Fortytwo: Decentralized AI Inference

• Fortytwo is a decentralized AI inference protocol that coordinates AI models via peer-ranked pairwise evaluation and reputation-weighted consensus.
• The protocol employs a modular architecture combining candidate generation, auxiliary processing, ranking engines, and encrypted communication for auditability.
• Empirical evaluations show Fortytwo achieves higher accuracy and robustness against adversarial inputs compared to centralized monolithic systems.

Fortytwo is a horizontally scalable, decentralized inference protocol that coordinates a heterogeneous swarm of AI models to collectively generate and peer-evaluate answers to complex queries. Drawing on biological swarm intelligence, tournament theory, and cryptographic systems, Fortytwo combines peer-ranked pairwise evaluation, reputation-weighted consensus, on-chain auditability, and Sybil resistance via proof-of-capability to deliver robust, high-accuracy, and attack-resistant inference services. Its architecture and protocol provide a foundation for democratizing access to reliable, high-quality AI inference in settings where centralized solutions face compute, trust, and governance limitations (Larin et al., 27 Oct 2025).

1. Motivation and Architecture

Fortytwo addresses critical limitations of centralized AI inference, including compute ceilings, single-point failures, oligopolistic model control, and diminishing performance returns from ever-larger monolithic models. Instead, it orchestrates a decentralized swarm of AI nodes, each performing dual roles as both generators (producing answer candidates) and judges (evaluating others' responses through detailed pairwise ranking).

The architecture comprises four principal modules:

• Primary Cognitive Module: An LLM or expert system responsible for candidate answer generation.
• Auxiliary Processing Unit: Handles pre/post-processing, tool usage, and caching.
• Ranking Engine: Samples pairwise comparisons, generates multi-token explanations, and aggregates votes using a Bradley–Terry optimizer.
• Communication Module: Employs gossip protocols, encrypted messaging, and on-chain coordination to manage query propagation, result distribution, and audit trails.

This structure enables horizontal scaling, heterogeneity of expertise, and systematic cross-evaluation, while embedding cryptoeconomic incentives and security primitives at the protocol layer (Larin et al., 27 Oct 2025).

2. Swarm Inference Workflow and Consensus

The Fortytwo protocol orchestrates the following workflow for each query:

1. Query Dissemination: Incoming queries are broadcast, using semantic-topology routing, to a selected sub-mesh of $N$ nodes.
2. Candidate Generation: Each node independently generates a proposed answer.
3. Pairwise Sampling: Upon receipt of all $N$ answers, nodes sample up to $3N$ random, cryptographically seeded pairwise comparisons (excluding their own).
4. Judgment and Reasoning Chains: For each pair $(i, j)$, nodes issue a 50–100-token reasoning justification and a binary vote $y_{ij} \in \{0, 1\}$ indicating preference.
5. On-Chain Commitment: All votes and reasoning chains are committed either on-chain or within an off-chain DAG, ensuring auditability.
6. Aggregation and Winner Selection: A reputation-weighted Bradley–Terry aggregation converts the set of pairwise preferences $\{y_{ij}^{(a)}\}$ into global quality scores $\pi_i$ for each answer.
7. Outcome and Reputation Update: The answer with highest $\pi_i$ is returned as the consensus result, with reputational scores and incentive distribution automatically updated (Larin et al., 27 Oct 2025).

3. Mathematical Foundations: Reputation-Weighted Bradley–Terry Aggregation

Fortytwo’s consensus mechanism is grounded in the Bradley–Terry model, extended with reputation-aware weighting:

For two items $i, j$ with latent qualities $\pi_i, \pi_j > 0$,

$N$0

Let node $N$1 with reputation $N$2 provide votes $N$3 for comparisons in $N$4. The log-likelihood of the observed pairwise data with reputation weighting is

$N$5

By reparametrizing with $N$6 and applying convex optimization, Fortytwo computes the global ranking scores $N$7. This aggregation, informed by peer reputation, produces a consensus that is less sensitive to outlier or malicious judgments compared to simple majority voting (Larin et al., 27 Oct 2025).

4. Reputation Dynamics and Economic Layer

Each node maintains a two-part reputation:

• Generation Success ($N$8): Fraction of rounds in which the node’s candidate wins consensus.
• Ranking Accuracy ($N$9): Kendall’s $3N$0 correlation between the node’s pairwise votes and the overall final ranking.

Combined reputation is computed as:

$3N$1

with updates using exponential moving averages. Substandard or adversarial behavior triggers “slashing” (loss of locked reputation), and inactivity leads to decay toward zero, requiring re-qualification. Nodes must stake a minimum reputation $3N$2 to participate, aligning economic incentives and filtering low-quality participants (Larin et al., 27 Oct 2025).

5. Proof-of-Capability and Sybil Resistance

To deter Sybil attacks, Fortytwo mandates that new identities demonstrate computational and epistemic capability rather than mere token or proof-of-work stake. Each aspiring node solves a dynamically generated suite of calibration test queries $3N$3, with

$3N$4

where $3N$5 is the per-query inference cost and $3N$6 is a calibrated capability threshold. The dynamic, parameter-varied, and paraphrased test queries preclude memorization, and vetted live user queries continually expand the calibration set. This forces adversaries to expend genuine compute resources for each Sybil attempt, making multi-identity attacks economically unattractive and further enriching overall domain coverage (Larin et al., 27 Oct 2025).

6. Empirical Evaluation and Robustness

Fortytwo was evaluated on six benchmarks, using a 35-node swarm comprising models across the 8–235B parameter range and domain specialists. Key results include:

Benchmark	Fortytwo Accuracy	Majority Vote / Baseline
GPQA Diamond	85.90%	68.69% (+17.21 pp)
LiveCodeBench	84.40%	~65% (best individual)
MATH-500	99.60%	—
AIME 2024	100%	—
AIME 2025	96.66%	—
HLE	24.84%	~23–26% (top models)

Robustness to adversarial prompts and Byzantine behavior was demonstrated: Fortytwo’s performance degrades by only 0.12% (extraneous info) compared to 6.20% for a monolithic single-model, and in injects up to 30% malicious nodes, accuracy remains at ~71% with reputation weighting, vs. ~52% for majority voting (Larin et al., 27 Oct 2025).

7. Security, Deployability, and Implications

Fortytwo extends classical Byzantine Fault Tolerance, tolerating $3N$7 misbehaving nodes through reputation weighting, and detects collusion via mutual-support metrics penalizing suspicious rankings:

$3N$8

On-chain storage of votes and reasoning chains enables forensic auditing. Practical deployment is feasible, with 2–5 s latency overhead suitable for many real-world AI tasks. Economic costs (∼40× single-model inference) are competitive with other verifiable compute systems given gains in reliability and robustness. The protocol further democratizes access, allowing small-scale participants with modest hardware to contribute and build reputation without reliance on large-scale infrastructure.

Together, these mechanisms make Fortytwo a practical blueprint for collective, decentralized AI inference, combining swarm intelligence, cryptoeconomic security, and scalable peer-evaluation to achieve high-accuracy, robust, auditable outcomes in adversarial and open environments (Larin et al., 27 Oct 2025).