{"id":"b16a6086-8c9a-4978-b47b-e608cf2d1821","arxiv_id":"2508.20538","paper_version":5,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper claims that a wrong cascade is breakable whenever p(1+V) >= 2c, but the derivation contains an arithmetic error that changes the condition.","lead":"This paper studies when costly, publicly observable fact-checking can overturn a false information cascade. It claims a simple inequality on the private reward to investigation determines whether wrong herds are broken, and derives policy levers from that condition.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 5.1's universal breakability claim fails in state θ=0: verification yields evidence only in θ=1, so at the upper boundary a disclosure event has probability zero even when (10) holds.","rationale":"The reader's algebraic concern about Eq. (8) is real but does not by itself break the knife-edge: Eq. (9), and hence condition (10), is the correct comparison if the failed-investigation branch is evaluated at the updated posterior, since the unconditional continuation value is 1/2. The reader's soft-channel concern about Theorem 4.2 is valid — nearest-grid rounding can make actions distinguishable even when γΔx<h, so the 'iff' is false — but that mainly affects the sufficient soft condition and is secondary to the hard-channel result. The decisive issue is that Theorem 5.1 quantifies over both classical boundaries while the evidence technology is one-sided. At μ=q with θ=0, no evidence can ever arrive, so the theorem's claim of a positive-probability disclosure event is false, and the upper-boundary half of the resilience frontier in Theorem 6.2 is unsupported. This is not a question of equilibrium selection or tie-breaking; it is a direct consequence of the model's own assumption that Pr(evidence | θ=0)=0. The paper's abstract even acknowledges that attempts can be finite almost surely and failures can lead to absorbing cascade traps, which is in tension with the unqualified Theorem 5.1. I therefore agree with the reader's overall REJECT verdict, but for a different and more fundamental reason; the reader's weakest-assumption focus on unobservability of failed attempts, while related, is not the same objection.","tokens_in":13737,"tokens_out":26455,"duration_ms":255204,"concrete_test":"Analytic check: fix any parameters satisfying (10), for instance q=3/5, p=1/2, V=1, c=1/2, and set the history at the upper boundary μ=q with true state θ=0. Since Section 3.5 gives Pr(evidence | θ=0)=0, the conditional probability of a finite-time disclosure is exactly zero. If Theorem 5.1 is read literally, it predicts a strictly positive probability of some finite-time disclosure; the model predicts 0. This contradiction is resolved by restricting the theorem to θ=1, as the proof implicitly does.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing gap is state asymmetry. Section 3.5: evidence arrives with probability p>0 only when θ=1 and with probability 0 when θ=0. Theorem 5.1 asserts that at any classical boundary μ∈{1−q,q}, condition (10) makes any wrong cascade breakable and, in particular, makes a finite-time disclosure event have strictly positive probability. Take the upper boundary μ=q with true state θ=0: this is a wrong cascade. The pro-truth signal is s=0, the posterior is x=1/2, and (10) can hold (e.g., p=1/2, V=1, c=1/2 gives equality). Agents with s=0 investigate, but because θ=0, the objective probability of evidence is zero; no disclosure ever occurs. The proof's step 'verifiable evidence arrives with probability p>0' silently assumes the true state is θ=1. The knife-edge is a condition on the agent's subjective expected payoff, not on the true state, so it cannot produce an objective positive-probability disclosure event in θ=0. Therefore the theorem's universal formulation, and the upper-boundary half of Theorem 6.2's resilience frontier, are false as stated; the result can hold only for wrong cascades in the evidence-generating state θ=1, if at all.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper models sequential social learning with continuous actions, conformity, coarse public observation of actions, and costly private investigation that can produce verifiable evidence only in state θ=1. Section 4 derives a soft-channel informativeness threshold (Theorem 4.2); Section 5 derives a boundary knife-edge for verification (Theorem 5.1) and claims that whenever p(1+V)≥2c, any wrong cascade at a classical boundary μ∈{1−q,q} is breakable with positive probability in finite time via disclosure. Section 6 combines this with the soft-channel condition into a resilience frontier (Theorem 6.2), and Section 7 translates the knife-edge into policy instruments. The central claim is that the private return to a one-shot investigation at a wrong boundary exceeds its cost exactly when p(1+V)≥2c, and that this guarantees eventual correction of false cascades.","tokens_in":14084,"tokens_out":13932,"duration_ms":126487,"significance":"If correct, the paper would offer a transparent, closed-form condition linking verification incentives to the collapse of false cascades, with clean comparative statics and policy implications; the soft-channel result on coarsened continuous actions (Theorem 4.2) is also interesting and largely elementary. The paper is self-contained, works out posterior-gap expressions explicitly, and does not fit free parameters. However, the central theorem is not established: the boundary calculation in Section 5.1 contains an algebraic error, and the state-asymmetry of the evidence technology invalidates the universal 'any wrong cascade' claim. These are load-bearing for Theorem 5.1, Theorem 6.2, and all policy results, so the contribution as stated is not sound.","major_comments":[{"comment":"The net-gain calculation is algebraically wrong as written. With the stated no-evidence continuation payoff U_no = 1/2, Eq. (8) gives U_inv = (p/2)(1+V) + (1−p/2)(1/2) − c = 1/2 + pV/2 + p/4 − c, so U_inv − U_no = pV/2 + p/4 − c = (p/2)(V + 1/2) − c, not (p/2)(1+V) − c. Consequently condition (10), p(1+V) ≥ 2c, does not imply that investigation is optimal. For example, with p=1/2, V=0, c=1/6, condition (10) holds (1/2 ≥ 1/3) but the true net gain is 1/8 − 1/6 < 0. If the authors instead intend the agent to update on the observed failure and choose an optimal action, then Eq. (8) is incorrect because the no-evidence continuation payoff should be 1/(2−p), not 1/2. Either way the derivation in the manuscript is internally inconsistent, and the error propagates into Theorem 5.1, Corollaries 5.3–5.4, Theorem 6.2, Proposition 6.4, and Section 7.","section":"Section 5.1, Eqs. (8)–(9)"},{"comment":"The theorem's universal formulation is false because of the one-sided evidence technology. Section 3.5 specifies that evidence arrives with probability p>0 only when θ=1 and with probability 0 when θ=0. At the upper boundary μ=q with true state θ=0, the cascade is wrong, the pro-truth signal is s=0, and condition (10) may hold, but the objective probability of a disclosure event is exactly zero. The proof's sentence 'Conditional on investigating, verifiable evidence arrives with probability p>0' silently assumes the true state is θ=1. By Corollary 4.3 the soft channel is locally mute at the boundary, so no positive-probability exit path exists. Thus Theorem 5.1 can hold at best for the lower-boundary wrong cascade (θ=1), and the hard-evidence half of Theorem 6.2 is not valid at the upper boundary. Footnote 3's asymmetric generalization does not rescue the result, since with p̃=0 the disclosure probability in θ=0 remains zero.","section":"Section 5.2, Theorem 5.1 and Section 3.5"},{"comment":"The payoff specification is internally inconsistent. Section 3.3 has the agent minimize γE[(a−θ)^2 | x] + (1−γ)(a−μ_t)^2, yielding the interior best response a⋆(x,μ)=γx+(1−γ)μ. Section 3.6, however, defines the per-period accuracy payoff as 1 if the chosen action matches the state and 0 otherwise. For continuous actions, the expected exact-match payoff of an interior action is generally not equal to the quadratic loss, and at the boundary posterior x=1/2 the value of the action rule (3) is not 1/2. Therefore the baseline value U_no=1/2 used in Eq. (7) does not follow from the specified primitives unless the action set is effectively binary or the payoff is redefined. The manuscript needs to state a single, coherent payoff function; as it stands, the verification calculation and the action rule are mutually incompatible.","section":"Sections 3.3 and 3.6"}],"minor_comments":[{"comment":"The title on the arXiv page, 'Two Motives for Verification in Information Cascades', does not match the title in the full text, 'False Cascades and the Cost of Truth'. The supplied abstract also describes a model with publicly observed investigation and a two-component investigation set, which does not match Sections 3.5 and 3.7 of the body, where investigation attempts are unobserved and the investigation set is not characterized. The abstract and body need to be reconciled.","section":"Title and abstract"},{"comment":"The 'if and only if' statement fails at the boundary case h=0. If h=0, the coarsening map is the identity, but when γ=0 the two signal-contingent actions coincide, so the public cannot infer the signal even though |Δa|≥0 holds. The theorem should either restrict to h>0 or use a strict inequality for distinguishability.","section":"Theorem 4.2"},{"comment":"Figure 2 plots the knife-edge in (p, c/V) space, but the ratio c/V is undefined when V=0, which is an allowed parameter value. The axes or the normalization should be defined so that V=0 is covered. The expression 'min{1−p, 2c/(1+V) − p}+' in Proposition 7.1 uses an unexplained '+' notation; the positive-part operator should be defined.","section":"Figure 2 and Section 7.2"},{"comment":"The corollary inherits the state-asymmetry problem: with evidence only in state θ=1, a positive probability of a 'high-V' or 'low-c' type does not generate disclosures when the true state is θ=0. The statement should be restricted to the evidence-generating state, or the model should be extended to symmetric evidence.","section":"Corollary 5.4"}],"recommendation":"reject","confidential_remarks":"The algebraic error in Section 5.1 could in principle be repaired by correcting the knife-edge condition, provided the intended continuation payoff is clarified. The state-asymmetry problem, however, is structural: with one-sided evidence, no disclosure can ever occur in θ=0, so Theorem 5.1 and the resilience frontier cannot be repaired by a local correction. The paper would need a substantive reformulation of the evidence technology or of the statement of the main theorems. I therefore recommend rejection in the current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The knife-edge condition in Theorem 5.1 is algebraically wrong, and the proof ignores the one-sided nature of evidence, so the universal breakability claim is false as stated. The paper is not ready for publication, but the underlying setup is sensible and one of its subsidiary results is worth keeping.\n\nWhat is new: the model puts continuous conformist actions, coarse observation, and one-sided verifiable disclosure into one sequential-learning framework. The soft-channel threshold in Theorem 4.2—actions are publicly informative iff γΔx ≥ h—is clean and correct, and explains why continuous actions don't automatically kill cascades when conformity is strong and actions are rounded. The policy levers (subsidies, verification quality, disclosure prizes) follow straightforwardly, and the comparative statics are reasonable.\n\nThe soft spots are where it hurts. In Section 5.1, equation (8) expands to U_inv = (1/2)p(1+V) − (1/4)p + 1/2 − c, so U_inv − U_no = (1/2)p(1/2+V) − c, not (1/2)p(1+V) − c as in (9). The claimed condition p(1+V) ≥ 2c is therefore not sufficient; the correct knife-edge is p(1/2+V) ≥ 2c. Parameters can satisfy the former while the latter fails, so Theorem 5.1's premise doesn't imply its conclusion. Second, even with the correction, the proof of Theorem 5.1 assumes evidence can arrive. But the model has evidence only in state θ=1. At the upper boundary μ=q with true state θ=0, a wrong cascade, the pro-truth signal leads agents to investigate, yet the objective probability of evidence is zero. No disclosure ever occurs. So the theorem can hold only for wrong cascades in the evidence-generating state, i.e., at the lower boundary; the universal formulation and the upper-boundary half of Theorem 6.2 are false. The paper's own abstract even contains a warning that positive chance of proof does not guarantee eventual discovery, which sits awkwardly next to Theorem 5.1.\n\nWho is this for? Someone working on cascades and verification might pick up the soft-channel threshold and the modeling approach, but they should not trust the main result until it's fixed. The paper deserves a serious referee because the question is well-posed and the fix is not obviously impossible; the algebra can be corrected and the theorem restated with the proper state restriction. My recommendation: send it to review, but expect a major revision; the current version shouldn't be accepted.","headline":"Two load-bearing flaws—an algebra error in the knife-edge and an unhandled state asymmetry—invalidate Theorem 5.1 as stated, but the soft-channel threshold and the unified setup give the paper enough substance to merit a major-revision path.","tokens_in":14513,"tokens_out":4312,"would_cite":false,"duration_ms":37839,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A simple inequality in verification costs and rewards determines when costly investigation can overturn a wrong cascade; below that threshold, coarse observation can make the error persist.","keywords":["social learning","information cascades","verification","disclosure","conformity","coarse observation","cascade breakability","resilience frontier"],"falsifier":"One way to test the central claim is to compare groups engaged in a sequential guessing task with costly verification: vary the cost \\(c\\) relative to the discovery probability \\(p\\) and reward \\(V\\), and measure whether wrong herds are eventually overturned. The theory predicts a sharp threshold: correction occurs with positive probability in finite time exactly when \\(p(1+V)\\ge 2c\\), and not when the inequality fails and observation is coarse. A second test targets the key assumption: make failed verification attempts public in one treatment and private in another; the model implies that visible failures reduce or eliminate the eventual correction.","tokens_in":13577,"feed_emoji":"🔍","tokens_out":6022,"duration_ms":51285,"temperature":0.7,"pith_summary":"The paper asks when a society that has herded on a wrong belief can correct itself if agents may pay to investigate and produce hard, publicly verifiable evidence. It shows that continuous actions do not by themselves kill cascades once conformity and coarse observation are present: at the classical cascade boundaries the soft action-based channel is locally mute. It then proves that if the expected private return to a one-shot investigation exceeds its cost, \\(p(1+V)\\ge 2c\\) at the boundary, then any wrong cascade is breakable with positive probability in finite time via disclosure. Combining the two margins yields a resilience frontier with comparative statics in signal quality, conformity, observability, verification quality, and investigation cost, together with least-cost policy instruments. A sympathetic reader would care because the result converts the vague question of whether false narratives can persist into a transparent inequality in primitives.","feed_headline":"One inequality decides whether false cascades break","feed_subtitle":"Costly verification overturns wrong herds exactly when its expected gain beats cost; otherwise they can persist forever.","key_machinery":"The load-bearing objects are the signal-induced action separation \\(\\$\\Delta$ a(\\mu,q,\\gamma)=\\gamma\\$\\Delta$ x(\\mu,q)\\), which must meet the observation grid width \\(h\\) for actions to be publicly informative, and the boundary knife-edge \\(p(1+V)\\ge 2c\\), which compares the expected private return from a one-shot investigation (probability \\(\\frac12 p\\) of discovering the truth and receiving reward \\(V\\)) with its cost \\(c\\). The first object fixes when soft learning works; the second fixes when hard evidence breaks a wrong cascade. The proof that the knife-edge is sufficient rests on the boundary property that the pro-truth posterior is exactly \\(1/2\\), so the investigation option has a transparent expected value, and on the assumption that a failed attempt leaves the public belief unchanged.","core_discovery":"Working in a sequential model with binary states, continuous actions, conformity, and a coarsened public observation of actions, the paper's central claim is that informational resilience has two separate channels. The soft channel is public only if the signal-induced action separation \\(\\gamma\\$\\Delta$ x(\\mu,q)\\) clears the observation granularity \\(h\\); because \\(\\$\\Delta$ x=0\\) at the classical boundaries \\(\\mu\\in\\{1-q,q\\}\\), actions are locally mute there for any finite responsiveness when \\(h>0\\). The hard channel operates through costly investigation that yields verifiable evidence only in the true state with probability \\(p\\), and the paper proves the knife-edge result: whenever \\(p(1+V)\\ge 2c\\), the expected private gain from investigating after the pro-truth signal at a wrong boundary exceeds the cost, so disclosure occurs with positive probability in finite time and resets belief to the truth. Section 6 elevates this into a resilience theorem: if at each boundary either the soft informativeness condition holds locally or the hard knife-edge holds, then no wrong cascade persists forever. The model also provides closed-form least-cost policies for subsidies, verification-quality investment, and disclosure prizes.","pith_inferences":["If failed investigation attempts were publicly observable, the repeated-trial argument would break down: each failure would move the public belief away from the boundary posterior, so the pro-truth posterior would no longer be exactly \\(1/2\\) and the incentive to investigate on later rounds weakens. The model's unobservable-attempt assumption is therefore not an innocuous detail.","The same boundary logic suggests an extension with false positives in the evidence technology: if evidence can arrive in both states with probabilities \\(p\\) and \\(\\tilde p\\), the knife-edge becomes \\((p-\\tilde p)(1+V)\\ge 2c\\), which the paper mentions in a footnote but does not develop.","A directly testable prediction is that making fact-checking attempts visible to other agents, for example through activity logs, should slow or prevent self-correction compared with private attempts, because failed checks push beliefs away from the boundary posterior that sustains the investigation incentive."],"forward_implications":["If \\(p(1+V)\\ge 2c\\), a wrong cascade at a classical boundary cannot persist forever: with strictly positive probability a disclosure resets public belief to the truth in finite time.","When the knife-edge fails and the soft channel is locally mute at the boundary, the wrong cascade is absorbing unless some other force intervenes.","Improvements in signal precision \\(q\\), responsiveness \\(\\gamma\\), or observation fineness (smaller \\(h\\)) expand the resilience region on the soft side; increases in verification quality \\(p\\), disclosure reward \\(V\\), or decreases in investigation cost \\(c\\) expand it on the hard side.","A planner can guarantee boundary breakability at least cost by choosing the cheapest single subsidy, verification-quality increase, or prize that satisfies the knife-edge; with convex costs an interior mix may be optimal.","With heterogeneous costs and rewards, it suffices that a positive mass of agents satisfy the knife-edge for a wrong boundary cascade to be breakable with positive probability in finite time."],"supporting_citations":[{"why":"supplies the classic herd-behavior benchmark that motivates the cascade question","marker":"[2]"},{"why":"defines information cascades and the boundary beliefs at which actions stop revealing signals","marker":"[5]"},{"why":"shows actions can be pathologically uninformative, the basis for the soft-channel muteness result","marker":"[18]"},{"why":"establishes verifiable-disclosure unraveling, the hard-evidence technology used for investigation","marker":"[12]"},{"why":"provides the representation-theoretic foundation for hard evidence and disclosure","marker":"[15]"},{"why":"models costly search in social learning, the baseline against which the investigation incentive is compared","marker":"[16]"}],"fun_headline_variants":["Two motives for verification: when investigation saves or sinks","A knife-edge condition decides if false cascades break","Costly proof can trap beliefs unless its gain outweighs cost","Verification's diagnostic and reward motives shape cascade traps","When does investigation break wrong herds? One inequality decides"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument's load-bearing premise is that a failed investigation attempt is not publicly observed unless evidence is disclosed, so after a failure the public belief stays at the boundary posterior and the same knife-edge condition applies in every subsequent period.","fun_headline_variants_meta":{"raw":{"variants":["Two motives for verification: when investigation saves or sinks","A knife-edge condition decides if false cascades break","Costly proof can trap beliefs unless its gain outweighs cost","Verification's diagnostic and reward motives shape cascade traps","When does investigation break wrong herds? One inequality decides"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000207,"raw_usage":{"total_tokens":1415,"prompt_tokens":972,"completion_tokens":443,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":365}},"tokens_in":588,"tokens_out":443,"duration_ms":5226,"temperature":1.0,"reasoning_tokens":365,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:45:03.247917+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One way to test the central claim is to compare groups engaged in a sequential guessing task with costly verification: vary the cost \\(c\\) relative to the discovery probability \\(p\\) and reward \\(V\\), and measure whether wrong herds are eventually overturned. The theory predicts a sharp threshold: correction occurs with positive probability in finite time exactly when \\(p(1+V)\\ge 2c\\), and not when the inequality fails and observation is coarse. A second test targets the key assumption: make failed verification attempts public in one treatment and private in another; the model implies that visible failures reduce or eliminate the eventual correction.","supporting_citations":[{"cited_title":"and Sørensen, P","cited_arxiv_id":null,"evidence_quote":"shows actions can be pathologically uninformative, the basis for the soft-channel muteness result"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"establishes verifiable-disclosure unraveling, the hard-evidence technology used for investigation"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the representation-theoretic foundation for hard evidence and disclosure"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"models costly search in social learning, the baseline against which the investigation incentive is compared"}],"review_version":2}