{"id":"a9fb8379-c66b-4e13-8c42-c985c680a5c0","arxiv_id":"2607.06570","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"Under a credal set, EVPI is concave so its lower envelope is generator-exact and its upper needs an LP, while Gamma-maximin VOI can exceed the whole classical envelope and is a distinct estimand.","lead":"The paper separates two ways to value information when probabilities are only known as a set: a decision-rule value (e.g. Gamma-maximin) and a range of classical VOI over every admissible measure. That split shows when research conclusions are robust across the evidence and when they hinge on one unsupported prior.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified to the central EVPI structural claim under the paper's stated assumptions.","rationale":"The reader's identification of the standing structural assumptions as the weakest point is accurate and already surface-level in the manuscript (Section 3 opening, Remarks 1-4, Proposition 2). Those assumptions are necessary for the exact endpoint statements and are not over-claimed for EVPPI/EVSI. The chemotherapy application correctly treats C0 as a convex hull and reports both generator max and LP upper endpoint, so the theory-application separation is already present. Because the load-bearing mathematics checks out under the stated hypotheses and the paper does not pretend the results hold without them, no verdict adjustment is warranted. The CONDITIONAL status correctly reflects pending public code and the methodological rather than transformative scope.","tokens_in":25043,"tokens_out":580,"duration_ms":6326,"concrete_test":"Independently re-derive representation (16) and the LP of Remark 2 from the definitions of f and g alone, without invoking any external Bayes-risk result; then recompute the two-state overshoot of Theorem 1(c) by hand. If both recover EVPI_LE = 2 > [0,1] and the LP optimum equals the claimed interior maximum, the central claim stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim (Theorem 1) is that EVPI is concave on a compact convex credal set, so for a finitely generated P = conv{G1,...,GK} the lower envelope equals min_k EVPI(Gk) exactly, the upper endpoint is a finite LP, and Gamma-maximin EVPI can strictly exceed the whole classical envelope. The proof is elementary and self-contained: representation (16) writes EVPI(P) = min_a EP[Da] with Da = M - NBa >= 0, each term affine, hence the pointwise minimum is concave; Jensen then pins the infimum at a generator; the LP of Remark 2 is the standard epigraph form of max_w min_a sum wk Dk,a; and the two-state mirror-image example of part (c) produces EVPI_LE = 2 > 1 = upper envelope under the deterministic named-strategy convention. The reader's weakest assumption (finite A, uniform bounds, compact convex P with continuous expectation maps) is exactly the standing hypothesis of Section 3 and is used only for Bauer's principle and continuity of the finitely many maps; it is not hidden. When those hypotheses fail the paper already flags that generator exactness and the LP need not hold (Proposition 2 for partial/sample information; nonconvex families noted in Section 5.2). No internal inconsistency or gap in the load-bearing argument was found.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper distinguishes two VOI estimands under a credal set of probability measures: a rule-specific value (e.g., Gamma-maximin / lower-expectation VOI) that fixes a decision criterion for acting under imprecision, and a fixed-measure envelope that records the range of classical single-measure VOI over the set. For EVPI it proves concavity on a compact convex credal set, exact attainment of the lower envelope endpoint at generators of a finitely generated set, a finite LP for the (possibly interior) upper endpoint, and an explicit two-state example in which Gamma-maximin EVPI strictly exceeds the entire classical envelope (Theorem 1). Continuity bounds, conditions under which EVPPI/EVSI endpoints remain generator-exact, nonnegativity of rule-specific VOI when constant policies are available, a unified outer-search estimation procedure, and a chemotherapy decision-model application complete the contribution.","tokens_in":25402,"tokens_out":1022,"duration_ms":9348,"significance":"If the structural claims hold, the paper supplies a clean and usable separation between robustness diagnostics and criterion-dependent decision values for VOI under imprecise probability. The concavity representation EVPI(P)=min_a E_P[D_a], the generator-exact lower endpoint, the LP for the upper endpoint, and the explicit overshoot example are elementary, checkable, and practically consequential: they show that Gamma-maximin VOI is not recovered from envelope endpoints and that vertex enumeration alone understates the upper EVPI range. The continuity bound, the EVPPI/EVSI counterexample, and the worked application that separates robust from measure-dependent research conclusions further strengthen the contribution for medical decision analysis and imprecise-probability decision theory. Reproducible code and explicit finite instances are additional strengths.","major_comments":[{"comment":"Section 5.2–5.3 and Table 2: the application treats C0 = conv{G1,...,G12} as the domain for exact EVPI endpoints, yet the continuous efficacy–harm correlation is handled only by a separate grid (an inner approximation). The paper correctly flags this, but the headline envelope [£84, £514] and the threshold classifications in Table 3 are therefore exact only for the zero-correlation slice. A short sensitivity statement quantifying how much the upper endpoint and the straddling of τ move when correlation is folded into the outer search (or a refined grid near the reported interior mixture) would make the applied claims load-bearing rather than illustrative.","section":"Section 5.2–5.3, Tables 2–3"},{"comment":"Section 4 and Algorithm 1: for continuous or finely discretized Z the rule-specific EVPPI/EVSI estimator (21) requires a controlled policy class, yet the manuscript only sketches the options (structured policies, DP, regression) and notes that Bayes-optimal policies under individual measures are a heuristic. Because the paper’s central distinction is that rule-specific value is not recovered from the envelope, a concrete, reproducible policy class (or an explicit statement that the chemotherapy EVSI numbers are fixed-measure only) is needed so that the estimation procedure actually delivers the rule-specific estimand it defines.","section":"Section 4, Eq. (21), Algorithm 1"}],"minor_comments":[{"comment":"Figure 1 is conceptually helpful but the threshold-classification bullets are dense; a one-line legend distinguishing robust vs. measure-dependent would improve readability.","section":"Figure 1"},{"comment":"Notation for the lower/upper expectations switches between E_P, E□, and V^LE; a short notation table or consistent subscripting would reduce cognitive load.","section":"Section 2.2"},{"comment":"The deterministic named-strategy convention that produces the overshoot is well explained in Remark 1, but a single forward pointer from the abstract or introduction would help readers who otherwise misread the claim as contradicting minimax theorems.","section":"Abstract / Remark 1"},{"comment":"Supplement Section 5 is cited for the full piecewise-linear EVPPI counterexample; ensuring that calculation is self-contained in the main text or clearly archived would aid verification of Proposition 2(a).","section":"Proposition 2"}],"recommendation":"minor_revision","confidential_remarks":"The central EVPI theory is sound and the overshoot example is clean; the two major comments are about making the applied and estimation claims match the precision of the theory, not about correcting errors. Fit for a methods-oriented statistics / decision-analysis venue is good. No novelty or citation concerns."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful bit is the formal separation of two estimands that people already mix up in practice: (i) VOI under a fixed decision rule for imprecision (Gamma-maximin etc.), and (ii) the range of classical single-measure VOI over a credal set. Theorem 1 is the load-bearing result and it holds under the paper’s stated assumptions. EVPI(P) = min_a E_P[D_a] with D_a = M − NB_a ≥ 0 is a pointwise min of affine maps, hence concave; Jensen pins the lower envelope at a generator when P = conv{G_k}; the upper endpoint is the standard epigraph LP over mixture weights; and the two-state mirror example shows Gamma-maximin EVPI can sit strictly above the whole classical envelope under deterministic named strategies. That overshoot is not a bug—it is the duality gap between max_a inf_P and inf_P max_a—and the paper is explicit that randomization would close the strict gap while still leaving the two objects distinct.\n\nWhat is new is not the affine/convex structure of Bayes values (DeGroot, Berger) or maximin information values (Szaniawski, Gilboa–Schmeidler, Bradley–Steele), but the packaging for applied VOI: which envelope endpoints are generator-exact, which need an LP or interior search, when EVPPI/EVSI lose vertex attainment (Prop 2, with an explicit interior-max counterexample), the sharp TV Lipschitz constant, and the constant-policy nonnegativity condition for rule-specific VOI. The chemotherapy example is honest: EVPI endpoints over the 12-generator convex hull are treated as exact (apart from MC error), EVSI as regression approximation, and the decision flip across generators is the point of the threshold table.\n\nSoft spots are real but proportionate. Finite A, bounded net benefits, and compact convex P are standing hypotheses, not hidden; if the admissible set is nonconvex the upper EVPI number drops from the LP value to the generator max, and the paper says so. Public code is promised rather than hashed here. Practical impact is methodological inside medical decision analysis and robust Bayes, not a new decision theory. Citation pattern is fair: prior information-aversion and dilation work is used, not papered over.\n\nThis is for people who already run VOI under structural or prior uncertainty and need to stop reporting one unsupported measure as if it were a robustness analysis. Math and examples check out. I would send it to referees.","headline":"Clean, checkable split between rule-specific VOI and fixed-measure envelopes, with real EVPI structure (concavity, generator lower endpoint, LP upper, Gamma-maximin overshoot) that the literature had not packaged this way.","tokens_in":26043,"tokens_out":640,"would_cite":true,"duration_ms":6857,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62C10","90B50","91B06"],"pacs":[],"model":"grok-4.5","headline":"When evidence leaves a whole set of probabilities, the value of information splits into a rule-specific number and a classical envelope that can disagree.","keywords":["imprecise probability","credal sets","value of information","EVPI","Gamma-maximin","lower previsions","probability bounds analysis","decision making under uncertainty"],"falsifier":"Construct any decision model and finitely generated convex credal set in which the classical EVPI lower envelope is not equal to the minimum generator EVPI, or in which the Gamma-maximin EVPI cannot exceed the classical upper envelope; either counter-example would refute the structural claims of Theorem 1.","tokens_in":25901,"feed_emoji":"📊","tokens_out":1053,"duration_ms":8982,"temperature":0.7,"pith_summary":"Standard value-of-information calculations pick one probability measure and report how much better decisions would be with more data. Real evidence often leaves a whole set of plausible measures (a credal set). This paper shows that under that set there are two different estimands that must not be mixed. A rule-specific value asks how much the information is worth to a decision maker who already uses a fixed rule for acting under imprecision (for example Gamma-maximin). A fixed-measure envelope simply records the range of ordinary VOI values as the measure runs over the set. For perfect information the classical functional is concave, so on a finitely generated set the lower envelope endpoint is exactly the smallest generator value while the upper endpoint is found by a linear program and can sit in the interior. The Gamma-maximin value can sit strictly above the whole envelope, so it is not recovered from the envelope endpoints. The paper supplies a continuity bound, conditions under which partial- and sample-information endpoints still come from generators, a unified estimation recipe that wraps ordinary single-measure estimators in a search over the set, and a chemotherapy example that separates conclusions true for every admissible measure from conclusions that hang on one unidentified choice.","feed_headline":"VOI under imprecise probabilities splits in two","feed_subtitle":"Rule-specific value can exceed the whole classical envelope; lower EVPI is exact on generators","key_machinery":"Concavity of EVPI(P) = min_a E_P[M - NB_a] (a pointwise minimum of affine maps), which yields exact lower-endpoint evaluation on the generators of a finitely generated credal set, an LP for the upper endpoint, and an explicit duality-gap example showing that Gamma-maximin EVPI can overshoot the classical envelope.","core_discovery":"Under a credal set the value of information must be split into a rule-specific quantity (what the information is worth once a decision rule for imprecision is fixed) and a fixed-measure envelope (the range of classical VOI over every admissible precise measure). EVPI is concave on the set, so its lower envelope endpoint equals the minimum over extreme generators while its upper endpoint may be interior and is given by a finite linear program; the Gamma-maximin EVPI can strictly exceed the entire envelope.","pith_inferences":["Health-technology assessment and research-prioritization bodies that already run probabilistic sensitivity analysis could treat the envelope as a routine robustness report without abandoning their existing single-measure estimators.","The same split between rule-specific value and fixed-measure envelope should apply to other imprecise-probability decision criteria (minimax regret, weighted lower-upper rules) once the aggregation functional is swapped.","When the admissible set is defined by p-boxes or moment constraints rather than an explicit finite generator list, the linear-program upper endpoint becomes an infinite-dimensional optimization that still inherits the same concavity structure."],"forward_implications":["A single reference VOI that lies on one side of a research threshold while the envelope straddles that threshold is assumption-dependent and should not be treated as robust.","Reporting only the largest classical EVPI over generators systematically understates the true upper envelope when the functional is concave.","Rule-specific lower-expectation VOI is nonnegative whenever the policy class still contains constant (ignore-the-signal) policies; without them free information can have negative value.","Partial- and sample-information envelopes generally require interior search; vertex enumeration is exact for their upper endpoints only when one action is pre-information optimal for every measure in the set."],"fun_headline_variants":["VOI on credal sets: rule-specific value can exceed full envelope","EVPI concave on credal set: lower end exact on generators","Gamma-maximin VOI can strictly beat classical envelope bounds","Fixed-measure VOI envelope vs rule-specific value under imprecision","Credal EVPI upper endpoint may be interior; solve by LP"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The argument needs a finite action set, uniformly bounded net benefits, and a nonempty compact convex set of measures in a topology that makes the relevant expectations continuous; if the real admissible set is nonconvex or unbounded the exact generator and linear-program endpoints need not hold.","fun_headline_variants_meta":{"raw":{"variants":["VOI on credal sets: rule-specific value can exceed full envelope","EVPI concave on credal set: lower end exact on generators","Gamma-maximin VOI can strictly beat classical envelope bounds","Fixed-measure VOI envelope vs rule-specific value under imprecision","Credal EVPI upper endpoint may be interior; solve by LP"]},"model":"grok-4.5","effort":"low","cost_usd":0.005794,"raw_usage":{"total_tokens":1553,"prompt_tokens":881,"num_sources_used":0,"completion_tokens":76,"cost_in_usd_ticks":57940000,"prompt_tokens_details":{"text_tokens":881,"audio_tokens":0,"image_tokens":0,"cached_tokens":0},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":596,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":881,"tokens_out":76,"duration_ms":5474,"temperature":1.0,"reasoning_tokens":596,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T11:34:08.949366+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Construct any decision model and finitely generated convex credal set in which the classical EVPI lower envelope is not equal to the minimum generator EVPI, or in which the Gamma-maximin EVPI cannot exceed the classical upper envelope; either counter-example would refute the structural claims of Theorem 1.","supporting_citations":[],"review_version":1}