{"id":"2aeb3d59-634d-485c-bd0b-671a3f9b747e","arxiv_id":"2505.14510","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"BACON trains logic aggregation trees with learned feature orderings, producing compact symbolic decision rules with accuracy near black-box models on small tabular datasets.","lead":"BACON is a trainable AI model that expresses its decisions as graded logic trees, combining features with adjustable conjunction and disjunction operators instead of a black-box neural net. The paper applies it to Iris, house buying, and breast cancer data, claiming accuracy near black-box models while keeping the decision logic readable.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed automatic explainability is not demonstrated: Appendix G.3 admits the published LSP tree is manually constructed, and G.4 says it was selected post hoc by accuracy, so the trained model and the explained logic may be different artifacts.","rationale":"The reader's stated weakest assumption is the intrinsic-logic premise (§3.3.3) and the stochastic permutation search. My read agrees those are risks, but the more directly load-bearing problem is empirical and admitted in the text: the published explanation tree is not what the trained model outputs. Section G.3 explicitly says the tree was manually constructed; Section G.4 says the final explanation was chosen for accuracy. If the central contribution is 'automatically training explainable AI models' with 'full structural transparency,' then the artifact shown to the reader must be the model itself, not a post-selected human simplification. The accuracy ambiguity (98.07% in §4.1 vs 98.95% for the explanation actually selected in G.4, and no stated data split) makes the headline number unverifiable from the paper alone. This is a correctness and completeness concern, not a novelty dispute. It can be settled by code execution and automatic extraction. Because the preprint is explicitly work in progress and the code URL is provided, the appropriate outcome is to keep the CONDITIONAL verdict pending that demonstration. I do not see an internal inconsistency in the GCD formulas or a mathematical flaw in the LSP aggregation itself; the concern is about what the experiment actually establishes. Hence no change to the reader's verdict, but the condition should be sharpened: the automatic pipeline, not just the code, must be released and run.","tokens_in":15803,"tokens_out":4368,"duration_ms":51798,"concrete_test":"Run the released anonymous code on WDBC with the Table 6 hyperparameters and a fixed seed, using a defined 80/20 train/test split. After training, automatically serialize the frozen hard-permutation model into a symbolic LSP expression using only the trained weights and andness values, without manual simplification or selection. Then compare (a) the reproduced held-out accuracy with the claimed 98.07%, and (b) the automatically extracted expression with Figure 4. If the expression is not produced verbatim by the trained model, or the held-out accuracy is materially lower, the central claim must be revised to describe a human-assisted interpretable pipeline with post hoc explanation selection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (abstract, §4.1) is that BACON automatically trains a fully explainable model whose trained structure provides precise logic-based symbolic explanations. The paper's own appendices undermine this. §G.3 states: 'we chose the top five important features as highlighted in Table 7 and manually constructed the LSP aggregation tree shown in Figure 4 and applied certain simplifications.' §G.4 states: 'In the paper, we chose the final explanation because it yielded the highest accuracy.' Thus the explanation displayed in Figure 4 is not automatically produced by the trained network; it is a human-curated, accuracy-selected simplification. This breaks the end-to-end traceability criterion from §1.1: the path from raw features to output is shown for a reconstructed tree, not necessarily for the model that achieved the reported accuracy. The accuracy reporting is also ambiguous: §4.1 claims 98.07% 'out of the box,' but Table 8 lists ten explanations with accuracies from 97.36% to 98.95%, and the selected final explanation has 98.95%, not 98.07%. No train/test split is described anywhere in §4 or Appendix G, so the 98.07% may be training-set accuracy. Combined with the stochastic Gumbel-Sinkhorn search over 30! orderings (and Appendix F noting failure after 5 attempts), the reported result may reflect selection over many runs rather than a robust property of the method. The underlying LSP math and the WDBC result are plausible, but the paper as written does not demonstrate that BACON automatically yields the transparent explanation that is its headline contribution.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BACON, a two-layer architecture combining a Gumbel-Sinkhorn permutation layer with a binary tree of Graded Conjunction/Disjunction (GCD) operators from LSP theory, and claims that it trains fully explainable decision models end-to-end while achieving high accuracy. The model is evaluated on Boolean expression recovery, Iris classification, house-purchase decisions, and the WDBC breast cancer dataset, where Section 4.1 reports 98.07% accuracy and a compact, clinically aligned LSP tree. The paper argues that BACON provides structural transparency, feature attribution, compositional transparency, and human-AI collaboration beyond what post-hoc explainability methods or decision trees offer.","tokens_in":16139,"tokens_out":3627,"duration_ms":33364,"significance":"If the central claim were fully substantiated, BACON would be a meaningful contribution: it combines a differentiable graded-logic architecture with a trainable permutation search, and it produces compact symbolic trees that are prunable, human-readable, and deployable without an ML framework. The LSP/GCD formalization is coherent and the code is made available. The Boolean experiments give some evidence that the architecture can recover nested logic expressions. However, the significance is conditional because the paper's own appendices show that the headline WDBC explanation is manually constructed and accuracy-selected, and because no held-out evaluation protocol is reported. With the current evidence, the paper demonstrates the feasibility of training LSP trees, but not the full 'automatically explainable' claim made in the abstract.","major_comments":[{"comment":"The central claim that BACON automatically produces the reported symbolic explanation is not supported: G.3 states that the authors 'chose the top five important features as highlighted in Table 7 and manually constructed the LSP aggregation tree shown in Figure 4,' and G.4 states that they selected the final explanation because it yielded the highest accuracy. The Figure 4 tree is therefore a human-curated, accuracy-selected reconstruction, not the trained network's own output. Since §1.1 defines end-to-end explainability as tracing the full decision pathway of the model, the manuscript must either (a) show that the automatically trained model produces the displayed tree, or (b) clearly present the manual tree as a post-hoc explanatory artifact and re-frame the central claim. In addition, §4.1 reports 98.07% accuracy 'out of the box,' while Table 8 lists Explanation 10 with 98.95% as the chosen final explanation; these numbers describe different models, so the main text must reconcile them.","section":"§4.1 and Appendix G.3/G.4"},{"comment":"No train/test split is described for the WDBC experiment; the text reports '98.07% (95% CI: [97.85%, 98.29%])' without saying whether the CI comes from cross-validation, a held-out test set, or resampling on the training data. Because pruning (Section 4.2) and explanation selection (G.4) are performed after observing accuracy, the reported performance is at risk of selection bias. The authors should specify the evaluation protocol, report held-out (or cross-validated) accuracy for the final selected explanation, and give the accuracy of the unpruned and pruned networks on the same split.","section":"§4.1 and Appendix G"},{"comment":"The stochastic search over 30! permutations is not characterized. Appendix F states that 'in most cases the training can finish under 5 attempts' but also that 'if a network fails to converge after 5 attempts, you may want to adjust other hyper-parameters and try again.' Since Table 8 lists ten explanations with different feature sets and accuracies, and G.4 says the final explanation was selected by accuracy, the reader cannot tell whether the headline result reflects a robust property of BACON or an artifact of multiple restarts and post-hoc selection. The paper should report the number of attempts used, the distribution of results across seeds, and the selection rule.","section":"§3.2, Appendix F, Table 8"},{"comment":"The Boolean approximation and Iris experiments are reported without sufficient rigor. Appendix B says synthetic data were generated 'by repeating all possible input combinations 100 times,' but no train/test partition, no comparison baseline, and no measure of how often training succeeds are given. Table 3 reports 96.67% accuracy for Setosa and Virginica, yet no description of the split or evaluation procedure appears, and the lack of a Versicolor model is interpreted as 'real-world ambiguity' (Appendix C) without supporting evidence. These experiments need a defined protocol to substantiate the claim that BACON recovers logic from data.","section":"Appendix B and Table 3"}],"minor_comments":[{"comment":"There are several typos: the table header 'Andess' should be 'Andness,' Section 3.1 has 'resursively' for 'recursively,' and Algorithm 1 has 'pemutation' for 'permutation.'","section":"Appendix E and Algorithm 1"},{"comment":"The prompt in Appendix H refers to the 'WDSD dataset,' which should be 'WDBC.'","section":"Appendix H"},{"comment":"The code link is an anonymous URL (https://anonymous.4open.science/r/bacon-net); the final version should provide a persistent repository identifier.","section":"General"},{"comment":"Table 2 reports thresholds 0.01, 0.46, and 0.63 with accuracies 63.97%, 98.77%, and 98.42%, which differ from the 98.07% claimed in §4.1; the relationship between these numbers and the main accuracy result should be clarified.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is honest about its own limitations in the appendices; my main concern is that the main text's claims (abstract, §4.1, §4.3) overstate what the appendices demonstrate. The revision should either restrict the claims or add the missing automatic-extraction and held-out evaluation evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"BACON is a real attempt to close the loop between trainable neural models and symbolic graded logic, and the core idea—using Gumbel-Sinkhorn to search feature permutations while learning LSP aggregation parameters—is genuinely new. On small synthetic and semi-synthetic tasks (Boolean expressions, house-buying rules) it recovers the underlying logic, which is a meaningful proof of concept. The appendices are refreshingly honest: they admit the permutation search can fail, that the breast cancer tree in Figure 4 was manually constructed (G.3), and that the presented explanation was chosen because it had the highest accuracy (G.4). That honesty is a point in the authors' favor.\n\nThe problem is that the honesty undercuts the paper's central claim. The abstract promises 'full structural transparency' from an automatically trained model, but the published WDBC explanation is a human-curated simplification of one of many runs. The accuracy numbers also don't line up: §4.1 says 98.07% 'out of the box,' while Table 8 shows the selected explanation at 98.95% and several other runs at 97-98%. No train/test split is described for WDBC, so it's impossible to tell whether the reported accuracy is on training data or held-out data. And with a stochastic search over 30! orderings, the risk of selection over many runs is real.\n\nThe LSP mathematics itself is solid and the clinical alignment of the tree is plausible, so I don't doubt the model can work. But the 'fully explainable' framing needs serious recalibration. Either the authors need to ship an automatic tree-extraction pipeline, or they need to present BACON as a human-in-the-loop tool where the expert constructs the final logic from a set of learned candidates. The current draft cannot support the stronger claim.\n\nThis deserves peer review, because the architecture is novel and the negative results (Iris Versicolor, training failures) are reported without spin. The referee should focus the authors on reproducibility: fixed random seeds, a proper validation protocol, and a clear separation between the trained network and the curated explanation. If those are addressed, this could be a solid XAI contribution.","headline":"Novel and honest, but the central 'fully automatic explainability' claim is undercut by manual tree construction and post-hoc selection; worthwhile for a serious referee after major revision.","tokens_in":16671,"tokens_out":2068,"would_cite":false,"duration_ms":20092,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"BACON trains a graded-logic aggregation tree that is transparent by construction, and on the WDBC breast cancer dataset it reaches 98.07% accuracy using all 30 features with no feature selection or fine-tuning.","keywords":["explainable AI","graded logic","LSP aggregation tree","logic scoring of preference","Gumbel-Sinkhorn permutation","feature attribution","breast cancer diagnosis","human-AI collaboration"],"falsifier":"Run BACON with many random seeds on a dataset whose optimal decision boundary is known to be non-logical (for example, a high-dimensional random-label set or a smooth nonlinear function without conjunctive/disjunctive structure) and check whether training consistently fails or produces unstable, contradictory trees. A sharper version: on the WDBC benchmark, compare repeated runs with different seeds to see whether the reported 98.07% accuracy and the 7-feature explanation are reproducible or whether the stochastic permutation search selects different, equally accurate trees.","tokens_in":15582,"feed_emoji":"🌳","tokens_out":8268,"duration_ms":70919,"temperature":0.7,"pith_summary":"BACON sets out to show that an automatically trained model can be fully explainable without giving up accuracy. Its central claim is that by replacing a neuron's weighted sum with a graded-logic aggregation operator, a binary tree can be learned that expresses the decision as an LSP graded-logic formula—an explicit conjunction/disjunction tree with per-input importance weights and andness values. On the WDBC breast cancer benchmark, the paper reports 98.07% accuracy using all 30 features with no feature selection or problem-specific tuning, while pruning leaves only 7 clinically meaningful features. If this holds, high-stakes domains like medical diagnosis could get black-box-level performance together with decision logic that a human expert can read, audit, and adjust.","feed_headline":"Logic-tree AI hits 98% accuracy while staying fully transparent","feed_subtitle":"BACON learns a human-readable graded-logic decision tree from raw features, no feature selection or fine-tuning needed.","key_machinery":"The load-bearing object is the andness-directed graded conjunction/disjunction operator $\\mathrm{GCD}_2(x,y;w,\\alpha)$, a two-input logic aggregation that interpolates continuously from pure disjunction ($\\alpha=0$) through the arithmetic mean ($\\alpha=0.5$) to pure conjunction ($\\alpha=1$), with hyper-conjunctive and hyper-disjunctive extremes. Composed in a left-associative binary tree, these operators form an LSP (Logic Scoring of Preference) aggregation tree that gives every internal node a semantic label such as 'must have all' or 'nice to have some'; the Gumbel-Sinkhorn layer supplies the search mechanism over input orderings. The same tree is simultaneously the model's explanation: its structure shows which features are mandatory, desired, optional, or sufficient, and its output is a closed-form arithmetic expression that can be evaluated without any ML framework.","core_discovery":"On its own terms, the paper's discovery is a training procedure that converges on an interpretable graded-logic aggregation structure rather than a statistical black box. A BACON network is a two-layer architecture: a Gumbel-Sinkhorn permutation layer that searches the $n!$ input orderings, followed by a left-associative binary tree whose internal nodes are GCD$_2$ aggregators $\\mathrm{GCD}_2(x,y;w,\\alpha)$, where $w$ sets relative importance and $\\alpha$ sets the degree of conjunction versus disjunction. Training alternates stochastic permutation exploration with a Hungarian-algorithm freeze of the best ordering, then fine-tunes weights and andness values. The trained tree is prunable and can be read as a symbolic LSP expression; in the breast cancer study it yields 98.07% out-of-the-box accuracy, identifies 7 of 30 features as decisive, and expresses malignancy as a strict conjunction between a large-cell group and a high-irregularity feature. The paper takes this as evidence that data-driven decisions can be made end-to-end transparent while staying competitive with optimized ensembles.","pith_inferences":["A testable extension: use non-convergence as a diagnostic signal for whether a dataset admits a transparent logical explanation at all; the paper's own assumption implies this, but the idea is not developed there.","A structural extension the paper leaves open: moving from the left-associative tree to a balanced or hybrid layout could let the model express group interactions and equitable feature participation that the current sequential tree cannot represent.","Because correlated features produce several near-equally accurate explanations in the breast cancer study, an editor's reading is that a single chosen tree may understate the ambiguity in the data; a small set of diverse trees would give clinicians a more honest picture.","The Boolean experiments suggest OR-dominated and deeply nested expressions are harder to learn; a natural stress test would quantify how learning time and convergence rate scale with the fraction of disjunctions in the target logic."],"forward_implications":["A pruned BACON tree can be reduced to a standalone symbolic expression, so deployment on edge devices or real-time controllers needs no neural-network runtime.","Feature importance comes from the tree structure itself: pruning shows which inputs actually change accuracy, giving a decision-centric ranking rather than a post-hoc approximation.","Changing one threshold maps the continuous truth value to a binary decision, letting a clinician favor recall during screening or precision during treatment planning on the same trained model.","Because the model's output is a degree of truth, BACON can be translated into natural-language reports, making the logic accessible to non-specialists.","On the benchmark, choosing all 30 features and doing no preprocessing other than normalization matches the practical accuracy of optimized black-box ensembles, while keeping the decision path visible."],"supporting_citations":[{"why":"Supplies the LSP graded-logic aggregation method and the andness-directed GCD family that BACON's tree is built from.","marker":"[Dujmović, 2018]"},{"why":"Supplies the Gumbel-Sinkhorn network used to search over input permutations during training.","marker":"[Mena et al., 2018]"},{"why":"Supplies the Hungarian algorithm that converts the soft permutation into a frozen hard assignment.","marker":"[Kuhn, 1955]"},{"why":"Provides the WDBC breast cancer dataset used in the main evaluation.","marker":"[Wolberg et al., 1993]"},{"why":"Provides the 99.89% ensemble baseline that BACON compares itself against in the breast cancer study.","marker":"[Reshan et al., 2023]"},{"why":"Supports the clinical plausibility of the selected features (large size and irregular shape).","marker":"[Yu et al., 2023]"},{"why":"Provides the andness-directed GCD table used to verbalize each aggregator in human-readable terms.","marker":"[Dujmović, 2025]"}],"fun_headline_variants":["BACON: Explainable AI hits 98% with no black box","Graded-logic AI promises full transparency at 98% accuracy","BACON: Human-readable logic trees, no feature engineering needed","Meet BACON: AI that explains every decision in plain logic","Fully transparent AI: BACON learns decision logic from raw data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise, stated by the paper, is that the data contains an intrinsic logical aggregation structure that a human can read; if no such structure exists, the model will not converge and no amount of training will produce a meaningful explanation.","fun_headline_variants_meta":{"raw":{"variants":["BACON: Explainable AI hits 98% with no black box","Graded-logic AI promises full transparency at 98% accuracy","BACON: Human-readable logic trees, no feature engineering needed","Meet BACON: AI that explains every decision in plain logic","Fully transparent AI: BACON learns decision logic from raw data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000387,"raw_usage":{"total_tokens":2038,"prompt_tokens":937,"completion_tokens":1101,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":1011}},"tokens_in":553,"tokens_out":1101,"duration_ms":10216,"temperature":1.0,"reasoning_tokens":1011,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:32:12.515569+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run BACON with many random seeds on a dataset whose optimal decision boundary is known to be non-logical (for example, a high-dimensional random-label set or a smooth nonlinear function without conjunctive/disjunctive structure) and check whether training consistently fails or produces unstable, contradictory trees. A sharper version: on the WDBC benchmark, compare repeated runs with different seeds to see whether the reported 98.07% accuracy and the 7-feature explanation are reproducible or whether the stochastic permutation search selects different, equally accurate trees.","supporting_citations":[{"cited_title":"Enhancing breast cancer detection and classification using advanced multi-model features and ensemble machine learning techniques","cited_arxiv_id":null,"evidence_quote":"Provides the 99.89% ensemble baseline that BACON compares itself against in the breast cancer study."}],"review_version":1}