acceptodds
Under review as a conference paper at ICLR 2027

CARTER Certifies Partial Taxonomy Repairs: Complete Decisions and Witnesses under Noisy Feedback

Abstract

Repairing a taxonomy requires deciding which parent relations to change and which to leave unresolved. With noisy feedback, agreement among surviving trees can increase after the target has been discarded. We formalize this failure as confident erasure and introduce CARTER, an exact interface that turns a calibrated family of primary-parent trees into complete edge certificates and opposing witnesses. A session-calibrated disagreement budget preserves target retention and all certified edits at every prefix with marginal probability at least , under candidate inclusion and exchangeable sessions from a frozen pipeline. Errors within a session may be dependent. Edge-additive disagreement reduces certification to constrained arborescence optimization, recovering every edge statement shared by the retained family without enumerating trees. An optional shared determinant circuit computes exact one-step query values. On a sealed structural holdout of 224 taxonomies, six-source macro-average whole-tree coverage is 91.5–93.5% at a 90% target, versus 34.6–41.2% for edge-pooled calibration. A separate development simulation adds 2.095 final certified edits per session over no feedback, with lifetime wrong edits in 2/148 sessions. These results connect structured uncertainty to inspectable partial repairs, with explicit witnesses for deferred decisions.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
conformal prediction, structured prediction, uncertainty quantification, arborescence optimization, version spaces

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_XEjmOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CARTER, a method for partially repairing taxonomy trees under noisy feedback. It retains a calibrated set of plausible trees and certifies only changes supported by every remaining tree, providing opposing example trees for unresolved decisions.

Critical strengths

The paper identifies an important and interesting failure mode: incorrect feedback can eliminate the correct taxonomy while increasing agreement among the remaining candidates in sec 2. This motivates preserving plausible alternatives rather than treating agreement as sufficient evidence for repair. CARTER addresses this problem by allowing a calibrated number of feedback disagreements and certifying only changes shared by all retained trees in sec 2 . In the simulated-feedback study in sec 3, it adds approximately 2.1 certified edits per session over no feedback, while producing an incorrect edit in 2 of 148 sessions. These results support its ability to make useful repairs while limiting errors.

Critical weaknesses

The practical evaluation is limited. Feedback comes from simulated answerers in sec 3, and CARTER makes no repairs on the six genuine taxonomy transitions in sec 6. It therefore remains unclear whether the method can make useful repairs with real feedback, rather than stay safe by avoiding changes.

The guarantee requires the correct taxonomy to be among the candidates. However, when candidate recall drops to 94.2%, the reported session-level wrong-certificate rate reaches 41.9% in sec 6. This raises concerns about reliability when candidate generation misses correct parent relations.

Questions for authors

The purpose and relevance of several implementation details in the appendix are difficult to understand. The authors should explain how these details support reproducibility or the paper’s scientific claims. Examples include:

  1. Lines 1010–1013 report an artifact-set SHA-256 hash and distinguish development checks from sealed-holdout evidence, but their relevance to the surrounding discussion is unclear.
  2. Lines 1024–1025 provide a session identifier digest without clearly explaining what it identifies or how readers should use it.
  3. Similar clarification is needed for lines 1105–1107, 1593–1597, and 1620–1622.
1 comment
AuthorsOpenReview ↗

Response to XEjm: Practical evidence, candidate failure, and appendix navigation

Thank you for identifying target retention as the key safety issue and recognizing the useful edits in the programmed-feedback study. We address the practical evidence, candidate failure, and appendix navigation below.

  1. Useful repair and genuine temporal changes

The 148-session programmed-feedback study adds 2.088 canonical-correct certified edits per session over no feedback, while a lifetime wrong edit occurs in 2/148 sessions. This measures useful committed repair under the declared simulated answer regime. The separate 224-taxonomy structural holdout evaluates repair without feedback.

On the six genuine MeSH/OEWN transitions, CARTER emits zero edits. The retained allowed graphs contain 68.6–100% of complete candidate edges, leaving no unanimous actionable change. This is the interface's deferral behavior under broad uncertainty, and its zero-error rate measures abstention. Genuine changes paired with real noisy feedback remain a distinct empirical question; the positive structural and programmed-feedback results establish utility in their respective tested settings.

  1. Why 94.2% candidate recall can coexist with 41.9% wrong sessions

The reported 0.941588 recall is child-weighted edge recall. The theorem's premise is whole-tree inclusion: every target parent edge must be available. In this stress test, 142 of 2,431 true-parent edges are removed, and initial whole-tree containment falls to 0.041. Average edge recall therefore does not measure the event protected by the guarantee.

The 0.419 wrong-session rate is a substantive failure: trees that all omit a true parent can still agree on an incorrect replacement or deletion. A calibrated feedback budget cannot restore an edge excluded upstream. The main experiments use complete candidates within each local node set, ensuring target inclusion by construction and isolating the calibration layer. Application to a restricted generator requires a separate audit of whole-tree inclusion and adequate candidate expansion; the stated guarantee applies when the target-containing candidate premise holds. The stress test makes this prerequisite measurable rather than treating per-edge recall as sufficient evidence of reliability.

  1. What the appendix identifiers mean and how to use them

The hashes identify specific experimental objects so a reader can match a reported result to its saved inputs or outputs. The scientific support comes from result tables, exactness checks, and paired implementation parity. Each identifier has a different role:

Lines 1010–1013: the artifact-set identifier in results/liar_query_dev_v1c/metadata.json binds the recorded trial, oracle, summary, and README files. Design and graph metadata have separate fields. The surrounding distinction tells readers that these are query/transcript implementation checks, with a different evidential role from the sealed structural holdout.

Lines 1024–1025: the digest identifies the sorted 148 reliability-calibration session IDs for iid_panel/ambiguity_block, recorded in results/joint_e2e_public_dev_v1/reliability_budgets.csv. It identifies which sessions selected the feedback budget of one; it is distinct from the 148 evaluation sessions and from a digest of their transcripts.

Lines 1105–1107: the identifier refers to the historical GNN multiseed aggregate result manifest. The released numerical table is results/public_gnn_multiseed_dev_v1/graph_seed_averaged.csv, which supports reaggregation of the learned-scorer results.

Lines 1593–1597: the two identifiers refer to the frozen configuration payload and output artifact manifest in results/adjoint_compiler_scaling_dev_v2b/. The configuration identifies the scaling grid; the manifest identifies saved measurements, summaries, and parity records. The relevant evidence for shared computation is agreement with the reference alongside the matched timing results.

Lines 1620–1622: this identifies the historical results/compiler_scaling_dev_v1c/ reference-compiler/certificate bundle underlying Table 11. It assigns those timings and declared computation caps to that implementation; the current shared compiler is evaluated separately.

For reproduction, start with REPRODUCE.md, select the study and its CSV/README, then use the associated identifier to match the saved object. A digest alone supplies identity, rather than evidence of accuracy or a fresh experimental replay.

Together, the useful-edit results and candidate stress test establish the capability and its operational prerequisite: calibrated partial repair over a target-containing family, with unresolved alternatives exposed through witnesses. We would appreciate reconsideration of the soundness assessment in light of this distinction.

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

This paper formalizes confident erasure: an update discards the correct taxonomy while creating new unanimous conclusions among the remaining candidate trees. To address this problem, it introduces CARTER, which maintains a calibrated family of primary-parent trees and tolerates a calibrated number of disagreements with noisy feedback. The method computes exact certificates for additions and deletions supported by every retained tree and provides opposing feasible trees to explain unresolved decisions. Under target-inclusion and exchangeability assumptions for a frozen pipeline, it guarantees simultaneous correctness of all certified edits. Experiments on 224 test taxonomies report source-macro whole-tree coverage of 91.5–93.5%.

Critical strengths
  1. The paper connects target retention to simultaneous correctness of all certified edits and provides explicit witness trees that make unresolved decisions inspectable.
Critical weaknesses
  1. The single-parent projection limits the practical meaning of certified repair. The guarantees concern a predefined primary-parent tree. For sources that permit multiple valid parents, the benchmark selects one parent using minimum root depth, with node identifiers breaking ties. Consequently, a deletion can be correct relative to this projection while removing a semantically valid relation from the original source hierarchy. The paper explicitly acknowledges this restriction, but the evaluation does not establish whether the chosen projection represents the intended taxonomy-maintenance task. An operational justification for the projection, together with a quantitative assessment of how certified deletions affect valid source relations, would clarify the applicability of the results.

  2. The genuine release-transition evaluation does not demonstrate useful certified repairs. In the six MeSH/OEWN release-transition instances, CARTER certifies zero edits. Its reported zero wrong-certificate rate therefore reflects abstention and does not establish that the method can make correct, useful changes in this setting. This result should be distinguished from the positive results obtained using artificial corruption and programmed feedback.

Questions for authors
  1. What practical maintenance requirement justifies the chosen single-parent projection, and how often would certified deletions remove relations that remain valid in the original source hierarchy?

  2. To what extent is CARTER’s absence of certified edits on genuine release transitions caused by weak parent scores versus insufficient calibration data?

1 comment
AuthorsOpenReview ↗

Response to eehG: Operational meaning of target and source of temporal abstention

Thank you for recognizing the target-retention guarantee and inspectable witnesses. Your questions concern the operational meaning of the target and the source of temporal abstention. We address them separately.

  1. What is being maintained, and can a deletion remove a valid source relation?

CARTER maintains a designated primary-parent tree view: every non-root concept has one assigned parent, as required by a tree-valued navigation or export. Within this task, a deletion certifies exclusion from that view. Semantic validity in the source polyhierarchy is a separate property. A source-valid alternative parent can therefore be absent from the maintained view while remaining valid in the source.

The benchmark's minimum-depth/identifier rule makes this target deterministic and independent of feedback. It provides a reproducible view definition; an application must designate the view it actually intends to maintain. The guarantee then applies to that declared target under the assumptions of Theorem 1.

We traced the 269 final certified deletions in the 148 canonical iid sessions back to source relations. One deletion removed a source-valid canonical parent, hence was a wrong certificate; zero removed a source-valid noncanonical parent. The exposure qualification matters: these tree-corruption inputs contained zero source-valid noncanonical edges in their observed graphs, although the local MeSH source DAGs contained 20 such alternatives. Thus, the zero count characterizes the tested tree-view repair task and does not estimate deletion risk when repairing a source DAG containing alternative parents.

The genuine-transition comparison makes the distinction concrete. Of 12 old-primary edges absent from the new canonical trees, two remain valid in the new source hierarchy. Those are changes of view membership. This is a reference-tree comparison: CARTER itself certified zero edits on these transitions and deleted neither relation.

  1. Why no certified edits on genuine transitions: scores or calibration size?

The immediate mechanism is a broad retained family with no unanimous actionable change. Actual calibration pooled five graphs at alpha_s=0.20 and used rank 5, the largest observed calibration score, giving a finite threshold. Across the six test instances, the allowed graphs retain 68.6–100% of the complete candidate edges. The certificate interface accordingly defers changes on which feasible trees disagree.

Both score quality and calibration resolution are relevant. The temporal local-argmin parent accuracy is 0.352, compared with 0.923 for retaining the old tree, indicating weak parent discrimination in this track. Five calibration graphs also force a coarse, maximum-score threshold at the actual level. The available runs contain no controlled scorer-by-calibration-size comparison, so they do not identify a numerical share attributable to either factor.

We agree that zero temporal errors here measure abstention. Useful committed output is established in separate settings: the 148-session programmed-feedback study adds 2.088 canonical-correct edits per session over no feedback, with a lifetime wrong edit in 2/148 sessions; the 224-taxonomy structural holdout evaluates repair without feedback. These studies quantify committed output under the structural-corruption and programmed-feedback protocols.

These results support CARTER's intended capability: useful partial repair when the retained family resolves an edit, and explicit deferral when it does not. The operational target and the two evaluation settings above delimit that capability precisely.

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

The paper studies partial repair of fixed primary-parent taxonomies under noisy feedback and defines confident erasure: an update removes the target tree from a previously covered family while adding new unanimous edge statements. CARTER improves whole-tree coverage and reduces wrong certificates relative to edge-pooled calibration in the structural holdout.

Critical strengths

Critical Strength 1: Clear safety goal and theoretical guarantee

The paper formalizes partial taxonomy repair under noisy feedback and distinguishes target loss from wrong certified edits. Under the stated assumptions, Theorem 1 provides a marginal guarantee for target retention and the correctness of all certified edits at every interaction prefix. This gives a clear safety goal for structured repair (Section 2.4; Appendix B.6).

Critical Strength 2: Complete edge decisions and feasible witnesses

Theorem 2 returns all mandatory, possible, and impossible edges of the retained tree family, with opposing feasible witness trees for unresolved edges. This supports partial repair without selecting one complete tree. The paper checks these decisions against enumeration and reference solvers. In the programmed-feedback study, the method adds 2.088 correct edits per session over no feedback, with wrong edits in 2 of 148 sessions (Section 2.5; Table 2; Appendix G).

Critical weaknesses

Critical Weakness 1: Limited new algorithmic contribution

The complete edge-classification method in Theorem 2 uses constrained minimum-arborescence tests, an approach already discussed for exact filtering of cost-bounded arborescences by Houndji et al. (2017, The Weighted Arborescence Constraint, Section 3). Houndji et al. define the same cost-bounded tree constraint and discuss exact filtering through repeated optimization, although their proposed filtering method mainly uses reduced costs. Using CARTER's disagreement counts as edge costs gives the same feasible tree family and supports the same edge classifications. The session-budget calibration is also shown to be equivalent to a special case of binary conformal risk control (Appendix B.6). Theorem 3 provides a specific shared computation method, but cost-layer tree counting and shared directed-tree derivative computation build on existing methods (Delaite and Pesant, 2017, Counting Weighted Spanning Trees to Solve Constrained Minimum Spanning Tree Problems; Zmigrod et al., 2021). Thus, although the complete interface is useful, its main algorithmic components are largely established, and the additional algorithmic contribution is limited.

Critical Weakness 2: Limited evidence of useful repair in realistic settings

The paper demonstrates the possibility of confident erasure, but does not establish its importance in real taxonomy repair or the practical benefit of protecting target retention beyond controlling wrong edits. The 224-taxonomy holdout contains no feedback, while the 148-session study uses programmed feedback without the binary query compiler. The integrated query experiments use five-node graphs with target inclusion imposed, and the repeated-calibration results do not establish an adjusted advantage over entropy (Table 22). In the six genuine temporal transitions, CARTER certifies no edits (Table 10). These experiments support the method under the tested conditions, but do not sufficiently demonstrate that the complete interface can produce useful repairs while maintaining safety on real taxonomy changes.

Questions for authors

Q1: Clarification for Algorithmic contribution

Houndji et al. (2017, Section 3) define the same cost-bounded arborescence constraint and discuss exact filtering through repeated minimum-arborescence optimization. Applying this approach to CARTER's disagreement costs appears to give the same feasible tree family and complete edge classifications. What additional algorithmic result does Theorem 2 provide? For Theorem 3, what is the specific new computation beyond the cost-layer and edge-inclusion counts of Delaite and Pesant (2017), and the established shared derivative computation for directed trees?

Q2: Real world importance of confident erasure

Figure 1 and Definition 1 show that confident erasure can occur, but the paper does not report how often it occurs in real taxonomy repair. Is there evidence from genuine taxonomy changes and noisy feedback that shows the frequency of this event? If real feedback data are unavailable, can the authors report its frequency in the existing programmed-feedback studies? Please distinguish confident erasure from target loss and wrong certified edits.

Q3: Practical benefit of preventing confident erasure

Table 2 shows that CARTER improves target retention, reduces wrong certified edits, and increases certified edits compared with hard filtering in the programmed-feedback study. However, these results do not separately establish the benefit of preventing confident erasure. Can the authors show whether preventing confident erasure leads to better subsequent repair decisions, even when no wrong edit has yet been emitted? For example, do the saved interaction results show that target loss affects later correct edits, wrong edits, or conflicts?

2 comments
AuthorsOpenReview ↗

Response to MSwg (Q1): Algorithmic contribution of Theorems 2–3

Q1. Algorithmic contribution of Theorems 2–3

CARTER contributes a session-calibrated contract for partial repair: preserve the designated target at every interaction prefix, return all unanimous edge decisions of the retained family, and provide opposing complete-tree witnesses for every unresolved edge. Theorems 1–3 make this contract both statistically protected and exactly computable. Under the stated target-inclusion and frozen-pipeline exchangeability assumptions, Theorem 1 gives one marginal event, of probability at least 1−αs−δh1-\alpha_s-\delta_h, on which the target is retained and every certified addition/deletion is correct, simultaneously across the entire session.

Theorem 2: a complete constructive repair result. Houndji et al. (2017, §3) are the direct antecedent for exact filtering of a cost-bounded arborescence family. Substituting our accumulated disagreement costs indeed gives the same feasible family and supports the same classifications. Theorem 2 specializes that principle to the repair contract. Repeated set-valued answers compile into additive tree costs. One baseline optimum plus one opposite restriction per allowed edge returns all mandatory, possible and impossible edges in at most ∣D0∣+1|D_0|+1 arborescence calls. Every unresolved edge comes with both a containing and an excluding budget-feasible tree. These witnesses respect the complete tree structure and the shared disagreement budget. The interface also handles the frozen universe CC versus allowed edges D0D_0 and suppresses operational certificates on conflict. The result delivers this complete, inspectable output within the session-protected family.

Theorem 3: exact compilation of the entire binary-query step. Delaite and Pesant (2017, §§3–4.1) already give cost-layer and edge-inclusion counts; Zmigrod et al. (2021, Proposition 3 and §5.2) establish shared directed-tree derivatives. Our specific computation starts from a repeated, possibly contradictory binary transcript, compiles its nonnegative integer edge costs, and evaluates one division-free determinant circuit plus one reverse sweep in Rb=Z[z]/(zbqry+1)R_b=\mathbb Z[z]/(z^{b_{\rm qry}+1}). It obtains every disagreement-layer count and every edge-layer count, then projects them into both horizon-(h−1)(h-1) Berlekamp successor weights for every admissible edge query, giving exact one-step minimax selection. The truncated-ring construction preserves all required coefficients even with zero divisors; the proof uses no division or cancellation.

With d=∣V∣−1d=|V|-1 and the edge costs supplied, the bound is O(d4(bqry+1)2+∣D0∣(bqry+1))O(d^4(b_{\rm qry}+1)^2+|D_0|(b_{\rm qry}+1)) exact-integer arithmetic operations, with O(d4(bqry+1))O(d^4(b_{\rm qry}+1)) integer tape slots. Transcript compilation adds O((t+1)∣D0∣)O((t+1)|D_0|); coefficient bit growth is excluded. This is the specific bounded-lie specialization and cost accounting supplied by Theorem 3 and Appendices A.5–A.6.

The core certificate interface supports different frozen query policies. The 148-session useful-repair result uses ambiguity-block queries; the binary compiler supplies an optional exact-computation layer. Table 22 evaluates the utility of separately calibrated query pipelines and establishes no adjusted advantage over entropy. Theorem 3's result is the exact shared computation of query values, which is separate from that empirical policy comparison.

The session-budget rule is also the binary CRC special case identified in Appendix B.6. CARTER connects that calibration to complete structured repair outputs and feasible explanations at every prefix. This statistical and structural result, together with the exact transcript-to-all-query computation above, is the contribution we ask the reviewer to assess.

References: Houndji et al.; Delaite and Pesant; Zmigrod et al..

AuthorsOpenReview ↗

Response to MSwg (Q2–Q3): Frequency and subsequent repair outcomes

Q2–Q3. Frequency and subsequent repair outcomes

The saved programmed-feedback records provide direct answers: strict confident erasure occurs beyond the schematic example, and some sessions subsequently enter conflict despite having emitted no wrong edit through the erasure prefix. We obtained the counts below by a post-hoc descriptive reanalysis of existing trajectories, without generating new feedback or sessions. Genuine taxonomy changes paired with real noisy feedback are unavailable, so these are programmed-feedback frequencies.

Q2. Strict confident erasure, target loss and wrong edits are different events. We apply Definition 1: a previously covered target is removed, the surviving family is nonempty, and at least one new unanimous edge statement appears. Target loss also includes empty-family updates; wrong edits concern executable additions/deletions; conflict means an empty family.

In the canonical iid-panel/ambiguity-block study, 7/148 sessions were initially uncovered. Among the 141 initially covered sessions, hard filtering causes 32 further target losses: 18 immediately enter conflict and 14 leave a nonempty family. At budget zero, a nonempty target-loss update makes an excluded target-parent edge impossible although it was previously possible, so it introduces a new unanimous statement. We verified the frozen-graph and nested-transcript prerequisites, giving exact strict CE of 14/148 (9.5%), or 14/141 (9.9%) among initially covered sessions. The calibrated arm has 3 new target losses: one empty-family loss and 2 strict CE sessions, 2/148 (1.4%). The zero-budget inference is needed because canonical records preserve executable edit sets rather than every non-edit edge label.

E2 preserves complete edge classifications, allowing direct adjacent-prefix checks of Definition 1. Under the fixed Berlekamp query policy and iid errors, hard filtering versus calibration gives strict CE in 20/60 versus 3/60 sessions, target loss in 22/60 versus 4/60, a wrong edit in 7/60 versus 1/60, and conflict in 10/60 versus 0/60.

Under persistent-child errors, the corresponding hard-filtering versus calibrated counts are 18/60 versus 7/60 for strict CE, 18/60 versus 10/60 for target loss, 11/60 versus 5/60 for a wrong edit, and 1/60 versus 0/60 for conflict.

Each entry counts sessions; wrong edits and conflict are lifetime events. Thus the 4 iid calibrated losses include only 3 strict CE events, and the 10 persistent calibrated losses include only 7. An independent enumerator reproduced all 7,200 saved E2 prefix states, complete edge classifications and edit sets.

Q3. What happens after CE while the edit history is still error-free? We selected hard-filtered sessions whose first nonempty target loss is strict CE, with no wrong edit at or before that prefix and at least one later saved prefix. Comparisons use the same saved query/answer transcript for the hard and calibrated certificate arms.

In the canonical iid setting, two OEWN sessions lose the target at round 1 without a wrong edit and enter conflict at round 2. Their matched calibrated arms retain the target at every prefix, have no wrong edits or conflicts, and yield a mean of one final correct edit per session versus zero with hard filtering. The corresponding two persistent canonical cases have no later wrong edit or conflict in either arm; their correct-edit means reverse, 2.5 hard versus 1.5 calibrated.

In the larger E2 iid cohort of 13 silent-CE sessions, hard filtering has later wrong edits in 0/13 and later conflicts in 5/13. The matched calibrated arm retains the target at every prefix in 11/13 sessions, has lifetime wrong edits in 0/13 and conflicts in 0/13, and increases mean final correct edits from 0.615 to 1.308.

In the persistent-child cohort of 9 sessions, hard filtering has later wrong edits in 2/9 and later conflicts in 0/9. The matched calibrated arm retains the target at every prefix in 3/9, has lifetime wrong edits in 1/9 and conflicts in 0/9, and yields 0.111 mean final correct edits compared with 1.889 under hard filtering.

These outcome-conditioned, fixed-transcript comparisons are descriptive; they do not identify a population causal effect or a uniform utility gain. They do show the operational issue raised in Q3: an error-free edit history can conceal a lost target, and the saved iid cases exhibit subsequent conflict and fewer final correct certificates under hard filtering. This is why the every-prefix guarantee protects the maintained repair state as well as its emitted edits.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author Overview: Calibrated partial repair and new trajectory evidence

Thank you for recognizing the every-prefix safety guarantee and the complete decisions with inspectable witnesses. CARTER's central result is an operational contract for partial repair: maintain a calibrated family of complete trees, certify every edge statement shared by that family, and supply opposing feasible trees for each deferred choice. Under candidate inclusion and exchangeable sessions from a frozen pipeline, the marginal guarantee protects target retention and all certified edits throughout interaction.

The computational results make this contract executable. At most |D0|+1 arborescence calls recover the complete certificate and witnesses. The optional shared compiler converts a repeated-query transcript into exact disagreement-layer counts and evaluates both one-step Berlekamp outcomes for every admissible query. Our response to MSwg places these results precisely relative to prior filtering, counting and derivative methods.

The empirical results quantify useful output as well as safety. The canonical programmed-feedback study adds 2.088 correct edits per session over no feedback, with a lifetime wrong edit in 2/148 sessions. Our new post-hoc analysis of the saved trajectories finds strict confident erasure in 14/148 hard-filtered sessions versus 2/148 calibrated sessions. It also identifies cases where retaining the target preserves subsequent repair options even before any wrong edit has appeared.

The individual responses provide the complete event counts and matched follow-up outcomes, explain the principal-parent maintenance task and temporal deferral, and clarify the candidate-inclusion premise and artifact navigation. These results support evaluating CARTER as a calibrated, inspectable partial-repair interface, with concrete evidence for the output and safety tradeoff in each tested setting.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.