strong truthfulness

The proof that the fused mechanism of the cybergraph — ICBS market, Bayesian Truth Serum scoring, and valence privacy — extracts truth, including in the perpetual market where no external oracle resolves anything.

This page settles the two results left open by truth-scoring and market: the joint mechanism and the perpetual game. It states what is proved, the one correction the naive composite needs, the architecture that makes the parts correct, and the residual obligations.

division of labor

The three layers carry different jobs, and the separation is what makes each one correct. The mistake to avoid is asking any single layer to do all three.

The serum is the oracle. The market is a liquidity skin over it. Privacy is the third leg that lets the no-external-oracle case stand.

the market is not a truth-scorer, and must not be asked to be one

ICBS has cost $C(s_Y, s_N) = \lambda\sqrt{s_Y^2 + s_N^2}$, so its spot prices satisfy $p_Y^2 + p_N^2 = \lambda^2$ — a circle of radius $\lambda$, not the probability simplex $p_Y + p_N = 1$. A risk-neutral trader with belief $\theta = P(\text{YES})$ buys YES while $\theta > p_Y$ and NO while $1-\theta > p_N$; the two optimality conditions $p_Y = \theta$ and $p_N = 1-\theta$ are simultaneously reachable only when $\theta^2 + (1-\theta)^2 = \lambda^2$, a measure-zero coincidence. Driving $p_Y$ to $\theta$ forces a reserve ratio (with $\lambda = 1$)

$$q = \frac{\theta}{\theta + \sqrt{1-\theta^2}}, \qquad \theta = 0.5 \;\Rightarrow\; q = \frac{0.5}{0.5 + 0.866} = 0.366 \neq 0.5.$$

The reserve-ratio report is a systematically biased function of belief. ICBS with unit settlement is not a proper scoring rule; the mark-to-market reading turns it into pure price speculation. Both readings agree: the market alone does not track truth.

This is the right behavior for a liquidity substrate. ICBS is kept for its self-scaling TVL, inverse coupling, and $\lambda$-range early-conviction payoff. Properness lives in the serum, where it holds: the meta-prediction's log-score term is strictly proper. The truthfulness guarantee routes through the serum, never through the market.

the separability lemma

The serum score of truth-scoring separates additively across its two reports. Expanding the information-gain term over outcomes $k$:

$$D_{KL}(p_i \| \bar m_{-i}) - D_{KL}(p_i \| \bar p_{-i}) = \sum_k p_i(k)\,\log\frac{\bar p_{-i}(k)}{\bar m_{-i}(k)} \;=\; \langle p_i,\,\ell\rangle, \qquad \ell(k) := \log\frac{\bar p_{-i}(k)}{\bar m_{-i}(k)},$$

linear in the first-order report. The prediction term

$$-\,D_{KL}(\bar p_{-i} \| m_i) = \sum_k \bar p_{-i}(k)\,\log m_i(k) + \text{const}$$

is the strictly proper log scoring rule for the meta-report $m_i$, with expected maximum at the truthful meta-belief. So the serum is linear in the first-order report and strictly concave in the meta-report, and the two never couple.

the correction the naive composite needs

The serum's first-order score is linear, so it cannot by itself pin a unique truthful report — a linear objective is maximized at a simplex vertex, not at the interior posterior. Two distinct settings resolve this, and they require different architectures.

Resolved markets. When a market settles against a real external outcome, the market profit becomes a strictly proper scoring rule (it is then a standard convex-cost maker over the simplex). The strictly-concave market term regularizes the serum's linear first-order term: the composite has a unique maximum at the true posterior. The market carries first-order truthfulness; the serum adds the meta channel. The factorization holds — route first-order signal through the market, score the valence with the serum.

Perpetual markets. With no external resolution the price is marked against its own converged value, which makes it a Keynesian beauty contest: any commonly-expected convergence point is self-fulfilling. The market cannot carry first-order truthfulness. The serum carries it instead, through the surprisingly-popular signal — proved next.

theorem — surprisingly-popular selection (perpetual market)

Claim. In the perpetual market the surprisingly-popular divergence selects the truthful equilibrium out of the beauty-contest manifold, provided reinforcement couples to that divergence rather than to the raw price.

Proof. A false focal point $p'$ is commonly expected by construction — agents coordinate on $p'$ because they predict everyone will report $p'$. So the meta-predictions also point there: $\bar m_{-i} \approx \bar p_{-i} = p'$, giving divergence

$$\Delta := \bar p_{-i} - \bar m_{-i} \approx 0.$$

This is the babbling case of truth-scoring, which scores exactly zero. A belief backed by private evidence is more popular than predicted — the surprisingly-popular effect — giving $\bar p_{-i} > \bar m_{-i}$, hence $\Delta > 0$ in its direction. The divergence is the coordinate that separates truth from mere agreement (Prelec–Seung–McCoy, 2017): popular-and-expected scores zero, popular-beyond-expectation scores positive.

Let reinforcement act on the surprisingly-popular estimate $\hat\theta$ derived from $\Delta$ rather than on the raw price. Then false focal points ($\Delta \approx 0$) receive no reinforcement and are not stationary under the dynamics; the truthful answer ($\Delta > 0$) is reinforced and is selected. Under the standard peer-prediction model — common prior, conditionally independent signals, stochastic relevance — $\hat\theta$ equals the truth, so the selected equilibrium is the truthful one. ∎

This is why the second dimension is essential, not redundant: the price can sit at a false focal point; the meta-report reveals whether that point is genuinely supported or merely coordinated. The two-dimensional signal of market is load-bearing in the perpetual case.

The remaining attack is a coordinated inversion — agents agreeing to report against their signal rather than merely coordinating on a focal point. Surprisingly-popular selection does not address this; it reduces to the honest-majority-by-stake condition with the multi-task Correlated Agreement structure of truth-scoring.

privacy completes the selection

The beauty-contest equilibrium is sustained by observability: agents coordinate on $p'$ because they can see the consensus and follow it. ZKP privacy removes that channel.

• A pump cannot recruit followers — no one can see the position to copy it — so it is faded by independent informed flow and reverts. Privacy permits secret accumulation but denies the cascade, and the cascade is where the systemic damage lives.
• Faced with a price it cannot attribute (one whale or a thousand honest agents), a rational agent has no informative social signal and falls back on its own belief, trading toward truth.
• A cartel cannot verify member compliance, so members defect to truth.

Privacy and the serum are complementary anti-consensus-fabrication mechanisms: the serum distinguishes coordinated from true; privacy removes the observability needed to coordinate. Residual single-agent manipulation is bounded by honest-majority-by-stake. Privacy is a defense, not a manipulation vector.

the two settlement modes

The mechanism is a multipurpose oracle: settlement is a free parameter, and the truth source differs by mode.

This is the precise sense in which Bayesian Truth Serum is the fundamental low-level oracle: it works exactly where markets fail, on the unresolvable and perpetual questions. The market is a liquidity skin; the serum is the part that cannot be removed.

the design rules this fixes

• Route truthfulness through the serum, never through ICBS — the market is liquidity and commitment, not a scoring rule.
• Couple reinforcement (rank, effective weight, reward) to the surprisingly-popular estimate $\hat\theta$, not to the raw price. This is what selects truth over coordinated consensus.
• Keep individual positions and reports behind ZKP; publish only aggregates. This removes the coordination that fabricates false consensus.

the Regime-B capital bound

Regime B — a cartel holding over half the stake-weight — cannot be defended by scoring, because the cartel is the reference the score is measured against. The defense is economic: the perpetual ICBS market makes holding a false consensus a continuous cost, and the bound states how much honest capital renders that cost prohibitive.

The peg-holding cost. Let the honest-signal-implied probability of edge $e$ be $q_T$ and the cartel's target false read be $q_F$, with lie size $g = q_T - q_F > 0$. In ICBS, parametrizing the state by an angle $\phi$ gives prices $p_Y = \lambda\cos\phi$, $p_N = \lambda\sin\phi$ and reserve-ratio probability $q = \cos^2\phi$. Honest informed capital sees YES underpriced whenever $q < q_T$ and buys toward $q_T$; the cartel must run a counter-flow to hold $q$ at $q_F$. Model the motion as

$$\dot q = \eta_H H\,(q_T - q) - u_C,$$

honest push proportional to deployed honest capital $H$ and responsiveness $\eta_H$, against the cartel's counter-flow $u_C$. Holding the peg ($\dot q = 0$ at $q = q_F$) forces a constant counter-flow $u_C = \eta_H H\, g$. The cartel buys the losing side at that rate against a price gap $g$, so it bleeds at rate

$$\dot L_C = u_C\, g + \delta D\, g = g\,(\eta_H H + \delta D),$$

the counter-flow cost plus the damper tax, where $D$ is market depth (capital to move $q$ by one unit, $D \propto$ TVL $= \lambda\rho$, largest near $q = \tfrac12$ by a $1/\sin 2\phi$ factor).

The bound. Let the cartel extract value at rate $\gamma(f)$ — the reward, rank, or influence it siphons by keeping $e$ mis-rated, increasing in its reach $f$. Both gain and bleed are rates, so viability is time-independent: the holding period cancels, and the attack either bleeds forever with no crossover or profits forever. It is a phase boundary, not a race. The attack is unprofitable exactly when $\gamma(f) \le g\,(\eta_H H + \delta D)$, which solves to the honest-capital threshold

$$H \;\ge\; H^*(f) \;=\; \frac{\gamma(f)}{\eta_H\, g} \;-\; \frac{\delta D}{\eta_H}.$$

Reading it:

• $H^*$ rises with the prize $\gamma(f)$ and falls with the lie size $g$. Big lies are cheap to deter — lucrative to bet against — and near-truth lies would need vast capital but extract almost nothing; the ratio $\gamma/g$ governs.
• the damper does part of the defense for free: if $\delta D \ge \gamma(f)/g$ then $H^* \le 0$ and the damper alone deters the attack. this sets the design target for $\delta$.
• $\lambda$ and $\delta$ are protocol levers ($D \propto \lambda\rho$); $H$, $\eta_H$, $\gamma$, $g$ are environment.

Two channels and antifragility. With the surprisingly-popular coupling and valence privacy, the cartel must corrupt both the price and the meta-aggregate to move $\hat\theta$. Corrupting the meta is the classic stake threshold — capital $\tfrac12 S_{\text{total}}$ to flip the reference; corrupting the price is the market cost above. The cartel must win both races; the honest side deters by holding either:

$$\text{secure} \iff H > H^*(f) \ \ \text{or}\ \ S_{\text{honest}} > \tfrac12 S_{\text{total}}.$$

Above the threshold the dynamics are self-reinforcing: betting against a false peg is profitable, so honest capital grows while the cartel's shrinks, widening the margin. $H^*$ is an unstable separatrix, not a finish line — the antifragility of market made quantitative.

what remains conjectural

• the capital bound is a mean-field model. The price ODE replaces the finite-population stochastic order flow with its drift; a full proof needs concentration (propagation of chaos). The geometric constants in $D$ and $\eta_H$ are order-of-magnitude, and the exact form of $\gamma(f)$ depends on the reward function of rewards.
• the surprisingly-popular selection rests on the standard peer-prediction model (common prior, conditional independence, stochastic relevance). Correlated evidence across agents weakens it; its interaction with the perpetual dynamics is unverified.

see truth-scoring for the static minority case, the babbling lemma, and the Correlated Agreement reduction. see market for the perpetual market and the liquidity damper. see epistemic-markets for the ICBS cost function. see serum for the original Prelec mechanism. see honest majority assumption for the stake condition.