| paper | d4-conf-aaai-2017-0899 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | no |
Theorem 1 gives a truthful-reporting condition and Theorem 2 gives a subsidy condition, but neither states a computational result under the stated gate. The type scan and fee calculation do not supply that missing named result, so there is no qualifying anchor for a continuous computational mirror.
fails bit a — no named computational result to mirror
The proposed calibration uses Theorems 1 and 2; it leaves the paper's remaining market and peer-prediction analysis untouched. Neither proposed anchor qualifies as a computational result.
There is a narrow affirmative case. The paper has no theorem classifying a problem as polynomial-time solvable or hard. Its best anchors are instead two named, proved-here guarantees that compute how much peer reward is needed and whether market fees can fund it. Those support a limited algorithmic mirror: calibrating truthful crowd arbitration from a type distribution.
A plausible regime is a decentralized prediction market that assigns each event to a large pool of vetted assessors. Many assessors share the same relevant features: signal model, budget bound, and market exposure. Let the population have \(N\) agents and \(\tau\) such types, with \(N\gg\tau\). A type records an agent’s role, budget and exposure class, and current binary signal; each individual still reports \(0\) or \(1\). If the arbiter share of the population is \(\rho\), then there are \(m_N=\rho N\) arbiters. Let \(r_t\) be the fraction of type \(t\) reporting \(1\); the market outcome is the mass fraction of arbiters reporting \(1\), as in the paper. This preserves its MSR, its fractional outcome, and its 1/prior-with-midpoint peer payments. Assume arbiters share the paper’s signal parameters \(\mu,\mu_0,\mu_1\), with update strength \(\delta=\mu_1-\mu_0>0\).
Theorem 1, proved in this paper, is the first anchor. It states that, when all other arbiters report truthfully, truthful reporting is a best response for arbiter \(i\) if
\[
k\ge \frac{2\max\{|\phi_b^-(B)|,|\phi_b^+(B)|\}}{m\delta}.
\]
For agents who may participate in the market multiple times, the paper gives the bound
\[
k\ge \frac{2B(1+f)}{fm\delta}.
\]
Here \(B\) bounds an agent’s market budget, \(f\) is the trading fee, and \(k\) scales peer payments.
Its continuous counterpart is Mass-Threshold Truthful Arbitration. The instance gives a finite rational support of arbiter types and their masses, a population scale \(N\), a common signal model, a fee \(f\), and type-specific budget bounds \(B_t\). The question is: what is the smallest common peer-payment scale \(k\) that certifies truthful reporting as a best response for every type when all other arbiters report truthfully? With \(B=\max_t B_t\), the paper’s repeated-participation guarantee gives
\[
k\ge \frac{2B(1+f)}{fN\rho\delta}.
\]
This is a well-posed type-compressed calibration problem, expected to be tractable by scanning the \(\tau\) types and evaluating the bound. The theorem supplies a guarantee, not a claim that this bound is always the exact minimum.
Theorem 2, also proved here, is the stronger anchor. It states that fee revenue suffices to fund truthful arbitration without an outside subsidy if
\[
fM\ge
\frac{2\max\{|\phi_b^-(B)|,|\phi_b^+(B)|\}}{\delta},
\]
or, for multiple market entries,
\[
fM\ge \frac{2B(1+f)}{f\delta}.
\]
Here \(M\) is the paper’s aggregate worst-case-loss measure that determines fee revenue. The theorem’s striking feature for this mirror is that its multiple-entry bound does not depend on the number of arbiters.
The lead mirror is Mass-Based Fee Calibration. Given the same type distribution and signal model, the arbiter fraction \(\rho\), budget caps \(B_t\), and aggregate fee base \(M_N\), find the smallest fee \(f>0\) for which the peer rewards can be funded while the paper’s truthful-best-response guarantee holds. Set \(B=\max_t B_t\). The paper’s condition becomes
\[
fM_N\ge \frac{2B(1+f)}{f\delta}.
\]
For \(M_N>0\), the least fee satisfying this guarantee is
\[
f_N^*=
\frac{B+\sqrt{B^2+2BM_N\delta}}{M_N\delta}.
\]
Thus this version remains tractable under type compression: it needs a maximum over the support and a scalar calculation. If per-capita market activity stays roughly constant so \(M_N=N\bar M\), the certified fee falls on the order of \(N^{-1/2}\). That gives a concrete finite-size question about how crowd size changes the subsidy requirement.
Theorem 2 is the stronger anchor because it directly supports a fee-minimization problem; Theorem 1 supplies the underlying incentive threshold. The authors could plausibly recognize both mirrors: their mechanism already makes the outcome a fraction of arbiter reports, and their incentive analysis explicitly tracks each report’s \(1/m\) effect. The proposed scope is only these threshold and subsidy results. It does not extend the paper’s conclusions to uniqueness: the authors explicitly leave open the possibility of uninformative equilibria.
The main weakness is that the paper studies a *small* set of distinct, verifiable arbiters. A large interchangeable assessor pool is a change in regime, even though the analysis is parameterized by \(m\) and decentralized platforms motivate crowdsourced reporting. There is also a limit issue: in a truly atomless society, an individual’s report has zero effect on the outcome. The proposed questions therefore retain a finite population scale \(N\) while representing the society by its type masses; their content is the high-multiplicity calibration rate, not a new atomless incentive phenomenon. Finally, these are quantitative mechanism-design results, not formal complexity classifications. If an eligible anchor must explicitly state a standard complexity result, this paper has none.
Theorem 1 and Theorem 2 are the only plausible anchors, but both turn on the same finite-agent effect: an arbiter’s report changes the market payout by an amount proportional to that arbiter’s holdings divided by \(m\). They give incentive and funding bounds, not algorithms or complexity results for a combinatorial problem. A continuous-population mirror therefore needs to supply a new computational question.
For Theorem 1, the proposed finite-scale version does not get much from the type masses. It substitutes \(m=N\rho\) and \(B=\max_t B_t\) into the paper’s existing bound, then scans the support for that maximum. In an atomless population with fixed per-arbiter budgets, an individual changes the reported outcome by zero. The manipulation term disappears, so truthful reporting is already a weak best response at \(k=0\); the threshold the theorem was designed to quantify degenerates. Keeping a finite \(N\) preserves the \(1/(N\rho)\) effect, but then the question is the paper’s finite-agent calculation expressed in population proportions.
Theorem 2 has the same problem, and its proposed fee question is already solved by the paper’s inequality. The stated \(f_N^*\) is a scalar closed-form calculation. If \(M_N\) is derived from type activity, that adds a sum over the support; if \(M_N=N\bar M\), the proposed \(N^{-1/2}\) scaling follows directly from the formula. That may be useful for platform calibration, but it does not create a continuous-computational problem. The paper already performs this kind of parameter calibration in its LMSR section.
A richer type model could assign different budgets or update strengths to different arbiter types. With the paper’s uniform payment rule, the robust condition still reduces to the most demanding type; with type-specific payments, the natural objective is a type-weighted sum subject to per-type incentive bounds. Those are plausible extensions, but the first is a maximum over the support and the second is separable or a small optimization problem. More elaborate peer assignment or coalition deviations might produce a harder problem, but they would change what is being analyzed: Theorem 1 is about one arbiter’s deviation, and Theorem 2 funds that guarantee. They are not mirrors of those results.
The scenario is also a stretch at the relevant population level. The paper explicitly uses a small set of distinct, verifiable arbiters. A large pool of similar arbiters is imaginable, but making market traders numerous does not help: the bounds depend on the arbiters’ count and stakes. I would not call this an airtight case against every future population model—someone may value the large-pool subsidy curve as mechanism-design work. But on the anchors raised here, the atomless version loses the strategic conflict, while retaining finite influence gives a direct reparameterization of results the paper already computes. That is a weak fit for ChoCo’s computational programme.
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.