{"id":"ed65719e-9b62-42fb-95a1-c1f72c559f7c","arxiv_id":"2501.03026","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper argues that Putnam's critical (prediction) and explanatory (explanation) tendencies in science necessarily depend on each other, and interprets this through the input-output structure of machine learning models.","lead":"This philosophy-of-science paper argues that Putnam's two scientific tendencies, prediction and explanation, necessarily depend on each other, and it maps that interdependence onto the input-output structure of machine learning models. A generalist might read it as a concrete attempt to connect philosophy of science with deep learning practice, though the author concedes the argument mostly reworks Putnam.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The necessity claim collapses under the paper's own 'successfully' qualifier: P1/P2 only license conditional interdependence, and the conclusion concedes that representation without the other tendency is possible.","rationale":"The reader's weakest_assumption identifies the reduction of epistemic values to predictive and explanatory power, which is indeed a genuine unsupported premise in the paper's support for P1. However, the more fundamental problem is the qualifier 'successfully' in both premises. Even if the epistemic-values reduction were granted, the argument still would not establish necessity, because a representation could be less successful yet still possible. The author's own conclusion concedes this possibility, so the strongest claim as stated is internally inconsistent with the author's caveat. My concern is therefore a distinct, more formal objection, though it overlaps with the reader's point: the reduction is one attempt to make 'must' do real work, and it fails. The reader's verdict of CONDITIONAL remains appropriate: the paper has a clear revision path, namely to state the qualified thesis as 'under successful representation, interdependence is required' and to either prove or drop the unargued reduction of epistemic values. I do not see the concern as requiring a full rejection, because the paper is openly exploratory and its own caveats point to the needed repair. Agreement is partial because the reader noticed the 'successfully' issue in the rationale but selected a different weakest assumption as the primary load-bearing point.","tokens_in":5295,"tokens_out":5281,"duration_ms":57271,"concrete_test":"Formalize the argument with explicit predicates: let S1(p) mean 'p is successfully represented by schema I', E(p) mean 'the auxiliary statements of p have explanatory power', and S2(p) mean 'p is successfully represented by schema II'. P1 is ∀p(S1(p)→E(p)) and P2 is ∀p(S2(p)→P(p)) for an analogous predictive-power predicate. C requires a necessity claim about the tendencies, but no premise of the form ∀p(E(p)↔P(p)) or a definition tying 'successful' to possession of both powers is supplied.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central conclusion, C, asserts that the critical and explanatory tendencies are necessarily interdependent. But both premises are conditional on a scientific problem being 'successfully represented' by one of the schemata. The inference from these conditionals to a necessity claim is valid only if 'successfully' is stipulated to include the presence of the other tendency; otherwise the argument shows at most that interdependence is possible or valuable, not necessary. The paper's own concluding admission makes this explicit: 'It can be put forth that it is possible to represent a scientific problem without the other tendency; however, I suggest that it is the worse for it.' That sentence directly contradicts C, since a merely worse representation is still a possible representation. The examples offered—the Earth-orbit prediction, the Uranus/Neptune auxiliary, the Copernican shift—establish that interdependence occurs and can be explanatory, but they do not establish that it is required. The attempt to support P1 by reducing all epistemic values to predictive or explanatory power is similarly asserted rather than demonstrated, and the author explicitly sets aside simplicity and scope as not yet approachable. The ML analogy does not repair the gap: the claim that a model needs an input is a computational-dependence claim, not a demonstration that the input has explanatory power. Thus the most load-bearing weakness is not any empirical detail but the modal structure of the argument itself.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reconstructs Hilary Putnam's distinction between the critical tendency (schema I: theory plus auxiliary statements yielding a prediction to be judged true/false) and the explanatory tendency (schema II: theory plus auxiliary statements explaining a fact). It then argues for a biconditional necessity claim (C): the two tendencies are necessarily interdependent. The argument proceeds through two premises, P1 (successful schema I representations require auxiliary statements with explanatory power) and P2 (successful schema II representations require conjoining theory with prediction), and is then illustrated through an analogy with machine learning models, where model parameters, inputs, and outputs are mapped onto the schemata. The conclusion reflects on machine learning as a 'wrench' in theory-choice debates and suggests the possible discovery of 'model-independent objective parameters.' The paper is candid about its limitations, including an explicit admission that the term 'successfully' may be a weasel word and that representation without the other tendency may still be possible.","tokens_in":5456,"tokens_out":3045,"duration_ms":28489,"significance":"If the necessity claim C were established, the paper would offer a substantive philosophical thesis: every predictive act in science carries explanatory presuppositions, and every explanation depends on predictive auxiliaries. That would give a principled unity to two tendencies usually treated as complementary but separable. The paper has notable strengths: it accurately reconstructs Putnam's schemata and examples (Earth orbit, Uranus/Neptune), uses standard literature (Kuhn, McMullin, Putnam), and introduces an original and thought-provoking machine learning analogy. It also explicitly flags its own weakest points, which is a commendable scholarly practice. However, the central modal inference is not demonstrated; the paper shows at most that interdependence is valuable and frequently present, not that it is necessary. The epistemic-values reduction that would support P1 is asserted rather than argued. The machine learning analogy, while suggestive, does not close the modal gap because it establishes a computational dependence on input rather than an explanatory power carried by that input.","major_comments":[{"comment":"The inference from the conditional premises to the necessity claim C is not valid. Both P1 and P2 are conditional on a scientific problem being 'successfully represented' by one of the schemata, but C asserts an unconditional necessity: that representation in either tendency is dependent on the other. The author's own concluding sentence concedes the gap: 'It can be put forth that it is possible to represent a scientific problem without the other tendency; however, I suggest that it is the worse for it.' That sentence directly contradicts C, since a representation that is merely worse is still possible. To repair this, the author must either define 'successfully' so that it already includes the presence of the other tendency (which would make the argument circular) or weaken C to a claim of valuable or typical interdependence.","section":"Introduction, argument block (P1, P2, C) and Conclusion"},{"comment":"The load-bearing premise that all epistemic values reduce to predictive or explanatory power is asserted rather than demonstrated. The author explicitly says that reducing scope or simplicity 'can't be approached in the same manner, but I believe it is possible.' Since P1's 'must' depends on this reduction (auxiliary statements must have explanatory power because only predictive and explanatory power matter), the premise is left unsupported at exactly the point where the necessity claim is supposed to enter. The paper needs at least a sketch of a reduction argument for simplicity and scope, or an alternative argument for why auxiliary statements are indispensable for successful schema I representation.","section":"Support for P1, third paragraph"},{"comment":"Putnam's Uranus/Neptune example, as reconstructed here, demonstrates that a schema II problem can be resolved by introducing a schema I-type auxiliary statement (S3), but it does not demonstrate that such a resolution is the only possible one or that every successful schema II representation must contain a predictive auxiliary. The later claim that 'the predictive power must be there if the representation is to be used by the scientific community' is a sociological or pragmatic assertion about use conditions, not an argument that a schema II representation without predictive power is impossible. The author also acknowledges that 'it is debatable whether a prediction must be based on observation,' which further weakens the necessity claim. P2 therefore remains a claim about a common and valuable pattern, not an established necessity.","section":"Support for P2, Uranus example"},{"comment":"The machine learning analogy does not repair the modal gap in the argument. The claim that a model cannot produce an output without an input establishes a computational dependence, but it does not establish that the input has explanatory power in Putnam's sense; an input vector or a prompt is not thereby an 'auxiliary statement' that explains where the theory is applied. Similarly, the explainability schema (pp. 4-5) shows that predictions can be used to probe model features, but that is a methodological point about model interpretation, not a demonstration that the critical and explanatory tendencies are necessarily interdependent. The analogy is suggestive, but it trades on an equivocation between 'input as a necessary condition for computation' and 'auxiliary statements as providing explanatory meaning.'","section":"Machine learning schematizations (pp. 3-5)"}],"minor_comments":[{"comment":"There is a typo in 'Khun's fruitfulness'; it should be 'Kuhn's fruitfulness.'","section":"Support for P1, p. 3"},{"comment":"The diagram heading reads 'Molel (Parameters)'; it should be 'Model (Parameters).'","section":"Machine learning schema, p. 5"},{"comment":"The in-text citation 'Curd & Clover, 1998' does not match the reference list entry 'Curd, Martin, J. A. Cover, and Chris Pincock'; the surname should be 'Cover', not 'Clover', and the year in the text should be reconciled with the reference entry.","section":"References and citations"},{"comment":"The final reference for the Explainability survey ends with a duplicated fragment '/10.48550/arXiv.2309.01029.' that appears to be a leftover from the previous line and should be deleted.","section":"References"},{"comment":"McMullin is cited in the text (in 'Rationality and Paradigm Change') but does not appear in the reference list; a full bibliographic entry should be added.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper reads more like a thoughtful essay or a working paper than a fully developed research article. The core issue is that the author has explicitly identified the gap in the argument (the 'successfully' qualifier) but has not resolved it; the conclusion concedes the very possibility claim that C denies. Since the author is aware of the problem, a revision that either weakens C to a claim of valuable or typical interdependence or supplies a genuine argument for the reduction of all epistemic values would be feasible within the paper's scope. The machine learning material is interesting and could be a distinctive contribution if it is used to sharpen, rather than merely illustrate, the philosophical claim. I would not reject the paper, but I would not accept it in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Net take: this is an honest, well-written philosophy essay that reconstructs Putnam's critical/explanatory schemata and maps them onto machine learning. The reconstruction is mostly accurate, and the author deserves credit for flagging the two weakest points himself. But the central thesis—biconditional necessity—is not established. The premises are conditional on 'successfully represented,' so they license at most 'if a problem is represented in either schema, it relies on the other tendency for success.' That is weaker than necessity. The author's own conclusion admits it: he says a problem can be represented without the other tendency, but it is worse off. That directly contradicts C. The examples (Earth orbit, Uranus/Neptune, Copernican shift) show interdependence is possible and sometimes explanatory, not required.\n\nThe paper's other soft spot is the reduction of all epistemic values to predictive/explanatory power, asserted to support P1. The author sets simplicity and scope aside with 'I believe it is possible,' which is not an argument. If that reduction fails, P1's 'must' loses its support. The ML section is suggestive but not empirical; the input-output analogy shows a model needs an input, not that the input has explanatory power. And the closing speculation about 'model-independent objective parameters' has no mechanism. Those are speculative, and the author presents them as such.\n\nWhat is genuinely useful: the paper gives a clear, fair exposition of Putnam's two schemata and the Uranus example, and it asks a real question—whether ML models change how we think about theory-ladenness. The parallels drawn are worth a short discussion paper, not a full argument for necessity. If the thesis were weakened to 'interdependence under successful representation' and the epistemic-value reduction dropped or defended, the essay could be a solid contribution to the philosophy-of-ML literature. As written, it is a promising draft.\n\nI'd read this carefully if I were refereeing a philosophy-of-science journal, and the internal candor makes it a good candidate for major revision. Serious referee: yes. I wouldn't cite it yet, but I'd bring it to a reading group for the discussion.","headline":"A readable, honest reconstruction of Putnam that overstates its case: the necessity claim doesn't follow from the premises, and the author's own concessions confirm it.","tokens_in":6089,"tokens_out":1964,"would_cite":false,"duration_ms":18493,"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":"This paper argues that Putnam's critical and explanatory tendencies are necessarily interdependent, and that deep learning models illustrate the interdependence in practice.","keywords":["Hilary Putnam","critical tendency","explanatory tendency","theory choice","scientific explanation","machine learning interpretability","deep learning","philosophy of science"],"falsifier":"Find one successful scientific prediction whose auxiliary statements contribute no explanatory content at all—for example, a purely conventional coordinate choice that makes a theory computable but that no one would claim explains anything—and the paper's claim that explanatory power is necessary for schema I collapses.","tokens_in":4964,"feed_emoji":"🔄","tokens_out":11611,"duration_ms":107694,"temperature":0.7,"pith_summary":"The paper argues that two ways of representing scientific problems that Hilary Putnam distinguished—the critical tendency, where theory plus auxiliary statements yields a prediction to be tested, and the explanatory tendency, where theory plus missing auxiliary statements yields an explanation of a fact—cannot function apart. Its central claim is a biconditional: any successful critical representation already depends on explanatory content in its auxiliary statements, and any successful explanatory representation already depends on predictive content. The author uses Putnam's own examples (predicting Earth's orbit; explaining Uranus's orbit) to show the two schemata interlock, then maps them onto deep learning: a model plus input produces a prediction, and explaining a model's trained features requires successful input–output predictions. If the argument is right, prediction and explanation are two aspects of the same underlying scientific act, and machine learning gives a concrete place to watch that interdependence.","feed_headline":"Prediction and explanation are necessarily interdependent","feed_subtitle":"Machine learning models illustrate why scientific prediction and explanation are inseparable.","key_machinery":"The central machinery is Putnam's paired schemata as modified by the author's biconditional reading. Schema I reads THEORY + AUXILIARY STATEMENTS → PREDICTION (true or false?), and schema II reads THEORY + ?????? → FACT TO BE EXPLAINED. The argument's engine is the identification of the missing clause in each schema with content of the other: the auxiliary statements in schema I are said to supply explanatory power, and the missing auxiliary in schema II is said to be a predictive statement of schema I form, as in the Uranus example. The author also introduces a machine-learning analogue: a trained model plus an input produces an output/prediction (schema I), and extracting meaning from learned parameters via successful input–output tests forms a schema II-like explanatory loop. This analogue lets the author claim that the interdependence is not just a philosophical reconstruction but observable in working AI systems.","core_discovery":"The paper's central claim is that the critical tendency and the explanatory tendency are necessarily interdependent, so that a scientific problem represented by one schema is always already entangled with the other. For the critical tendency this means the auxiliary statements that make a prediction possible must also carry explanatory power—otherwise the prediction is not meaningful—and for the explanatory tendency it means the missing auxiliary statements used to explain a fact must themselves have predictive power, often taking the form of a lower-level schema I prediction. The author supports the first half by extending Putnam's observation that a theory never predicts alone: just as a theory needs auxiliaries to yield a prediction, it needs auxiliaries with explanatory content to make the prediction meaningful, illustrated by the Copernican case where two theories had equal predictive power but different explanatory reach. The author supports the second half through Putnam's Uranus example, where the auxiliary that completes the explanation is itself a schema I prediction about the existence of a planet. The paper then reads deep learning models through the same two schemata, arguing that model predictions depend on contextualizing inputs and that model explainability depends on predictive success, making machine learning a live instance of the interdependence.","pith_inferences":["A testable extension the paper does not pursue: if the biconditional holds, removing the contextual prompt from an LLM should degrade prediction quality in a way that tracks the loss of explanatory auxiliary content, not just loss of information.","The author's reduction of epistemic values could be made quantitative by measuring explanatory and predictive power per parameter in trained models, turning simplicity and scope into efficiency measures rather than separate values.","The parameter-ladenness idea suggests a sharper research question than the paper states: which scientific explanations would change if a foundation model's internal representation, rather than an explicit law, became the standard auxiliary statement?","If the interdependence is real, purely rationalizing AI explanations that make no predictive difference should be judged as failed science, a criterion the paper gestures at but does not state."],"forward_implications":["If every successful prediction already rests on explanatory auxiliaries, debates about theory choice cannot reduce to predictive accuracy alone; explanatory value is already inside the predictive act.","If every successful explanation already relies on a predictive auxiliary, then explanations that float free of any possible prediction are not usable by a scientific community.","Machine learning use, schematised as model plus input to output, inherits the same structure: an LLM's output is not a prediction from the model alone but from model plus prompt, so the prompt functions as an explanatory auxiliary.","Model explainability, interpreted as a schema II problem, depends on predictive success: we explain model features by testing what input–output pairs they make true, so the two tendencies interlock inside AI systems as well.","If machine learning models continue to improve as the paper assumes, future extraordinary science may be driven by new parameter structures rather than by new observations, a possibility the author calls the 'parameter-ladenness of theories'."],"supporting_citations":[{"why":"Supplies the two schemata (critical and explanatory tendencies) and the orbit examples (Earth and Uranus) that the paper reconstructs and builds its biconditional argument on.","marker":"Putnam, 1979"},{"why":"Provides McMullin's 'Rationality and Paradigm Change', which contributes the Copernican case where equal predictive power is distinguished by explanatory power, a key support for P1.","marker":"Curd & Clover, 1998"},{"why":"Cited as the seminal transformer paper establishing rapid machine learning capability growth, used to justify the analogy between theories and machine-learning models.","marker":"'Attention is All You Need' (2017)"},{"why":"Supplies empirical evidence that AI-generated output is indistinguishable from human work and often preferred, supporting the claim that models produce meaningful predictions.","marker":"Porter and Machery, 2024"},{"why":"A foundation model trained on physical sensor data that predicts phenomena without explicit laws, used to argue machine learning may eventually surpass theory-based prediction.","marker":"Archetype AI (2024)"}],"fun_headline_variants":["Prediction and explanation: necessarily interdependent","ML shows prediction and explanation are inseparable","Putnam's tendencies: prediction and explanation must unite","Science's prediction-explanation link, confirmed by ML","Why prediction and explanation cannot be pulled apart"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that every scientific virtue can be boiled down to either predictive power or explanatory power; if that reduction fails, the claimed necessity in each schema loses its support.","fun_headline_variants_meta":{"raw":{"variants":["Prediction and explanation: necessarily interdependent","ML shows prediction and explanation are inseparable","Putnam's tendencies: prediction and explanation must unite","Science's prediction-explanation link, confirmed by ML","Why prediction and explanation cannot be pulled apart"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000668,"raw_usage":{"total_tokens":2989,"prompt_tokens":827,"completion_tokens":2162,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":443,"completion_tokens_details":{"reasoning_tokens":2094}},"tokens_in":443,"tokens_out":2162,"duration_ms":16539,"temperature":1.0,"reasoning_tokens":2094,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:58:32.963242+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Find one successful scientific prediction whose auxiliary statements contribute no explanatory content at all—for example, a purely conventional coordinate choice that makes a theory computable but that no one would claim explains anything—and the paper's claim that explanatory power is necessary for schema I collapses.","supporting_citations":[],"review_version":1}