The method treats behavioral evaluation as three-level nested sampling over templates, generations, and edits, with the design structure determining effective sample size.It defines the estimand, estimator, variance decomposition, cost constraint, and calibration strategy.Source: §3 · p. 2

3.1 EVALUATION TARGET AND FINITE EDIT POPULATION

  • The target R_s is average behavioral-failure probability across a randomly sampled template, model generation, and valid edit state; audit depth changes precision and cost, not the target.Single-program certification, all-state pass rate, and worst-case risk are outside the paper's scope.Source: §3.1 · pp. 2–3

3.2 THREE-LEVEL SAMPLING AND RISK ESTIMATION

  • The design samples q templates, g programs per template, and k frozen edit states per program, with each level controlling a different source of uncertainty.Templates control task coverage, generations control within-task generation uncertainty, and audits control measurement error in an individual program's risk.Source: §3.2 · p. 3

3.3 THREE-LEVEL VARIANCE DECOMPOSITION 

  • Increasing audit depth k reduces only edit-state sampling error; reducing between-template variation or generation stochasticity requires increasing q or g.For a finite reference pool, resampling uses a finite-population variance estimate, while future-task uncertainty is bootstrapped by template clusters.Source: §3.3 · pp. 3–4

3.4 COST MODEL AND JOINT ALLOCATION

  • The cost model includes template preparation, generation and initial build, and each edit execution; discrete search selects feasible q, g, and k under the budget.The search accounts for template saturation, finite-population corrections, and residual integer budget.Source: §3.4 · p. 4

3.5 CALIBRATION COST AND USAGE STRATEGY

  • Calibration estimates variance and cost using independent templates; pooled configurations need target-system cost measurement, while system-level calibration can diagnose deviations and reuse scenarios.Systems can have different generation and execution costs.Source: §3.5 · p. 4, §3.4 · p. 4

4 EXPERIMENTAL DESIGN

The experiments compare evidence-allocation strategies across two CAD environments and five generation systems under a fixed evaluation budget. They use frozen edit states, a validated automated judge, held-out calibration, and metrics that separate program-level from model-level error.

4.1 DATA AND EVALUATION ENVIRONMENTS

  • Five systems are evaluated in primary environment A and cross-task-family environment B using a frozen, fully audited pool of real build, edit, and judgment outcomes.A tests prediction on held-out templates; B tests transfer across disjoint task families.Source: §4.1 · pp. 4–5

4.2 COUNTERFACTUAL EDIT STATES AND AUTOMATED JUDGE

  • Each template has 16 pre-frozen edit states, and an automated judge checks build, geometry, parameter response, relational, and semantic constraints.Two experts independently and blindly review a stratified sample of 800 records before consensus labeling.Source: §4.2 · p. 5

4.3 DATA SPLITS AND CALIBRATION SETUP

  • Family-stratified calibration and test splits keep each template and its generations together, while test data do not select variance parameters, depth, or costs.Full calibration audits all 16 states; smaller pilots test shrinkage and reuse.Source: §4.3 · pp. 5–6

4.4 COMPARISON METHODS AND BUDGETS

  • Comparisons include fixed depths, pooled and system-level allocation, shrinkage, edit-type stratification, and paired allocation for model pairs.The formal budget is C0=1024; generation costs vary by system and environment.Source: §4.1 · p. 5, §4.4 · p. 6

4.5 EVALUATION METRICS AND STATISTICAL INFERENCE

  • Evaluation reports program-level MSE and cost-scaled model-level J, then tests whether calibration predicts the direction and cost of deeper-audit benefit.Auxiliary measures include regret, robust and interval coverage, judge validity, and paired model decisions.Source: §4.5 · p. 6

5 EXPERIMENTAL RESULTS

Across systems, deeper edit auditing does not uniformly improve model-level estimation: its value depends on the uncertainty source, while calibration can predict benefit direction but does not consistently beat a strong fixed-depth baseline. Calibration expense, transfer stability, judge validity, and paired decisions further qualify the allocation results.

5.1 EVIDENCE HIERARCHY AND EVALUATION RELIABILITY

  • Three-level design variance raises mean interval coverage from 91.81% to 95.26%; gains are largest for S1 and S2, which have stronger template-level correlation.Their coverage rises from 89.06% and 86.82% to 94.28% and 94.42%, respectively.Source: §5.1 · p. 6
  • At k=16, S1's program-level MSE falls from 0.014284 to zero, but model-level J rises from 0.0974 to 0.1649 as generated programs drop from 184 to 48.S4 shows the opposite pattern: J falls from 1.2881 to 0.5378 when depth rises from 4 to 16, despite templates decreasing from 46 to 30.Source: §5.1 · pp. 6–7, Figure 1 · p. 7
  • Calibration predicts the direction of benefit correctly in nine of ten environment-A comparisons and all ten after local calibration in environment B.Mean wrong-recommendation loss across the ten environment-A comparisons is 0.000389; the authors describe the 20 comparisons as limited mechanism validation.Source: §5.1 · p. 7
  • Fixed k=16 has five-system mean J of 0.471, lower than stratified at 0.481, system-level at 0.502, and pooled at 0.533.The authors do not interpret calibration-driven joint allocation as outperforming the best fixed depth on average.Source: §5.1 · p. 7

5.2 CALIBRATION COST AND TRANSFER STABILITY

  • Full calibration costs about 2.04–5.67 times the formal budget, while small calibration costs about 28.1%–76.6%.Across random template splits, median relative regret is 1.041 and mean relative regret is 1.098; leave-one-task-family-out P90 regret reaches 1.709 for S1 and 1.608 for S2.Source: §5.2 · pp. 7–8
  • In environment B, small calibration with four-way stratification has mean J=0.667, compared with 0.742 for fixed k=10 and 0.756 for direct transfer from A.The primary-environment configuration can transfer its frozen g and k, but optimality on a new task set is not guaranteed.Source: §5.2 · p. 8, Figure 3 · p. 9

5.3 JUDGE VALIDITY AND MODEL DECISIONS

  • The automated judge preserves expert-consensus risk ranking across all five systems but underestimates S4 and S5 risk by about 1.97 and 3.65 percentage points.S2's automated labels exactly match expert consensus.Source: §5.3 · p. 8
  • Paired allocation offers only a small improvement over a strong fixed-depth baseline, with its benefit largely disappearing at high budgets.Its value is concentrated in small risk gaps, low paired variance, and reusable calibration cost.Source: §5.3 · p. 8

6 DISCUSSION

The discussion argues that evaluation evidence should match the dominant source of uncertainty, while calibration is worthwhile only when its gains and reuse justify its cost. The conclusions are bounded by the finite states, sampling assumptions, judge coverage, and tested settings.

6 DISCUSSION

  • Template heterogeneity favors more tasks, generation variability favors independent programs, and substantial edit-state variation with expensive generation favors deeper auditing.Joint allocation explains system-specific audit preferences but does not consistently outperform strong fixed-depth baselines.Source: §6 · p. 9
  • Calibration is useful only when its gains and reuse justify its additional cost, and transfer beyond the tested task composition warrants caution.The tested families, systems, and kernels constrain generalization.Source: §6 · p. 9
  • The estimand covers 16 frozen states rather than continuous parameter spaces or open-ended interaction, and average failure risk does not establish engineering safety or worst-case correctness.The variance model assumes a prespecified sampling design and within-level exchangeability; small calibration sets and limited judge semantic coverage are additional constraints.Source: §6 · p. 9

7 CONCLUSION

Under a fixed budget, deeper CAD-program auditing can reduce model-evaluation reliability by sacrificing task coverage and independent generations. The appendix derives how variance components and execution costs inform allocation, while noting that calibration and finite data constrain practical efficiency.

7 CONCLUSION

  • Under a fixed budget, deeper auditing can reduce evaluation reliability by sacrificing task coverage and independent generations; no depth is universally best.Joint allocation does not consistently outperform strong fixed-depth baselines, and practical efficiency depends on calibration cost, task composition, and reuse.Source: §7 · p. 9

A.1 DESIGN-BASED DERIVATION OF THE THREE-LEVEL VARIANCE

  • Increasing edit depth k reduces only within-program edit-state sampling error; template and generation uncertainty require more templates q or generations g.The decomposition separates template heterogeneity, generation heterogeneity, and finite-population variation among frozen states.Source: §A.1 · p. 14

A.2 METHOD-OF-MOMENTS ESTIMATION OF VARIANCE COMPONENTS

  • Full calibration records all M=16 states per program and subtracts lower-level measurement noise to avoid misattributing edit variation to generation or template heterogeneity.Negative method-of-moments residuals are truncated at zero so optimization uses finite, nonnegative variance components.Source: §A.2 · p. 14

A.3 COST CONSTRAINT, CONTINUOUS APPROXIMATION, AND INTEGER SEARCH

  • Larger within-program variance or fixed generation cost favors deeper auditing, whereas larger template/generation variance or per-edit cost favors new templates or generations.The continuous relaxation interprets the mechanism; reported allocations use integer search, which selects the minimum-variance feasible design under the budget.Source: §A.3 · pp. 14–15

A.4 CALIBRATION COST AND BREAK-EVEN CONDITION

  • Separate system-level calibration is worthwhile only when its cost can be amortized across enough evaluations and its cost-scaled MSE is below the shared rule's.If the pilot rule's cost-scaled MSE is at least the shared rule's, no positive break-even count exists; calibration cost is reported separately from the formal evaluation budget.Source: §A.4 · pp. 15–16

B TASK FAMILIES AND CONSTRUCTION OF FROZEN EDIT STATES

The protocol constructs frozen edit states from task specifications and evaluates candidate programs through isolated execution and ordered automated checks. It also specifies state acceptance criteria, quality checks, tolerances, and handling of failures.

B TASK FAMILIES AND CONSTRUCTION OF FROZEN EDIT STATES

  • State construction reads only the task specification, excluding candidate implementation, model identity, and runtime outcomes from state generation.The environments contain eight task families with nine variants each in A and six disjoint families with eight variants each in B; variants vary dimensional ratios and avoid uniformly scaled copies.Source: B TASK FAMILIES AND CONSTRUCTION OF… · p. 16, §B.1 · p. 16

B.1 FOURTEEN TASK FAMILIES AND THEIR EDIT SEMANTICS

  • Environment A has eight task families and B has six disjoint families, with variants changing multiple dimensional ratios rather than uniformly scaling one geometry.The task-family table is described as linking families, parameters, and high-level semantic contracts.Source: §B.1 · p. 16, Table 4 · p. 16

B.2 CONSTRUCTION AND ACCEPTANCE CRITERIA FOR THE FOUR STATE TYPES

  • Each template has 16 frozen parameter vectors: four states of each of four types, with no-ops, out-of-range states, and duplicate vectors rejected.Each state records units, valid ranges, changed parameters, expected relations, and tolerances; evaluation samples states without replacement and does not adapt them to candidate programs.Source: §B.2 · p. 17, Table 5 · p. 17

B.3 STATE-QUALITY AND SPLIT-CONSISTENCY CHECKS

  • Quality checks cover 120 templates and 1,920 frozen states to ensure deeper audits sample the same target population rather than easier or rewritten states.Source: §B.3 · p. 17, Table 6 · p. 17

C ISOLATED EXECUTION AND AUTOMATED JUDGE

C.1 ISOLATED WORKER AND GEOMETRY PROBES

  • Infrastructure errors may be retried once, but timeouts and model or geometry failures are not; an invalid nominal program makes all 16 counterfactual states failures.Source: §C.1 · p. 18

C.2 SIX-STAGE AUTOMATED-JUDGMENT CONTRACT

  • The judge uses fixed geometric tolerances and routes records within a 0.90T–1.10T threshold gray zone to manual review.Binary-outcome analyses exclude gray-zone records until a review label is available; discrete topology counts use exact rules.Source: §C.2 · p. 18, Table 8 · p. 18

D SYSTEM MAPPING, RUNTIME ENVIRONMENT, AND FROZEN PROTOCOL

  • The execution protocol records state identifiers, edit types, labels, failure stages, timing, and diagnostics after nominal and selected-state builds.The system-mapping table identifies five generation systems.Source: §C.3 · p. 18, Table 9 · p. 19

D.2 ENVIRONMENTS AND DATA MATRIX

The supplied materials identify tables for the complete pools in the two environments and the frozen execution protocol. They also describe runtime-cost measurement, template-level data splitting, and expert validation.

D.2 ENVIRONMENTS AND DATA MATRIX

  • Table 10 presents the complete pools for the two frozen environments.Source: Table 10 · p. 19

D.3 EXECUTION, COST, AND RANDOMNESS PROTOCOL

  • Per-run wall-clock time is the raw cost measure, and calibration/test splits keep all records from a template together.The main cost estimate uses a 10% trimmed mean, with ordinary mean and median as sensitivity checks.Source: §D.3 · p. 19, Table 11 · p. 19

E EXPERT VALIDATION

  • Expert validation uses a stratified sample of 800 records, independently reviewed by two parametric-CAD experts and reconciled through predefined adjudication.Source: E EXPERT VALIDATION · p. 20

E.1 WEIGHTED METRICS AND EXPERT AGREEMENT

The automated judge's main bias is missed failures, with recall decreasing as system difficulty increases despite precision of 1.000 for S3–S5. Supplementary analyses describe weighted expert comparisons, calibration costs, and split stability and task-family extrapolation.

E.1 WEIGHTED METRICS AND EXPERT AGREEMENT

  • The 800 doubly labeled expert records yield Cohen's κ≈0.797, and expert-reference and automated-judgment risks use the same weighted sample.The weighted confusion metrics and both risks therefore share a target population and weighting scheme.Source: §E.1 · p. 20
  • Precision remains 1.000 for S3–S5, while automated-judge recall decreases as system difficulty increases, making missed failures its main bias.The authors report automated-judgment risk as a reviewable model-level estimate alongside expert-reference risk.Source: §E.1 · p. 20, Table 12 · p. 20

F SUPPLEMENTARY EXPERIMENTS, ROBUSTNESS, AND MODEL COMPARISON

  • The supplementary experiments collect cost, stability, and model-comparison results omitted from the main text, with exact main-text figure values in Appendix F.6.Source: F SUPPLEMENTARY EXPERIMENTS,… · p. 20

F.1 CALIBRATION COST

  • Full calibration serves as a mechanism-analysis reference, while small calibration tests whether system-level configurations can amortize added cost across repeated evaluations.Source: §F.1 · p. 20, Table 13 · p. 20

F.2 SPLIT STABILITY AND TASK-FAMILY EXTRAPOLATION

  • Random-split stability uses 40 family-stratified splits, while leave-one-family-out validation measures changes in task composition rather than ordinary split noise.The random splits produce 200 system-by-split outcomes; each calibration split uses three templates per family and holds out the remaining six.Source: §F.2 · p. 21, Table 14 · p. 21

F.3 HIERARCHICAL INTERVAL EXPERIMENT

Across systems and environments, no audit depth is universally best: variance sources and execution costs determine whether deeper auditing helps. The section also reports interval and allocation protocols, paired-comparison results, and supporting materials.

F.3 HIERARCHICAL INTERVAL EXPERIMENT

  • Interval replays fix q=16, g=3, and k=8 across 5,000 without-replacement repetitions, using identical sampled records and risk estimates for both coverage methods.The methods differ only in whether they retain template-level correlation.Source: §F.3 · p. 21

F.4 FORMAL SELECTION CONTRACT FOR ALLOCATION STRATEGIES

  • Allocation searches share variance estimates, costs, budget, and objective; joint search varies g and k, while fixed-depth search constrains k.The held-out oracle is post hoc and does not select configurations; transfer freezes (g,k) and remaps q using measured costs.Source: §F.4 · p. 21

F.5 PAIRED MODEL DECISIONS

  • Paired allocation yields at most sub-percentage-point gains at low and medium budgets, with gains largely vanishing at high budget.It is positioned as a focused review tool for closely matched systems rather than a standard-leaderboard default; pairing can reduce comparison variance, but calibration value depends on separation, cost shift, and reuse.Source: §F.5 · p. 22, §I.1 · p. 27

F.6 EXACT DATA FOR MAIN-TEXT RESULT FIGURES

  • The main-text result figures are backed by appendix tables that report estimation efficiency, candidate configurations, and cross-task-family transfer results.The supplied captions identify table scope but do not provide cell values.Source: §F.6 · p. 22, Table 18 · p. 22, Table 19 · p. 23

G ANONYMOUS EXECUTABLE MATERIALS AND CONSISTENCY CHECKS

  • Automated consistency checks are intended to prevent mismatches in task definitions, state populations, record hierarchy, and analysis inputs, not to replace the statistical argument.The supplementary materials list reproduction entry points and validation contracts.Source: G ANONYMOUS EXECUTABLE MATERIALS AND… · p. 23, §G.3 · p. 25

H.1 EXECUTABLE CAD GENERATION AND EDIT EVALUATION; H.2 BUDGET-CONSTRAINED EVALUATION AND MULTISTAGE SAMPLING

  • The evaluation draws on executable CAD work and multistage sampling, while combining automated execution logs with expert review for semantic or expert-judged labels.The statistical setup accounts for nested templates, generations, and edits, as well as template-preparation, generation, and execution costs.Source: §H.1 · p. 26, §H.2 · p. 26

I.1 MAIN FINDINGS AND PRACTICAL IMPLICATIONS

  • No audit depth is universally best; calibration-based variance and cost estimates explain differing preferences, but joint allocation does not consistently beat strong fixed-depth baselines.Deeper fixed auditing is a strong baseline in the primary environment, while task-family composition can change preferred depth; calibration cost, task composition, and reuse also affect net efficiency.Source: §I.1 · pp. 26–27

I.2 LIMITATIONS

The experiments find that preferred audit depth varies with system, environment, variance structure, cost, and task composition, while joint allocation does not consistently outperform fixed depth. The conclusions are bounded by finite edit-state coverage, modeling assumptions, judge coverage, and calibration costs.

I.2 LIMITATIONS

  • The 16 audited edit states form a predefined finite population and cannot represent continuous parameter spaces or open-ended design interactions.The estimand also depends on the quality of the state generator, valid ranges, and task constraints.Source: §I.2 · p. 27
  • The variance formula assumes prespecified sampling and within-level exchangeability; correlated task difficulty, failure, and cost or nonrandom stopping may require design weights or unit-level cost optimization.Small calibration sets can also yield substantially fluctuating variance components.Source: §I.2 · p. 27
  • The automated judge has limited design-semantics coverage, and validation spans few task families, systems, and execution environments.Transfer across benchmarks and CAD kernels requires additional real experiments; expert review cannot fully remove differences between state definitions and professional judgment.Source: §I.2 · p. 27
  • The estimand covers average failure risk and pairwise model-risk differences, not safety certification, worst-case or continuous-parameter verification, or unified leaderboard ranking.Those settings require critical-state weighting, worst-case analysis, or multi-model uncertainty and multiple-comparison treatment.Source: §I.2 · p. 27

J SUPPLEMENTARY MOTIVATION AND PROTOCOL EXPLANATION

  • Under a fixed budget, more edit checks can reduce total model-level estimation error when template or generation differences dominate, while high edit-state variation and generation cost favor deeper audits.Calibration-based benefit directions agree with held-out observations; three-level evidence distinguishes task, generation, and within-program variation.Source: J SUPPLEMENTARY MOTIVATION AND… · p. 28

J.1 EXTENDED CALIBRATION AND SPLIT PROTOCOL; J.2 EXTENDED CONCLUSION

  • The protocol defines average failure risk across templates, generations, and edits, incorporates finite-population variance and evaluation costs, and jointly searches q, g, and k.Calibration is separated from test-set configuration selection, and pilot overhead is included when assessing whether allocation is worthwhile.Source: J SUPPLEMENTARY MOTIVATION AND… · p. 28, §J.2 · p. 29, §J.1 · p. 29

J.2 EXTENDED CONCLUSION

  • No audit depth is universally best across systems and environments; calibration estimates explain the direction of audit benefits, but joint allocation does not consistently beat strong fixed-depth baselines.The value of system-level and stratified configurations lies mainly in diagnosing deviations and guiding budget adjustment, subject to calibration cost, task composition, and reuse count.Source: §J.2 · p. 29

Related papers

Loading 2609.15122v1…