{"id":"882840b9-4d5f-4d92-a07d-64adbddb2476","arxiv_id":"2606.04779","paper_version":2,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":2,"one_line_summary":"A tree-based formalization proves that reliance-based HAIs cannot be complementary under a pointwise-min benchmark, reduces regression complementarity to distance minimization, and shows classification is obstructed for internal aggregation rules.","lead":"This paper builds a tree-based mathematical framework for human-AI complementarity, in which a team is an ordered binary workflow of prediction vectors. It proves that selection-style interactions cannot beat the best per-case prediction, and that in regression complementarity reduces to getting closer to the truth, while in classification it is impossible under natural aggregation rules.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Impossibility results hinge on the pointwise-min benchmark; under the standard aggregate benchmark, internal classification rules can achieve complementarity, so the 'obstructed in classification' claim is benchmark-conditional.","rationale":"The reader's weakest_assumption correctly identifies the pointwise-min benchmark as the fragile premise. I agree: the core theorems are correct as stated, but the central claim's significance for HAI depends on the normative benchmark choice. The paper is transparent in §3.2.3, which is why I do not recommend rejection; the math checks out. However, the abstract and conclusion present the classification impossibility as a general obstruction, omitting the benchmark qualifier. This can mislead readers who use the standard aggregate complementarity definition (e.g., Donahue et al. 2022). A concrete counterexample with Brier loss and internal averaging demonstrates that C>0 while Ψ<0, so Theorem 4 cannot be extended to the aggregate benchmark. Therefore, the paper should be accepted only with the condition that all impossibility claims are explicitly and consistently stated as relative to the pointwise-min oracle benchmark, both in the abstract and in the final summary. This is a framing/correctness-of-interpretation issue, not an algebraic error.","tokens_in":31289,"tokens_out":7619,"duration_ms":69560,"concrete_test":"Reproduce the two-agent binary classification example with Brier loss: y=(1,0), H=(0.9,0.1), A=(0.1,0.9), internal rule m(u,v)=(u+v)/2. Compute Ψ per Eq. (6)–(7) and C per Eq. (8). The computation yields Ψ=−0.24 and C=+0.16, showing that Theorem 4's obstruction vanishes under the aggregate benchmark. If a reviewer or reader independently verifies these numbers, the concern lands: the paper's unqualified 'obstructed in classification' claim must be restated as 'obstructed relative to the pointwise-min benchmark.'","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central negative results—Theorem 1 and Theorem 4—are proved for Ψ with the pointwise-min benchmark Φ of Eq. (6), not for the aggregate benchmark C of Eq. (8). The paper itself notes in §3.2.3 that C can report complementarity while Ψ fails. This is not a minor technicality: the difference is consequential even for internal rules and endpoint-monotone losses. Concretely, with Brier loss, y=(1,0), H=(0.9,0.1), A=(0.1,0.9), and the internal rule m(u,v)=(u+v)/2, the protocol output is (0.5,0.5). The pointwise-min benchmark gives Ψ=0.01−0.25=−0.24, so Theorem 4 applies and complementarity is 'obstructed.' But the aggregate benchmark gives L_H=L_A=0.41 and L_T=0.25, so C=0.16>0: the same internal rule, the same endpoint-monotone loss, and the same dataset achieve positive complementarity under the standard CTP-style benchmark. Thus the impossibility is an artifact of choosing the pointwise-min oracle; it does not hold for the aggregate benchmark that defines complementarity in most HAI literature. The abstract and conclusion state 'complementarity is obstructed in classification' without this qualifier, which overstates the scope of the theorem. The paper does disclose the benchmark dependence in §3.2.3, so the mathematics is internally sound; the load-bearing concern is that the headline claim's practical import is contingent on a normative benchmark choice that many HAI researchers may not accept.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a tree-based formal framework for studying complementarity in multi-agent human–AI interactions. An HAI protocol is represented by an ordered configuration of agents and a rooted planar binary tree whose leaves carry prediction vectors; a local binary composition rule is evaluated recursively to produce the protocol output. Complementarity is defined via the pointwise-min oracle benchmark (Eq. 6), and the main results are: (1) selector rules cannot be complementary (Theorem 1); (2) under squared-loss regression, maximizing complementarity is equivalent to Euclidean distance minimization, with a closed-form optimal pooling weight for N=2 (Props. 2 and 4); (3) under linear pooling, protocol trees induce barycentric coordinate charts and Tamari-cover reparameterizations preserve complementarity, with a pentagon identity for N=4 (Prop. 6, Thms. 2 and 3); (4) in binary classification with endpoint-monotone losses, no internal local rule can achieve positive complementarity (Theorem 4), with an analogous multiclass obstruction under cross-entropy. Numerical illustrations accompany the regression and classification results.","tokens_in":31641,"tokens_out":8690,"duration_ms":89133,"significance":"If the results are correct, the paper makes a substantive formal contribution: it moves complementarity from two-agent loss comparisons to a setting with explicit workflow topology, and it gives clean geometric and algebraic characterizations. The closed-form N=2 regression optimizer and the classification impossibility theorem are crisp and falsifiable. The paper is also careful in presenting proofs in appendices, and the key computations I checked (Theorem 2's parameter transport, Proposition 5's coefficient identities, and Theorem 4's endpoint-monotone argument) are sound. The main caveat, already acknowledged in §3.2.3, is that all impossibility results are relative to the pointwise-min benchmark, not the aggregate benchmark C in Eq. (8); this matters for interpreting the practical scope of Theorem 4. Overall, the framework is a valuable addition to the formal HAI literature.","major_comments":[{"comment":"The headline claim that 'complementarity is obstructed in classification' is stated in the abstract and in the fourth discussion message without the explicit qualifier 'relative to the pointwise-min benchmark'. Section 3.2.3 correctly shows that the aggregate benchmark C (Eq. 8) can report positive complementarity for the same internal rule, dataset, and endpoint-monotone loss, even when Ψ is negative. Since Theorem 4 is a statement about Ψ, the unqualified sentences overstate the theorem's scope. I recommend adding the benchmark qualifier wherever the impossibility is advertised outside the formal theorem statements.","section":"Abstract and §9"}],"minor_comments":[{"comment":"The notation 'CmT_{N,T}' in Eq. (8) is slightly confusing because the subscript already contains T; consider using C^{m_T}_{N,T} or a clearer symbol such as C_N^{m_T}.","section":"§3.2.3"},{"comment":"The sentence 'selectors cannot yield complementarity' in the text preceding Theorem 1 is accurate only for Ψ. Since the aggregate benchmark C can be positive for a casewise selector, please add a forward reference to the benchmark discussion in §3.2.3 at this point.","section":"§4.2"},{"comment":"The bibliography lists 'Hemmer et al.' with inconsistent first-author naming: Patrick Hemmer et al. (2021) and Philipp Hemmer et al. (2025). Please standardize.","section":"References"},{"comment":"The caption says 'blue points satisfy nΨ>0' while the color scale is not fully accessible in black-and-white print. Consider adding a symbol or grey-scale pattern to distinguish the two classes.","section":"Fig. 9"},{"comment":"The numerical experiments are explicitly synthetic illustrations, but they lack error bars or repeated-seed sensitivity analysis. A short sentence acknowledging that the plots are deterministic illustrations of the theorems would be sufficient.","section":"Appendix E/F"}],"recommendation":"minor_revision","confidential_remarks":"The paper is mathematically sound as far as I checked. The main risk is interpretive: the negative classification result depends on the pointwise-min benchmark, which the authors disclose but which is not the standard aggregate benchmark in much HAI work. This is a modeling choice rather than an error, and the body of the paper handles it honestly. The requested revision is mostly a matter of carrying the same qualifier into the abstract and conclusion. I would not block acceptance on the benchmark issue, provided the language is tightened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth reading. Ferrario gives HAI theory something it lacked: a precise, tree-based language for multi-agent protocols, plus several real results. The regression story is the strongest part—complementarity as Euclidean distance minimization (Prop. 2), the closed-form N=2 optimizer with the residual-correction interpretation (Prop. 4), and the Tamari reparameterization/pentagon identity are clean and correct. The selector impossibility (Thm. 1) and the endpoint-monotone classification obstruction (Thm. 4) are also proved properly; I spot-checked the key algebra and it holds. The paper is honest enough to disclose the benchmark choice in Section 3.2.3. That is where the trouble starts. The impossibility theorems are all relative to the pointwise-min benchmark, not the aggregate benchmark used in most of the HAI literature. And this is not a minor technicality. Under the aggregate benchmark, selectors can be complementary (pick the better agent per case), and internal rules can be complementary too. Simple example: y=(1,0), H=(0.9,0.9), A=(0.1,0.1), m=(u+v)/2. Then L_H=L_A=0.41, L_T=0.25, so C=0.16>0, while the pointwise-min benchmark gives Φ=0.01 and Ψ=-0.24<0. Same internal rule, same endpoint-monotone loss, same dataset. So the headline 'complementarity is obstructed in classification' is only true for one benchmark, and the abstract/conclusion state it without the qualifier. The stress-test note gets this right in substance, though its concrete numbers are off (it claims L_H=L_A=0.41 for H=(0.9,0.1), A=(0.1,0.9), which are not right). The fix is easy: qualify all negative claims as relative to the pointwise-min oracle, and move the benchmark discussion to the front of the paper rather than burying it halfway through. Minor issues: no machine-checked proofs (though the proofs are short and checkable), and the numerical illustrations carry no error bars but are explicitly illustrative, so I am not worried. The distinction between the two benchmarks is genuinely important for HAI practice, and the paper's framework is the right tool to have that conversation in. This deserves a serious referee and, with the benchmark qualifier added to the main claims, acceptance. For me, it is a definite cite in the next twelve months.","headline":"A genuinely useful formal framework for multi-agent complementarity, with solid math and a real framing problem: the negative results only hold under the pointwise-min benchmark, and the abstract overstates them.","tokens_in":32102,"tokens_out":4219,"would_cite":true,"duration_ms":40997,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A tree-based theory shows complementarity is achievable in regression but blocked in binary classification.","keywords":["complementarity","human-AI interaction","multi-agent protocols","planar binary trees","protocol trees","pointwise-min benchmark","Tamari lattice","classification impossibility"],"falsifier":"Run a binary classification experiment with two probabilistic predictors and any internal local rule, such as the coordinatewise arithmetic mean, using an endpoint-monotone loss like cross-entropy, on a dataset where for every case the pooled probability is strictly closer to the true label than both inputs. Theorem 4 predicts the tree-relative complementarity Ψ is always ≤ 0; finding one such dataset with Ψ > 0 would refute it.","tokens_in":31157,"feed_emoji":"🌳","tokens_out":4152,"duration_ms":43562,"temperature":0.7,"pith_summary":"The paper tries to establish a formal theory of complementarity — a human-AI interaction beating the best available prediction — for workflows with more than two agents. It models an interaction as a protocol tree whose leaves carry prediction vectors and whose internal nodes combine them with a local rule. Against a pointwise-min benchmark, it proves that selector-style reliance can never be complementary, that in squared-loss regression complementarity reduces to Euclidean distance minimization with a closed-form optimal weight for two agents, and that in binary classification no internal averaging rule can achieve complementarity under endpoint-monotone losses such as cross-entropy. If correct, it would reorient empirical HAI research: complementarity is a genuine regression phenomenon, but classification studies must look beyond interpolating between human and AI probabilities.","feed_headline":"Regression enables human-AI complementarity; classification blocks it","feed_subtitle":"Why it matters: many classification workflows interpolate human and AI probabilities, yet interpolation can't beat the best prediction per c","key_machinery":"Protocol trees: rooted planar binary trees whose leaves are decorated by ordered agent prediction vectors and whose internal nodes carry a local binary composition rule; evaluating the rule recursively yields the protocol output. The tree-relative complementarity functional Ψ = Φ − Θ compares the pointwise-min benchmark loss Φ (Eq. 6) with the protocol loss Θ (Eq. 7). The argument is carried by the internality property (coordinatewise outputs remain between the input probabilities) combined with endpoint-monotone losses, and by the barycentric coordinate map φ_T from local linear-pooling weights to the simplex of leaf weights, with Tamari-cover reparameterizations preserving protocol output","core_discovery":"The central claim is a structural containment result: with the pointwise-min oracle benchmark (Eq. 6), the class of complementary multi-agent protocol outputs is empty for selectors and for internal rules in binary classification. Regression under squared loss is the positive counterpart: the complementarity functional satisfies nΨ = nK_n − ∥y − ŷ_T∥², so maximizing complementarity is equivalent to moving the protocol output close to the ground-truth vector; for N=2 the optimal linear weight is α* = Π[0,1](−B_n/A_n) with B_n = ⟨ŷ_H − ŷ_AI, ŷ_AI − y⟩, giving a residual-correction interpretation. Thus complementarity is attainable only when aggregation is non-internal or the loss is not endpoi","pith_inferences":["If the appropriate deployment benchmark is the best fixed predictor averaged over cases rather than the pointwise-min oracle, the classification impossibility dissolves; empirical complementarity findings may therefore be highly sensitive to this normative benchmark choice.","A practical design prescription follows: classification workflows seeking complementarity should deliberately extrapolate beyond the convex hull of agent probabilities — e.g., amplified logarithmic pooling — rather than averaging calibrated probabilities.","The Tamari-reparameterization result suggests a broader equivalence principle: under linear pooling, workflow topology matters only through the induced barycentric leaf weights, so two protocols with related parameter maps could be tested for equal loss on real teams.","A testable extension would add interaction costs (depth, monitoring burden) to the optimization, likely shifting the optimal protocol tree away from the complementarity-maximizing topology identified here."],"forward_implications":["Selector-based HAIs, including self-reliance and AI-reliance, never achieve complementarity relative to the pointwise-min benchmark, for any task, loss, or prediction quality.","In squared-loss regression, maximizing complementarity is Euclidean distance minimization from the ground-truth vector; for N=2, the optimal linear-pooling weight has a closed form and depends on whether the human-AI disagreement direction corrects the AI residual.","In regression with linear pooling, Tamari-cover reparameterizations of protocol trees preserve complementarity; for N=4, the two directed Tamari paths from the left comb to the right comb induce the same reparameterization, satisfying the pentagon identity.","In binary classification, no internal local rule can achieve complementarity under endpoint-monotone losses, including Bregman and many finite Bernoulli f-divergence losses; an analogous obstruction holds for coordinatewise-internal multiclass aggregation under cross-entropy.","Non-internal rules such as amplified logit pooling can escape the classification impossibility, although the relation between local outside-interval rates and global complementarity is not deterministic under unbounded cross-entropy."],"fun_headline_variants":["Regression enables complementarity; classification blocks it","Complementarity: regression yes, classification no","Classification obstructs complementarity; regression doesn't","Tree proof: complementarity attainable only in regression"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The impossibility results rest on measuring complementarity against the pointwise-min oracle benchmark (Eq. 6); if the right benchmark for a given HAI is instead the best fixed predictor's average loss, then Theorem 1 and Theorem 4 no longer apply and classification complementarity is not ruled out.","fun_headline_variants_meta":{"raw":{"variants":["Regression enables complementarity; classification blocks it","Complementarity: regression yes, classification no","Classification obstructs complementarity; regression doesn't","Tree proof: complementarity attainable only in regression"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001164,"raw_usage":{"total_tokens":4721,"prompt_tokens":878,"completion_tokens":3843,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":622,"completion_tokens_details":{"reasoning_tokens":3786}},"tokens_in":622,"tokens_out":3843,"duration_ms":26638,"temperature":1.0,"reasoning_tokens":3786,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T12:22:38.398672+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a binary classification experiment with two probabilistic predictors and any internal local rule, such as the coordinatewise arithmetic mean, using an endpoint-monotone loss like cross-entropy, on a dataset where for every case the pooled probability is strictly closer to the true label than both inputs. Theorem 4 predicts the tree-relative complementarity Ψ is always ≤ 0; finding one such dataset with Ψ > 0 would refute it.","supporting_citations":[],"review_version":2}