| paper | d4-conf-ijcai-2019-0614 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | medium |
| authors would recognise it | no |
Theorem 4 is computational, but its explanation length counts premise occurrences in a fixed knowledge base. The proposed consultation model makes endorsement prevalence a fractionally selectable proof resource; the paper gives no semantics for that, and the proponent acknowledges this changes what counts as a short explanation. I do not find a population mirror of the anchored result that the authors would recognize as their problem.
fails bit b-changes — the continuum limit changes the question
The proponent notes that length differs from mass, but does not establish why repeated endorsements should become a scarce, fractionally selectable resource when the paper treats each premise as a preference constraint.
fatal: True
The proposal targets Theorem 4's bounded-length explanation decision; it leaves the inference and preference-structure results without a population mirror.
The strongest positive case is the paper’s explanation-complexity result, Theorem 4, proved here. It says that, given alternatives \(x,y\) and positive integers \(n,m\), deciding whether there is an elliptic cancellative explanation of \((x,y)\) of length at most \(m\), with at most \(n\) repetitions of the conclusion, is NP-complete. The result remains NP-complete when \(n=1\). The paper proves hardness by reduction from Vertex Cover, citing Karp (1972).
A plausible high-multiplicity setting is a large public consultation on policy packages. Each person is asked to give a holistic comparison between two packages described by fixed criteria. A type is the complete comparison the person gives; people giving the same comparison are interchangeable for the explanation task. The mass \(\mu_t\) of type \(t\) is its fraction of respondents. A consultation might have hundreds of thousands of participants but only dozens or hundreds of distinct comparison types, because everyone answers from the same finite questionnaire.
Here is a continuous version of Theorem 4’s underlying problem, which I would call Mass-Bounded Cancellative Explanation. An instance has finite ordered feature domains, a target comparison \(q=(x,y)\), a finite support \(T\) of stakeholder comparison types with rational masses \(\mu_t\) summing to one, and a rational budget \(B\). Ask whether there are real numbers \(\lambda_t\), with \(0\leq\lambda_t\leq\mu_t\) and \(\sum_t\lambda_t\leq B\), such that the selected fractions of the comparison types support a cancellative explanation of \(q\), allowing the paper’s implicit dominance statements. A solution is the allocation \((\lambda_t)_{t\in T}\), together with the dominance completion. Equivalently, minimize the total cited population mass \(\sum_t\lambda_t\). The cancellation requirements can be expressed as the paper’s finite threshold-difference inequalities over the feature values appearing in the instance, so this version is an LP with size polynomial in the feature count, support size, and input bit length. I expect this mirror to be tractable.
The interpretation is that the explanation task remains the same: justify a preference comparison by combining reference comparisons and dominance. What changes is the supply of premises and the measure of explanation length. A decision maker’s finite list of statements becomes a large pool of people supplying those statements; citing \(2\%\) of a cohort is meaningful in the consultation setting, just as “\(2\%\) of residents support this comparison” is. The mirror covers Theorem 4’s short-explanation problem, especially its \(n=1\) hard case. It does not claim to continuize the paper’s entire theory of preference inference.
The weakest point is that the paper measures explanation length in premise occurrences, not in the mass of people who endorse them. Allowing fractional use of a cohort changes what counts as a short explanation, and an author might reasonably find “use this fraction of an endorsement type” less cognitively natural than citing a whole statement. Also, the Vertex Cover reduction’s combinatorics live in the feature structure, so hardness may survive if premises remain indivisible. Thus the LP would show that a meaningful population-based explanation problem is tractable; it would not show that Theorem 4’s original single-decision-maker problem becomes tractable. The next questions are whether continuous witnesses round to short discrete explanations, what integrality gap this creates, and which consultation settings make fractional evidence genuinely useful.
Theorem 4 is a real computational anchor, so the negative case is not that the paper lacks a complexity result. The problem is what that result computes: the shortest cancellation certificate for a particular decision maker’s fixed knowledge base \(P\). If a consultation supplies those premises, the natural pooled knowledge base is the set of distinct comparisons respondents make. Inference depends on that set, not on how many people repeat each comparison. If instead each person’s complete type is their own knowledge base, the explanation problem remains separate for each type; population masses do not change its answer.
The proposed mass-bounded explanation gives mass a job the paper’s logic does not assign it. A premise is a comparative constraint, not a unit of corroboration: one respondent’s statement and a thousand identical statements impose the same constraint. Selecting a fraction of a type then means using a fractional copy of a premise. That may define a weighted or probabilistic argument scheme, but it changes what counts as an explanation. The paper’s LP for inference uses real coefficients to test whether a conclusion follows; that does not make those coefficients fractions of a population or explain why prevalence should price an argument.
Keeping premises indivisible does not resolve the mismatch. In a continuous population, citing one person of each needed type has zero mass; charging the full mass of each type instead makes a popular premise more expensive to cite, even though its repeated endorsement adds no new reason. A new rule could give mass a different interpretation, but that rule—not the paper’s cancellation scheme—would supply the substantive model. Asking which fraction of a population prefers a comparison would likewise be a population aggregation question, rather than a mirror of Theorem 4’s explanation problem.
So the strongest case against is that no population quantity naturally enters the paper’s existing computational task: prevalence is irrelevant to its inference, while making prevalence matter requires changing the meaning or cost of an explanation. This is a substantial objection, but not an airtight universal one. A collective decision process that explicitly treats endorsement mass as usable evidence might make the proponent’s variant worthwhile if the authors accept that new semantics. The paper alone gives little reason to think they would.
The adversarial triple: the proponent anchors on up to three named results; the opponent sees that case and must defeat every anchor; the judge decides which case convinced it. These are the pipeline’s own outputs, generated by tools/triple_run.py — no human edited them. The paper’s own text is not reproduced here beyond the quoted statement above.