acceptodds
Under review as a conference paper at ICLR 2027

Record-Edit Robustness under Shared Corruption Budgets

Abstract

Damaged audit records can make a report ambiguous, reverse it, or leave it without a compatible explanation. When missing and incorrect fields share a damage budget, resolving ambiguity requires excluding a label while retaining a feasible world. We characterize this coupled optimization for fixed finite worlds and binary labels. Our key insight is to choose one world compatible with the final record. Any edit reaching a singleton report can be replaced by writes that agree with this survivor and by deletions, without increasing cost or introducing any edited field outside the original edit set. This survivor-aligned normal form preserves both the minimum cost of reaching a singleton target and whether a protected field set blocks that target within the edit budget. We establish its charging boundary and show when a cumulative damage guarantee permits pure deletion. On the same distance-two tables, ordinary anchoring is fixed-parameter tractable in anchor size, whereas shared-budget resolution is W[1]-hard in budget. A supplied narrow joint program provides a tractable route to exact computation. Independently checked computations examine the structure and its costs. A measured recovery case exhausts all six equal-cost field protections: delaying the first nonempty report change and blocking an opposite definite report have no common optimum. Thus the reporting contract determines both edit structure and protection requirements.

Primary area:
causal reasoning
Keywords:
evidence corruption, robust explanations, parameterized complexity, causal auditing, partial identification, certified computation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_E5QsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates record-edit robustness under shared corruption budgets, addressing how to certify or manipulate report outcomes when audit logs contain both missing values (erasures MM) and explicit mismatches (corruptions {0,1}\{0, 1\}).

Critical strengths

S1. Elegantly Formalized Theoretical Mechanics: The paper establishes a sound, mathematically rigorous framework for coupled corruption-edit dynamics. The Survivor-Aligned Normal Form is a genuinely neat theoretical insight that simplifies target search spaces, and the authors commendably prove its precise boundary conditions (Proposition 1) rather than ignoring edge cases.

S2. Sharp Parameterized Complexity Characterization: The paper offers a complete theoretical treatment of the problem's computational hardness, establishing a clear contrast with standard anchoring via a W[1]-hardness proof for explicit tables alongside a working FPT dynamic programming route over ordered binary decision branching programs.

Critical weaknesses

W1. Venue and Topic Fit. The paper's core contributions lie in formal logic, parameterized complexity, abductive explainability, and database causality. It contains no learning dynamics, representation learning, or neural architectures. Submitting this work to mainstream machine learning venues (such as ICLR, NeurIPS, or ICML) is a clear mismatch. The paper is vastly better aligned with formal methods, logic, and database theory venues—such as KR, CAV, SAT, PODS, ICDT.

W2. Restricted Real-World Applicability. Binary Feature Abstraction: Restricting the feature space strictly to binary vectors (H⊆{0,1}r\mathcal{H} \subseteq \{0, 1\}^r) limits direct deployment to real-world database records, financial transaction logs, and audit trails, which are inherently multi-valued, continuous, or relational.

W3. Fragile Unit-Cost Requirement (λ=1\lambda = 1): The core theoretical engine depends entirely on erasing a field (MM) costing the exact same unit budget as intentionally falsifying a value (0↔10 \leftrightarrow 1). Proposition 1 explicitly proves that when λ≠1\lambda \neq 1, the survivor-aligned normal form breaks down. In real-world security and data auditing, erasing a record entry and actively forging a field value carry fundamentally different operational costs, adversary constraints, and detection risks.

W4. Empirically Weak Evaluation Lacking Scalability. In-Memory Baselines Outperform Symbolic Solvers: Experiments are restricted to tiny dimensions (r≤16r \le 16 for SCMs, r≤24r \le 24 for DP programs). At these small scales, a naive in-memory Bitset search dramatically outperforms both General and Normal symbolic SAT solvers (e.g., at r=16r = 16, Bitset finishes in 0.32s vs. Normal SAT's 1.02s and General SAT's 1.27s). The PySAT/Glucose3 solver setup introduces overhead that defeats the purpose of symbolic execution on the tested benchmarks.

W5. Missing Standard Optimization Baselines: The evaluation compares symbolic search against a brute-force Bitset search, but completely ignores standard formal verification and optimization baselines—such as MaxSAT, Integer Linear Programming (ILP/Gurobi), or SMT solvers (Z3, CVC5)—leaving the actual utility of the proposed DP/SAT implementation unverified at scale.

Questions for authors

Critical weaknesses W1--W5.

3 comments
AuthorsOpenReview ↗

Response to Reviewer E5Qs

Thank you for the questions about learned models and optimization baselines.

W1. ... venue ... topic fit ...

Answer: Different inputs compatible with a damaged record can receive different predictions from a trained classifier. We added a Connect-4 experiment (Section 6.4, Table 5, p. 10) asking which squares protect a frozen classifier’s win/non-win report. We erased one square on each of 128 held-out boards. At filter budget one, 118 initially definite reports could reverse within two edits. Two protected squares could block reversal on 56 of those 118 boards. On 27 of the 56, all 861 pairs tied for delaying the first nonempty report change, yet only one to four pairs per board blocked reversal. On the other 29, every first-change optimum blocked it. Appendix I.2, Table 17, identifies the unique blocking pair on one board. A first-change tie can therefore conceal a consequential protection choice.

W2. ... Binary Feature Abstraction ...

Answer: Appendix I.1 extends Theorem 2 to arbitrary finite categorical fields with binary labels and equal missing and mismatch charges within each field. Fix a world surviving the final edit. Erasing a new write that disagrees with this survivor preserves its charge and never lowers any other world’s charge. When erasing an observed field costs at most rewriting it, costs are positive and additive over changed fields, and the replacement is admissible, this preserves the singleton target without increasing cost or introducing any new edited field. Both the minimum cost and protected-set blocking are preserved. The appendix also gives the unit-cost DP bound for a supplied complete ordered multiway program. In the Connect-4 experiment, an observed blank is a third square value, distinct from missing. Spambase uses threshold predicates on numerical features (Appendix I.3, Table 19). Raw continuous edits and edits subject to relational constraints need their own admissibility conditions.

W3. ... Fragile Unit-Cost Requirement ...

Answer: The auditor counts damaged fields. Its charge can vary by field, with missing and wrong values charged equally within a field. For the editor, Appendix I.1 allows erasing an observed value to cost less than rewriting it, provided the erasure is admissible, change costs are positive and additive, and unchanged fields cost zero. Proposition 1 shows why unequal within-field damage charges can break normalization. Operational detection risk is outside this model. We also ran an additional sensitivity experiment at all breakpoints of the auditor's missing charge λ\lambda, keeping mismatch and edit costs at one (Appendix I.5, Tables 21–22). The two protection objectives in Table 4 have disjoint optima throughout 0<λ≤10<\lambda\le1. At λ=1/2\lambda=1/2, protecting trial and restore blocks both reversal and empty support within two edits. At λ=1\lambda=1, every pair permits empty support after one edit. All 18 SCM queries retain their initial reports and target radii for 3/4<λ≤13/4<\lambda\le1. At λ=1/2\lambda=1/2, ten of the eighteen admit no nonempty change within budget.

W4. ... Scalability ... In-Memory Baselines ...

Answer: In the original SCM comparison, Bitset beat both symbolic methods on 16/18 queries. Both symbolic methods were faster on two 16-root layered queries, and Normal had no stable timing advantage over General (Table 2). The original DP study reached 512 fields on supplied programs of width five or seven. All three queries at that size had radius two and took 0.014–0.068 seconds for inference over roughly 1.18×101071.18\times10^{107} admissible strings (Appendix D). Section 6.3 adds two source-compilation experiments. All eight bounded P protocol configurations passed source checking. A shared check settled 32/64 queries, and DP was faster than enumeration on 12 of the 32 requiring search. Four of nine Spambase tree/order configurations passed source checking. Two hit the node cap and three timed out during proof checking. Enumeration was faster on all 32 completed comparisons requiring search (Appendix I.3, Tables 18–19).

W5. ... Standard Optimization Baselines ...

Answer: Section 6.2 adds a weighted partial MaxSAT baseline (RC2) on the original 18 SCM queries. We ran all four methods three times per query with the same external limits of 30 seconds per producer or checker stage, 512 MiB of memory and 32 MiB of output. A common checker rebuilt the labels, verified each witness and exhausted all lower-cost records. All 216 runs passed. Comparing per-query median worker times for preparation, solving and checking, MaxSAT was faster than General on 15/18 queries and Normal on 14/18. Bitset was faster than MaxSAT on all 18. With process startup included, MaxSAT beat Bitset on 5/18. Appendix I.4, Table 20, reports both measures and describes the encoding and proof costs.

Thank you for the detailed rebuttal and additional experiments. While the clarifications and new results are helpful, they do not fully address my main concerns.

First, the proposed model remains restrictive compared to real-world data auditing scenarios, particularly regarding data types and corruption costs. Second, although the additional experiments improve the evaluation, they still do not convincingly demonstrate scalability or clear practical advantages over simpler search methods and standard optimization baselines.

Overall, I appreciate the authors' efforts, but the rebuttal does not sufficiently change my assessment. I will maintain my original rating.

Thank you for considering the additional results and clarifying the remaining concerns. Appendix I.1 extends the setting to finite categorical fields and specifies admissible action costs; equal missing and mismatch charges within each field remain essential. The central contribution is the normal form’s preservation of minimum target-edit cost and protected-set blocking, together with the same-table complexity separation. The additional experiments examine the resulting protection decisions and computational costs, including cases where simpler methods are faster.

Reviewer_q2UZOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies robustness of audit-style explanations when records may contain missing or incorrect fields under a shared corruption budget. Given a finite set of possible worlds and their labels, it asks how many record edits are required to change an ambiguous report into a desired singleton-label report while retaining at least one compatible explanation. The main contribution is a survivor-aligned normal form, showing that optimal singleton-target edits can be restricted to survivor-consistent writes and deletions without increasing cost or expanding the edited field set. The paper further characterizes when this normalization holds, analyzes the computational complexity of shared-budget resolution, and provides tractable algorithms under additional representation constraints. Experiments on synthetic instances and controlled file-recovery scenarios illustrate the theoretical properties and protection trade-offs studied in the paper.

Critical strengths
  1. The normal form preserves both optimization and protection. Choosing a final survivor permits a simple but consequential dominance argument without assuming a persistent core or initial compatibility of the survivor. Preserving the edited-field set gives a stronger conclusion than merely preserving the minimum radius: it also preserves whether a protected set blocks a singleton target throughout the edit budget (Theorem 2; Appendix B.3).
  2. The theoretical boundaries are explicit and informative. The paper proves where equal missing/mismatch charges are essential, separates ordinary-anchor FPT from shared-budget W[1]-hardness on the same table, and gives a tractable algorithm under a clearly stated joint-width assumption. These results provide useful structural knowledge independently of whether the Normal solver is faster (Proposition 1; Theorems 3–4).
Critical weaknesses

The application significance remains narrow relative to the theoretical characterization. Equal charging of erasure and mismatch is essential, and the stronger factual guarantee uses a cumulative budget condition. The most distinctive file-protection conflict is demonstrated with a fixed filter outside that sufficient cumulative guarantee. Moreover, the eight generated plans yield only three nonconstant response functions and one constant-safe function; the thousands of query rows do not represent thousands of independent workflows (Section 6.4; Appendices H.1–H.3). This limits the evidence that the proposed protection choices are consequential for practical auditing, although it does not invalidate the finite-domain theorems.

Questions for authors
  1. Scope of the world model assumption. The formulation assumes a fixed finite world domain HH and a label function y:H→{0,1}y:H→\{0,1\}, meaning that the possible explanations and their labels are available before analyzing corrupted records. In realistic auditing scenarios, however, the main difficulty is often that the hypothesis space itself may be incomplete or unknown. Could the authors clarify whether the goal is to provide robustness guarantees within a given formal model, rather than to infer real-world causes from incomplete evidence? What assumptions are required for the proposed guarantees to remain meaningful when the candidate world space is only an approximation?

  2. Positioning relative to existing reasoning frameworks. The proposed formulation appears closely related to hypothesis elimination, model checking, database consistency, and robust explanation frameworks. Could the authors clarify what aspects of shared-budget resolution fundamentally differ from these existing perspectives? In particular, is the primary contribution the survivor-aligned normal form itself, the corruption-budget formulation, or the protection analysis enabled by the formulation?

3 comments
AuthorsOpenReview ↗

Response to Reviewer q2UZ

Thank you for asking how the protection result carries over when the world model or the reporting rule changes.

W1. ... application significance ... independent workflows ...

Answer: The original eight file plans yield three nonconstant response functions and one constant-safe function (Figure 2 and Appendix H.2). The reported query counts measure coverage within those plans. The conflict in Table 4 occurs at a fixed filter of b=2b=2, with initial damage bound b0=2b_0=2 and two further edits. In an additional missing-charge experiment, the two protection objectives retained disjoint optima throughout 0<λ≤10<\lambda\le1 with unit mismatch charge (Appendix I.5, Table 21). We also added a Connect-4 experiment on a frozen classifier. Table 5 (Section 6.4, p. 10) compares two filter budgets on the same 128 held-out boards. At b=1b=1, 118 of 128 initially definite reports could reverse within two edits. On all 56 boards where a pair could block reversal, the two objectives shared an optimum. On 27 of those boards all 861 pairs tied on the first nonempty change radius, yet only one to four blocked reversal. In the example in Appendix I.2 (Table 17, p. 34), protecting either (a1, a2) or (d1, e4) delayed that change until edit two. Only (d1, e4) blocked reversal, and both permitted empty support after one edit. At b=3b=3, the cumulative bound (b0=1b_0=1, a=2a=2) retained every original board. Those same 118 reports then started ambiguous. The remaining ten could lose certainty, and no pair delayed their first loss.

Q1. ... world model ... approximation ...

Answer: The radii are exact for the specified finite world domain and label map. A factual interpretation of a singleton requires that the model include the factual world with its correct label and that the world satisfy the damage bound at the edited record. We clarified the case of an approximate domain in Section 3. If it includes every admissible world, preserves their labels, and uses the same fields, compatibility rule and edit permissions, its report contains the true-domain report at each allowed record. If the larger model excludes a label at every allowed edit, the admissible domain excludes it at every such edit as well. The radii and optimal field choices can still change. Omitting a compatible world of the opposite label can turn an ambiguous report into a spurious singleton.

Q2. ... existing reasoning frameworks ... contribution ...

Answer: The central contribution is the survivor-aligned normal form and its charging boundary. Hypothesis elimination filters candidate worlds, model checking can verify an edit, consistent query answering ranges over admissible repairs, and robust explanations identify features sufficient for a prediction under perturbations. Our optimization edits the record while holding the worlds, their labels and the filter fixed. Each erasure consumes budget that could otherwise cover a mismatch. The same edit must retain one world and exclude every world of the opposite label. We clarified these two constraints beside Eqs. (11)–(12) in Section 4.2. Theorem 2 shows that the final survivor determines a normalized witness of no greater cost using a subset of the edited fields. It follows that any protected fields avoided by a successful edit are also avoided by its normalized witness (Appendix B.3). On identical distance-two tables, ordinary anchoring encodes Vertex Cover and shared-budget resolution encodes Independent Set (Section 4.3).

Thank you for the detailed response and additional experiments. The clarification of the finite-world assumption and its implications under approximate world models addresses my first question. The explanation of how shared-budget resolution differs from hypothesis elimination, model checking, and robust explanation frameworks also helps clarify the paper's main contribution.

The additional Connect-4 experiment provides useful evidence beyond the original file-recovery scenarios, although the practical evaluation remains somewhat limited in scope. Overall, I appreciate the authors' efforts to address my concerns and clarify the theoretical contribution and its applicability.

Thank you for taking the time to revisit the paper and the Connect-4 experiment. We’re glad our response clarified how the guarantees depend on the choice of world model and how the normal form relates to existing reasoning frameworks.

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

The paper studies binary audit-style reports computed from damaged records. Records have Boolean fields, some possibly missing (M) or simply wrong; a world counts as compatible with the record if the number of missing plus wrong fields stays within a single shared budget b, and the record's "report" is the set of labels over its compatible worlds. The main object of study is the edit radius: how many field edits (fill, flip, delete) it takes to change the report to some target, e.g. to make an ambiguous report definite. The authors prove a characterization of labels that survive any legal edit (a "persistent core"); a normal form saying every successful edit can be rewritten as aligning to one surviving world plus deletions, which rather surprisingly works only when missing and wrong fields are charged the same (λ=1); a W[1]-hardness result in the budget, with a neat same-table contrast against ordinary anchors (vertex cover vs independent set on the same table); an exact dynamic program that is FPT in the program width and the budgets; and a protection analysis showing that the fields one would freeze to delay any report change are not the ones one would freeze to block the opposite report. The experimental side is small-scale certified computation: 18+54 synthetic queries and one measured 5-field file-recovery case study.

Critical strengths

S1. The theory is careful and honest about its own limits. The normal form (Thm 2) comes with the exact boundary of its validity — Prop 5 shows completeness holds iff λ=1 and gives counterexamples at every other value — and Cor. 1 / Prop. 6 pin down precisely when pure deletion is enough. The proofs are theoretically correct; I also verified the main claims numerically, checking Thm 1's core formula, Thm 2's radius equality, Cor. 1, and both hardness constructions against brute-force computation on small random instances, and found full agreement.

S2. The same-table contrast (Prop 7) is a nice observation. On one explicit labeled table, the ordinary anchor query boils down to vertex cover (FPT in k) while the shared-budget radius query boils down to independent set (W[1]-hard in k). That is a convincing, concrete way to show that sharing the corruption budget between missing and incorrect fields changes the complexity landscape for real, rather than just renaming an existing problem.

Critical weaknesses

W1. My main concern is the fit with this venue. The motivation is audit-log integrity; there is no learning anywhere in the paper, and the validation is tiny certified computations (edit balls of ~2k records, one 5-field case study). Since the general problem is W[1]-hard, whatever practical mileage exists sits in the narrow-width DP regime, and the paper gives no evidence that realistic models compile to small width. As written, this is more like a theoretical CS / database-theory paper, and the transfer to problems the ICLR community works on is asserted rather than shown.

W2. The paper is somewhat difficult to read, well beyond what the material requires. The prose is telegraphic, and much of the real content is carried by notation alone — for instance, the correctness of Theorem 2 hinges on the retention and exclusion constraints in Eqs. (11)–(12), specifically on which deleted positions agree with each world, but this is never explained in plain terms. The same goes for the basic intuitions: why deletion dominates filling under unit charging, and what the three budgets each mean, are left for the reader to reverse-engineer from the algebra. I needed the appendices and several passes to assemble the picture, and I expect many readers will not invest that effort.

Questions for authors

Q1. Everything assumes λ=1, which Prop 5 shows is the only point where the normal form is complete. But intuitively an incorrect field is worse than a missing one — with a missing field you at least know where the damage is (cf. the erasure-vs-error asymmetry in coding theory). How sensitive are the protection conclusions (Table 4) and the Sec. 6 findings to λ≠1? If the answer is "very", that weakens the practical claims; if there is a weighted normal form, that strengthens the paper.

Q2. Thm 4 is FPT in the joint width w with compilation excluded. Is there any evidence that realistic SCMs compile to narrow joint programs? Otherwise the theorem is more of a possibility result.

2 comments
AuthorsOpenReview ↗

Response to Reviewer uNjb

Thank you for independently checking the normal form and hardness constructions against exhaustive search.

W1. ... fit ... venue ...

Answer: The ML question is which registered fields protect the report of a trained classifier when several inputs fit the record. We added a Connect-4 experiment using the predictions of a frozen random forest on 67,557 registered boards (Section 6.4, Table 5, p. 10). After one square was erased on each of 128 held-out boards, all reports were definite at b=1b=1. Within two edits, 118 could reverse. On 56, a protected pair could block reversal while also maximizing the first nonempty report-change radius. On 27 of those 56 boards, all 861 pairs tied on that radius, yet only 1–4 blocked reversal. Uniform tie-breaking would block reversal with probability below 0.5% on each of these boards. Exhaustive enumeration checked all 7,876 candidate records per board without a width assumption (Appendix I.2, Table 17, p. 34).

W2. ... retention ... exclusion constraints ...

Answer: Section 4.2 now explains Eqs. (11)–(12) beside Theorem 2. In Figure 1, writing zero in the first field of MM0MM0 moves 000 from distance two to one. Erasing the last field brings 000 back to two and moves 100 from two to three, excluding it at b=2b=2. Each aligned write gains the survivor one agreement. Each deletion removes an agreement only for worlds that matched the erased value. An opposite-label world at distance k≤bk\le b after the writes must lose at least b−k+1b-k+1 agreements. Erasing a new write that disagrees with the survivor leaves its distance unchanged and cannot readmit an excluded world. Undoing a fill saves an edit, and replacing a flip by a deletion costs the same. The example added in Section 3 distinguishes the initial damage bound b0b_0, edit budget aa, and filter budget bb. When b≥b0+ab\ge b_0+a, a factual world within the initial bound survives every allowed edit.

Q1. ... sensitive ... protection conclusions ...

Answer: We added an exact sensitivity experiment on the missing charge λ\lambda for Table 4 and the 18 SCM queries, holding mismatch and edit costs at one. We checked every breakpoint and intervening interval (Section 6.4 and Appendix I.5, Tables 21–22). The protections optimal for delaying the first change and for blocking reversal are disjoint in the file case throughout 0<λ≤10<\lambda\le1. At λ=1/2\lambda=1/2, protecting trial and restore blocks both reversal and empty support within two edits. At λ=1\lambda=1, every pair permits empty support in one edit. Above one, the initial file support is empty. All 18 SCM queries retain their four target radii for 3/4<λ≤13/4<\lambda\le1. At λ=1/2\lambda=1/2, ten admit no nonempty change within budget. The supplied-program and Connect-4 experiments use unit charging. Proposition 1 allows different positive weights across fields when missing and mismatch charges agree within each field. Universal completeness at unit edit cost requires λ=1\lambda=1. Appendix I.1 also allows positive additive action costs when unchanged fields cost zero, erasure costs at most rewriting, and the normalized record is admissible.

Q2. ... realistic SCMs ... narrow ... programs ...

Answer: We added two source-compilation experiments in Section 6.3 (pp. 8–9). In eight bounded P TwoPhaseCommit configurations, we ran factual and intervened executions with shared controls, suppressing one participant’s halt in the intervention. All eight yielded source-checked programs of width 3–7 over 19 or 31 fields. From the built executable, building and checking the program took 1.70–18.68 s per configuration. DP and enumeration certified the same 64 answers. Among the 32 requiring search, DP was faster on 12 (median checked worker-time ratio, DP/enumeration, 1.02). On learned Spambase trees, four of nine tree/order combinations passed source checking. Only one depth-six order passed, at width 118 over 36 threshold fields in 47.24 s. Two combinations hit the node cap and three timed out in proof checking. Enumeration was faster on all 32 completed tree comparisons requiring search (Appendix I.3, Tables 18–19). This establishes narrow width for the bounded protocol. Typical production SCM widths remain unknown. Theorem 4’s FPT bound in (w,a)(w,a) or (w,b)(w,b) applies after compilation.

Thank you again for your careful review. Regarding the connection to ML you raised in W1, the example below Table 5 (p. 10) shows why the choice of protected inputs matters for a frozen classifier. Two pairs of Connect-4 squares delayed the first nonempty report change equally, but only one blocked an opposite definite report within two edits. We verified this directly by enumeration, without assuming a narrow program. If you have any remaining questions about this example or the compilation results in Section 6.3, we’d be happy to clarify.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the author response and revisions

Thank you for the careful reviews. The central result is that every edit reaching a singleton report has a survivor-aligned witness of no greater cost using only fields the original edit touched. This preserves both the minimum edit cost and which protected field sets block the target. The revised manuscript marks additions in blue.

We added a Connect-4 experiment to show how this distinction affects a trained classifier. After one square was erased on each of 128 held-out boards, 118 reports could reverse within two edits at the fixed filter budget b=1b=1. Protecting two squares could block reversal on 56 of those boards. On 27 of the 56, all 861 pairs tied for delaying the first nonempty report change, yet only one to four pairs per board blocked reversal. At b=3b=3, the cumulative budget retained every original board. Table 5 compares the two budgets, and Table 17 gives a concrete choice of squares. We checked the protection results by enumeration without assuming a narrow program.

We also examined how the results depend on field types and costs. Appendix I.1 extends the normal form to finite categorical fields with binary labels under the stated charging, edit-cost, and admissibility conditions. It also gives a unit-cost DP bound for a supplied multiway program. In an exact sensitivity analysis, the two protection objectives in the file-recovery case had disjoint optima at every missing-field charge 0<λ≤10<\lambda\leq1. All 18 SCM queries retained their four target radii for 3/4<λ≤13/4<\lambda\leq1, with mismatch and edit costs fixed at one (Tables 21–22).

Two new compilation studies measure the cost of building and checking joint programs from a bounded two-phase commit protocol and learned Spambase trees. All eight protocol configurations yielded source-checked programs of width 3–7 (Tables 18–19). We also added a weighted partial MaxSAT baseline on the original 18 SCM queries. A common independent checker verified all 216 runs across four methods (Table 20).

Finally, Section 3 distinguishes the initial damage bound, edit budget, and filter budget, and explains which guarantees carry over when the candidate domain includes every admissible world with its correct label. Section 4.2 now uses Figure 1 to explain the two constraints behind the normal form in plain terms: the edit must retain a compatible world of the target label and exclude every compatible world of the opposite label.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.