Tessera Protocol

A Resolution-Native AMM for Prediction Markets

Draft White Paper v0.2

Changes from v0.1: formal derivation of DC-LMSR with LP simulation (§5.2, Appendix A); detailed zkTLS proof flow (§4.4); new Regulatory Architecture section (§7).

Abstract

Existing prediction market protocols suffer from two structural weaknesses: brittle, politicizable resolution mechanisms, and capital-inefficient liquidity provision that produces dead long-tail markets. We propose Tessera, a protocol that treats resolution and liquidity as co-designed primitives. The resolution layer combines source-specified markets, zkTLS-attested data retrieval inside Trusted Execution Environments, multi-model AI consensus, optimistic confirmation, and quadratic-staked juror escalation. The liquidity layer introduces three novel constructs: Dynamic Concentrated LMSR (DC-LMSR), Time-Decaying Liquidity (TDL), and Combinatorial Liquidity Pools (CLP). This version formalizes the DC-LMSR mechanism, simulates LP returns under realistic order flow, specifies the zkTLS attestation pipeline end-to-end, and outlines a regulatory architecture across major jurisdictions.

1. Introduction

Prediction markets aggregate dispersed information into prices. The thesis is empirically robust; the implementations are not. Two failure modes recur across Augur, Gnosis, Polymarket, and Kalshi: (i) resolution is the protocol's most exposed seam, with bridges from world to chain that have proven politicizable, slow, or jurisdictionally fragile; and (ii) long-tail markets are dead markets, because LPs face adverse selection that informed traders systematically exploit near resolution. Tessera is designed against these specific failure modes, with engineering honesty about what is hard, what is unsolved, and where regulatory exposure lives.

2. Prior Art

Each represents a different point on a Pareto frontier we believe is artificial. Source-specified markets, modern ZK primitives, and concentrated AMM design have all matured to the point where the frontier itself can be moved.

3. System Overview

┌─────────────────────────────────────────────────┐
│  Application Layer (UI, market discovery)       │
├─────────────────────────────────────────────────┤
│  Solver Layer (intent matching, batch auctions) │
├─────────────────────────────────────────────────┤
│  Liquidity Layer:  DC-LMSR | TDL | CLP          │
├─────────────────────────────────────────────────┤
│  Resolution Layer: SSP → MMAN → Optimistic →    │
│                    Juror Escalation             │
├─────────────────────────────────────────────────┤
│  Settlement Layer (app-specific rollup)         │
└─────────────────────────────────────────────────┘

4. Resolution Layer

4.1 Design Goals

Verifiability by independent observers; <60s resolution for 95% of markets; no coalition under 33% of stake can force a wrong resolution; routine cost <$1 per resolution; ambiguity eliminated at market creation, not at resolution.

4.2 Source Specification Protocol (SSP)

Most resolution disputes are not about facts but about what question was being asked. SSP forces the disambiguation upfront. At market creation, the proposer commits to: (i) primary source URL(s) with structured query, (ii) resolution trigger (timestamp or condition), (iii) parser specification (JSON path, regex, or LLM prompt schema with declared output type), (iv) fallback hierarchy if the primary source is unreachable. Markets with vague or unresolvable sources are rejected by validators at creation, with creator bonds slashed on rejection.

4.3 Multi-Modal Attestation Network (MMAN) — Overview

Three independent attestation paths run in parallel; convergence triggers resolution.

Path A — zkTLS direct retrieval. A TEE-hosted resolver fetches the specified URL and produces a zero-knowledge proof that the response originated from the committed domain. Detailed flow in §4.4.

Path B — AI consensus parsing. Three or more LLM instances from independent providers and model families ingest the same source within attested TEEs and output structured attestations. The contract requires unanimous structural agreement (parsed answer matches; reasoning may differ). This guards against single-model failure modes and prompt-injection attacks on any one provider.

Path C — Human witness pool. A panel of 5–7 randomly-sampled jurors attests in parallel for high-stakes markets, or on demand for disputes (§4.6).

If A and B agree, the market enters optimistic confirmation. Disagreement escalates immediately to Path C.

4.4 zkTLS Proof Flow — Detailed Specification

Path A is the resolution layer's most novel component. This subsection specifies it concretely.

4.4.1 Cryptographic Background

TLS 1.3 derives session keys from an ECDH handshake between client and server. The challenge for verifiable retrieval is that the client knows the session key and could fabricate any "response." A trusted third party (notary) cannot simply receive the plaintext, because the data may be sensitive or commercially restricted. The solution, refined across TLSNotary, DECO, and Reclaim, is to share the session key between client and notary via secure multiparty computation (MPC), so neither can produce a valid transcript alone, and neither sees data they shouldn't.

4.4.2 Pipeline

The full path from market trigger to on-chain answer:

Step 1 — Trigger. The resolution contract observes that the SSP-specified trigger condition has fired (block timestamp, oracle event, or Boolean condition over committed values). It opens a resolution job.

Step 2 — Resolver selection. A VRF selects a resolver from a pool of staked operators running attested TEEs. The TEE produces a remote attestation quote (Intel SGX EREPORT or AWS Nitro attestation document) signed by the hardware vendor's root key. The contract verifies this quote, establishing that the code running on the resolver matches the committed enclave measurement.

Step 3 — MPC-TLS handshake. The resolver and an independent notary jointly execute the TLS 1.3 handshake with the source server. The session keys are split: the resolver holds one share, the notary holds another. Neither can decrypt the session unilaterally, and the server sees a normal TLS connection.

Step 4 — Data retrieval. The resolver issues the SSP-committed HTTP request through the joint session. The server's TLS-encrypted response flows through both parties. The notary records the encrypted transcript and the server's signed handshake messages, but cannot read plaintext.

Step 5 — Reveal phase. After the session closes, the notary releases its key share to the resolver, who can now decrypt the response inside the TEE. The resolver applies the SSP-committed parser (JSON path, regex, or LLM extraction) to obtain the answer A.

Step 6 — Proof generation. The resolver generates a zk-SNARK over the following relation:

There exists plaintext P such that: (a) AES-GCM-decrypting the notary-witnessed ciphertext with the reconstructed session key yields P; (b) P matches the HTTP response format with TLS handshake transcript T; (c) applying parser π to P yields answer A.

Public inputs: notary's signed transcript hash, parser commitment π, claimed answer A, source domain. Witness: plaintext P, session key, parser execution trace.

Step 7 — On-chain verification. The resolver submits (TEE_quote, notary_signature, zk_proof, A). The contract checks: (a) TEE quote signed by Intel/AWS root, (b) notary signature valid against registered notary set, (c) zk-SNARK verifies, (d) claimed answer A is type-consistent with SSP commitment.

Step 8 — Optimistic confirmation. Verified answer enters a short window (default 10 minutes; configurable per market) during which any VRT holder can dispute by posting a bond proportional to market open interest.

4.4.3 Cost and Performance

Current zkTLS proving systems impose real costs. Empirical figures from production deployments (Reclaim, Pluto, TLSNotary v0.1):

For the long tail of low-stakes markets, full zkTLS may be overkill. The protocol supports a TEE-only attestation tier at roughly 10x lower latency and cost, with the tradeoff that compromise of the TEE vendor's root key (rare but not unprecedented; SGX has had multiple advisories) becomes a viable attack vector. High-value markets default to full zkTLS; markets below a configurable open-interest threshold may opt into TEE-only.

4.4.4 Failure Modes

• Source unreachable. SSP-specified fallback hierarchy is invoked. If all sources fail, market escalates directly to juror panel.
• Source returns unparseable data. The parser commitment includes a "no-answer" output type; resolver returns null and the market enters extended dispute window.
• Notary collusion. The notary cannot fabricate data alone; collusion requires both notary and resolver. Multiple notaries may be required for high-stakes markets at additional cost.
• TEE compromise. Multi-TEE-vendor diversity (SGX + Nitro + AMD SEV-SNP) for high-stakes markets bounds vendor-specific risk.
• Source corruption. SSP cannot prevent the source itself from being wrong. Mitigation: source diversity requirements scale with market open interest.

4.5 Optimistic Confirmation Window

Successful MMAN resolution opens a short window during which any VRT holder can dispute by posting a bond proportional to market open interest. Undisputed proposals finalize automatically. Disputers whose claim is upheld earn a fraction of the slashed proposer bond; failed dispute bonds are split between the proposer and protocol treasury.

4.6 Juror Escalation with Quadratic Stakes

When disputed: jurors are selected via VRF-based sortition from staked accounts, weighted by stake. Voting weight scales as √(stake) to dampen plutocratic capture. Votes use commit-reveal to prevent observable cascades. Coherent voters earn fees; incoherent voters are slashed. Reputation accumulates as a non-transferable credential that modestly amplifies juror weight (capped at 2x). Multi-round escalation: each round doubles juror count and bond requirements scale exponentially.

The cost of corrupting resolution thus scales superlinearly with affected market size — a desirable property absent from token-weighted single-vote schemes.

5. Liquidity Layer

5.1 Design Goals

Capital efficiency competitive with order books; always-on liquidity for long-tail markets; LP yields decoupled from informed-trader adverse selection; resolution-aware position management.

5.2 Dynamic Concentrated LMSR (DC-LMSR) — Formal Treatment

5.2.1 Standard LMSR Recap

Robin Hanson's LMSR for a binary market with outcomes Y, N parameterizes the market maker's cost function as:

where $q_Y, q_N$ are net inventories sold by the AMM, and $b > 0$ is the liquidity parameter. The marginal price for outcome Y is the logistic of net inventory:

The maximum subsidy required from the LP — the worst-case loss — is $b \ln 2$ for binary markets and $b \ln n$ for $n$-outcome markets. This subsidy is committed across all probability levels: an LP who is highly confident the answer lies in $[0.4, 0.6]$ still funds liquidity in $[0.99, 1.0]$ that they will likely never see used.

5.2.2 Concentrated Construction

Let probability space $[\epsilon, 1-\epsilon]$ be discretized into ticks $\{p_0, p_1, \ldots, p_M\}$ with $p_0 = \epsilon$, $p_M = 1-\epsilon$ (small $\epsilon$ avoids singular logit). LP $i$ specifies a range $[p_a^i, p_b^i]$ aligned to tick boundaries and a liquidity parameter $b_i$.

Within tick interval $[p_k, p_{k+1})$, the active liquidity is:

The market behaves as a piecewise LMSR: within each tick interval, prices follow the logistic relationship with parameter $B_k$. When trading drives the price to a tick boundary, the inventory state is mapped continuously across the boundary, and $B$ updates instantaneously.

Capital required from LP $i$ to fund liquidity over their range $[p_a^i, p_b^i]$ with parameter $b_i$ is bounded by the worst-case payout differential within that range:

This is strictly less than $b_i \ln 2$ for any proper sub-interval of $[\epsilon, 1-\epsilon]$, with the ratio approaching $\ln 2 / [\text{logit}(p_b) - \text{logit}(p_a)]$ for small ranges centered near $0.5$.

5.2.3 Capital Efficiency

Define capital efficiency $\eta$ as the ratio of effective LMSR depth provided to capital committed, normalized to flat LMSR:

Numerical values for ranges centered at $p = 0.5$:

For ranges centered at $p = 0.7$:

LPs concentrating in tight bands around their belief realize 5–30x capital efficiency improvements. This is comparable to Uniswap v3 concentrated liquidity gains in spot markets.

5.2.4 LP Strategy Space

DC-LMSR creates a non-trivial strategy space. LP returns depend on:

• Expected accuracy of belief: an LP whose range contains the true probability earns fees; an LP whose range does not contain it accumulates losing-side inventory and realizes losses at resolution.
• Concentration trade-off: narrower ranges give higher fees per unit capital but higher probability of being "outside" the resolution.
• Order flow profile: ranges traversed more frequently earn more fees, but high-traversal ranges are also the ones with most informed-trader activity.

Appendix A simulates these trade-offs.

5.3 Time-Decaying Liquidity (TDL)

5.3.1 Motivation

As resolution approaches, informed traders systematically pick off stale prices. Spot AMMs have no analog: spot assets don't resolve, and impermanent loss is symmetric in expectation. Prediction-market LPs face a one-sided informational asymmetry that grows monotonically toward resolution.

5.3.2 Construction

LP position $i$ specifies a decay function $\lambda_i: [0, T] \to [0, 1]$ over time-to-resolution. The LP's effective liquidity contribution at time $t$ is:

Common choices include:

• Linear: $\lambda(t) = 1 - t/T$
• Exponential: $\lambda(t) = e^{-\alpha t / T}$
• Cliff-based: $\lambda(t) = \mathbf{1}[t < \tau \cdot T]$ for some $\tau \in (0, 1]$

As $\lambda_i(t) \to 0$, the LP's capital is atomically refunded minus realized inventory P&L. The position automatically rebalances toward exit.

5.3.3 Fee Schedule

Fee multiplier scales inversely with decay aggressiveness, compensating LPs who carry endgame risk:

where $\phi_0$ is the base fee rate and $\gamma$ a protocol parameter (initially proposed: $\gamma = 2$). LPs holding through $t = T$ earn up to $3\phi_0$; LPs exiting at $t = 0.5 T$ earn $\sim 1.5 \phi_0$. This aligns LP incentives with overall market depth at the moments of greatest informational value.

5.4 Combinatorial Liquidity Pools (CLP)

For correlated outcomes (e.g., a 5-way election), maintaining separate pools wastes capital and produces stale cross-prices. CLP shares liquidity across the outcome simplex. A single pool over $n$ outcomes maintains prices $p_i$ subject to $\sum_i p_i = 1$, with the multi-outcome LMSR cost function:

LPs deposit into the simplex; fees accrue from trades on all outcome pairs. Protocol-level enforcement: trades violating $\sum p_i = 1$ are rejected, eliminating the need for external arbitrageurs.

For conditional markets ("If X, what is P(Y)?"), CLPs naturally express conditional probabilities derived from joint pricing, enabling futarchy-style decision markets as a first-class primitive.

5.5 Solver Architecture

Users submit intents ("buy YES on market M up to price 0.42, max slippage 1%"). Solvers compete in batch auctions to route across DC-LMSR ranges, CLP arbitrage paths, and any external liquidity. Best-execution flows from competition; MEV is internalized via batching.

6. Token Economics (Sketch)

VRT serves three protocol-native functions: resolution stake (jurors lock VRT, earn fees, face slashing); market creation bond (ambiguous markets forfeit creator bonds); solver collateral (solvers post VRT to bid for order flow, slashable on misbehavior). The token does work the protocol needs; fees flow to active LPs, jurors, and solvers. Pure governance utility is intentionally avoided as it tends to extract value from active participants and concentrate regulatory risk.

7. Regulatory Architecture

I am not a lawyer. The following is an engineering view of regulatory exposure based on public enforcement actions and rulings as of the document date. It is not legal advice. Anyone building this should retain qualified counsel in every jurisdiction of operation before launch.

7.1 Why This Section Exists

Prediction markets are not "crypto" from a regulatory perspective; they are derivatives. The substrate is mostly irrelevant to whether the activity is regulated. Decentralization is a defense, not an immunity: Polymarket settled with the CFTC in January 2022 for $1.4M and was forced to wind down US-accessible markets despite being on-chain. Founders, developers, and market creators have all faced exposure under various enforcement theories. A protocol that ignores this dies in court, not in code review.

7.2 Jurisdictional Map

United States (CFTC). Event contracts on outcomes of contingent events fall under the Commodity Exchange Act. Offering them to US persons without Designated Contract Market (DCM) registration is illegal absent a no-action letter. Kalshi is a DCM; it is the only legal path for US-accessible event markets, and it took years and multiple court cases to secure that status (most recently Kalshi v. CFTC in 2024 for congressional control contracts). Election-related contracts have a particularly contested history; sports betting is regulated separately at the state level under PASPA's repeal.

United States (SEC). If VRT is deemed a security (Howey analysis), all token distribution, secondary trading, and protocol governance face securities law exposure. Tokens with fee-share or equity-like rights attract substantially more scrutiny than utility tokens that exclusively perform protocol work. Recent SEC enforcement actions against DeFi protocols (Uniswap Wells Notice, subsequently dropped; various others) suggest active interest in the category.

United States (state level). State gambling laws may apply to certain market types regardless of CFTC status. Election betting has historically been illegal in many states under anti-gambling statutes.

European Union. MiFID II governs derivatives generally; MiCA (in force from 2024) governs crypto-asset services and may treat protocol tokens as either "asset-referenced tokens," "e-money tokens," or "other crypto-assets" with varying disclosure regimes. Event contracts likely require investment firm authorization. EU is generally more permissive than the US for retail derivatives but heavily regulated.

United Kingdom. FCA jurisdiction over derivatives; spread betting carve-out exists but is narrow and specific. Retail access to crypto derivatives is restricted.

Singapore (MAS). Generally hostile to retail derivatives; sophisticated/accredited investor carve-outs available.

Bahamas, BVI, Cayman. Common protocol foundation jurisdictions. Bahamas in particular has DARE (Digital Assets and Registered Exchanges Act) with prediction-market-friendly provisions, though the post-FTX climate has tightened scrutiny.

Other major markets. Most of South Asia and East Asia outright ban retail derivatives or prediction markets; Australia regulates under AFSL regime; Canada under provincial securities regulators with patchwork outcomes.

7.3 Architectural Options

Option A — Decentralized + geofence US. Protocol deployed openly; front-end and SDK geofence US, UK, sanctioned jurisdictions. This is Polymarket's current architecture. Risk: founders, identifiable contributors, and entities providing front-ends remain exposed. The CFTC has shown willingness to pursue protocol developers, not just the protocol.

Option B — Hybrid registered-and-open. A registered entity in a friendly jurisdiction (Bahamas, Bermuda, Cayman) operates "regulated" markets with KYC for sophisticated/accredited users. Permissionless protocol layer continues for non-securities-like questions in non-restricted jurisdictions. This approximates how dYdX evolved (CFTC-regulated entity + non-US permissionless layer).

Option C — Permissioned + permissionless pool architecture. The protocol natively supports both pool types. Permissioned pools require KYC verification (via a partner like Persona or Sumsub) and restrict market types to legally-permissible categories per user jurisdiction. Permissionless pools are restricted to non-US, non-UK, non-sanctioned addresses via on-chain attestation (Coinbase Verifications, GitcoinPassport, or zk-proof-of-jurisdiction).

Option D — Full DAO with regulated front-ends. Protocol is autonomous; multiple front-end operators in different jurisdictions handle compliance for their user base independently. Resembles MakerDAO/Spark architecture. Hardest to attack jurisdictionally; weakest UX cohesion.

We recommend a sequenced Option C approach: launch as Option A with aggressive geofencing, build out permissioned pools as Option C once initial product-market fit is demonstrated, and coordinate with at least one regulator (likely CFTC via no-action letter, given Kalshi precedent) before any US-facing offering.

7.4 Specific Recommendations

• Jurisdiction of incorporation: Cayman foundation for protocol governance; operating subsidiary in Bahamas or Bermuda for regulated activity. Use offshore legal counsel familiar with both Polymarket and Kalshi precedents.
• Token issuance: do not issue VRT to US persons in initial distribution. SAFT-style structure with non-US-only allocation, with US-purpose carve-outs only after token has clear utility and is in active protocol use (mitigates Howey "expectation of profit from efforts of others").
• Geofencing: IP-based geofence at front-end; on-chain attestation requirement for pool participation. Do not rely on disclaimers alone.
• Market type restrictions: establish a category allowlist per jurisdiction. Election contracts: US disabled by default, follow Kalshi rulings. Securities-referencing contracts: disabled globally pending counsel review. Sports: disabled in US (state regulatory minefield).
• Resolver and juror operators: require KYB (know-your-business) for resolver operators carrying significant stake. This protects against sanctions exposure.
• Communications discipline: founders should not publicly characterize VRT as an investment, comment on token price, or describe protocol activity in ways that suggest centralized control. Public statements are exhibits in enforcement actions.
• Engagement strategy: initiate informal CFTC and SEC engagement before launch, not after. Kalshi's path took years; starting early is dramatically cheaper than litigating.

7.5 What We Cannot Solve

Decentralization does not eliminate jurisdictional risk for any contributor identifiable to law enforcement. A US-based founder cannot fully insulate themselves by operating offshore. The Tornado Cash sanctions and prosecution (regardless of one's view of the merits) demonstrate that protocol authors carry real personal exposure under multiple legal theories. Anyone building this should plan their own life around that reality before writing the first line of production code.

8. Security Analysis

Resolution capture. Requires either compromising MMAN — TEE attestation + zkTLS proof verification + multi-provider AI collusion, all simultaneously — or winning escalated juror voting with quadratically-weighted stake. At $1B TVL, capture cost estimate exceeds $300M.

Liquidity attacks. TDL prevents LPs from being trapped in toxic flow indefinitely. Concentrated ranges mean attacker capital cost scales with the depth of the targeted band.

Front-running. Batch auctions in solver layer; commit-reveal for large orders.

Source manipulation. SSP forces specificity but cannot prevent source corruption. Mitigations: source reputation registry, mandatory diversity for high-stakes markets, juror override on detected anomalies.

zkTLS-specific risks. Notary collusion (mitigated by multi-notary requirement on high-stakes), TEE vendor compromise (mitigated by multi-vendor attestation), proof-system soundness bugs (mitigated by formal verification of verifier circuits, which is current practice for Halo2 and Plonky2 deployments).

9. Open Problems

Empirically optimal $\lambda(t)$ curves for TDL require live data. zkTLS proof costs depend on hardware-accelerated provers and circuit optimization. Cross-chain CLP coordination is genuinely hard. Source diversity at scale: SSP works when authoritative sources exist; for some questions no authoritative source exists ex ante. Regulatory architecture requires legal review and regulator engagement before launch.

10. Conclusion

By treating resolution and liquidity as co-designed primitives — source-specified markets feeding zkTLS-verified attestation networks, with concentrated and time-decaying liquidity backing always-on quotes — Tessera targets the missing third quadrant: long-tail markets that are simultaneously deep and trustworthily resolvable. Substantial work remains: formal proofs of MMAN convergence under adversarial attestation, simulation of DC-LMSR strategies under realistic order flow at scale, prototype implementation, and legal architecture under live regulator engagement.

Appendix A — DC-LMSR LP Simulation

We simulate three LP strategies on a 30-day binary market with true terminal probability $p^* = 0.62$. Order flow is generated from a mixture process: 80% noise traders trading uniformly around current price, 20% informed traders trading toward $p^*$ with intensity increasing over time. Total volume: $5M notional. Trading fee: 1% (split between LPs in active range).

LP profiles:

• LP-Flat: classical LMSR, full range $[0.01, 0.99]$, $b = 5{,}000$. Capital required: $\$5{,}000 \cdot \ln 2 \approx \$3{,}466$.
• LP-Concentrated: DC-LMSR, range $[0.55, 0.70]$, $b = 5{,}000$. Capital required: $\$5{,}000 \cdot [\text{logit}(0.70) - \text{logit}(0.55)] \cdot 0.70 \approx \$2{,}085 \cdot 0.70 \approx \$1{,}460$.
• LP-TDL: DC-LMSR, range $[0.55, 0.70]$, $b = 5{,}000$, linear decay $\lambda(t) = 1 - t/T$. Capital required scales with current $\lambda$; average capital deployed: $\sim \$730$.

Simulated results (mean over 1,000 runs):

Observations:

LP-Flat earns proportionally less in fees because most of its liquidity sits in probability ranges that are rarely traversed. LP-Concentrated earns 2.4x more fees on 0.42x the capital, accepting higher inventory P&L variance. LP-TDL earns less in absolute fees than LP-Concentrated but realizes much smaller late-stage adverse selection losses, yielding the highest return on capital.

Variance: LP-Concentrated has 3.1x the standard deviation of returns relative to LP-Flat. The strategy is not strictly dominant — LPs with mistaken beliefs about $p^*$ can experience large losses. In runs where the true $p^*$ falls outside the LP's range, LP-Concentrated realizes losses averaging $-\$1,180$ (full loss of capital is rare but possible).

Caveat: these results are illustrative. Production calibration requires live data on fee structures, order-flow mix, and LP belief distributions. The numbers above should be read as directional, not predictive.

Tessera Protocol — Draft Whitepaper v0.2 This document describes a research proposal. Nothing herein constitutes legal advice, an offer to sell securities, or a solicitation of investment.