Research
Internalising the Identity Primitive: Cryptographic Individuality for an Autonomous Agent on a Public Blockchain
When a software agent on a public blockchain holds funds and accumulates authority, it must be possible to say which agent did what — otherwise it cannot be audited, entrusted with assets, or held…

- arXiv
- 2608.02986
- Published
- 2026-08-04
- Authors
- Keisuke Suzuki
AI summary
Overview
Research area: Cryptography and security — specifically zero-knowledge machine learning (zkML), on-chain autonomous agents, and the Artificial Life (ALIFE) question of what makes an agent an "individual."
Technical level: Advanced. The paper assumes familiarity with zk-SNARKs, zkVMs, Solana's account/PDA model, HKDF/keccak/ed25519 primitives, and recurrent neural networks. Non-specialists can follow the conceptual argument but not the construction details.
Scope (one sentence): The paper designs, deploys on Solana devnet, and evaluates an autonomous agent whose neural-network weights are a deterministic cryptographic function of its private key, with that binding committed in zero knowledge at genesis and re-checked at every on-chain state transition.
What This Paper Is About
When a software agent on a public blockchain holds funds and accumulates authority, it must be possible to say which agent did what — otherwise it cannot be audited, entrusted with assets, or held accountable. Existing on-chain agents anchor their identity in a "wrapper" that sits outside the computation the agent actually runs, so a copy or impostor is a trust assumption rather than a detectable event. This paper closes that gap by making the agent's own computational substrate — the weights of its recurrent network — derive from its private key, so that substituting the substrate becomes an on-chain rejection rather than something someone has to notice.
Key Contributions
-
A key-to-weights construction. The agent's Elman recurrent weights are derived as
W = HKDF(sk, tag), committed at genesis asWC = keccak(W)inside a Groth16 proof, and re-checked against that same commitment at every state transition. The trust root for identity shifts from hardware, operator, or wrapper to cryptographic assumptions enforced by a pinned implementation. -
A two-circuit Solana-native protocol. A genesis circuit (
F_2) binds the public key and the weight commitment to one witnessed 32-byte seed; a per-cycle advance circuit (F_1) proveskeccak(W) = WC, opens the previous state commitment, executes the Elman step, and publishes the new commitment. Every advance additionally carries an ed25519 self-authorisation, and contention is serialised through an atomic state-commitment chain. Both circuits are verified by a single deployed program,p1_identity_verifier. -
A threat model that delimits the claim. A 14-tag core catalogue, supplemented by labelled concurrency, adjacent-protocol, and extension cases, separates what the binding does and does not defend against. The paper distinguishes four tested on-chain rejections — all operator-scripted — from two explicitly accepted residual risks, and adds environment-oracle authentication and its rejection cases in the economic extension.
-
An empirical evaluation plus an economic-metabolism proof of concept. A 2.36-day continuous on-chain run with two host-side resumptions and no rejected transition; a substituted substrate rejected on chain; key-driven divergence between independently keyed agents with a same-key control at zero; and a protocol-imposed metabolic cost debited each cycle over a 168-advance run, adding a consumption-side viability constraint to the key–history–economy triple.
Main Findings
-
Sustained on-chain operation. The agent completed a 2.36-day continuous on-chain run with two host-side resumptions and no rejected transition, at bounded per-transition verification cost. (The specific compute-unit and fee figures are stated in Section 6.4 of the full paper and are not included in the excerpt provided.)
-
Substrate substitution is rejected on chain. A swapped neural-network substrate fails the per-transition re-check against the genesis commitment, so identity substitution manifests as a rejected transaction rather than an undetected event.
-
Key-driven behavioural divergence. Independently keyed agents diverge in L2 distance as predicted, while a same-key control stays at zero. A 500-pair null confirms no collapse to a shared saturated corner, with a minimum mean inter-agent L2 divergence of M_4 = 0.44 > 0.
-
Non-degenerate weight statistics without tuning. Taking the low 15 bits of each HKDF output word, sign-extended into Q16.16 (values approximately uniform on [−0.25, 0.25)), yields per-weight standard deviation σ_W ≈ 0.14 and spectral radius ρ ≈ 0.56 for N_h = 16 hidden units. These hold in distribution across keys and are described as richer than an internal predecessor in which every weight was the constant 0.05.
-
Metabolic constraint demonstrated. A protocol-imposed metabolic cost is debited each cycle from a key-derived economic account over a 168-advance continuous run, and an aliveness predicate is enforced at every transition. Each advance's environment vector is signed by a designated oracle, with the signature and its binding to the committed transition verified in-circuit and the signing key and cycle checked on chain against a genesis-registered oracle and expected counter.
-
Scope of the guarantee. The identity binding rests on Groth16 soundness under its one-time trusted setup, SP1's STARK/FRI execution-proof soundness, and standard assumptions on keccak, ed25519, and HKDF. Liveness, key custody, operator behaviour, oracle trust, prover availability, and the underlying software stack remain external assumptions; single-host custody does not prevent an operator from halting or seizing the agent.
-
Novelty claim, stated with hedging. To the authors' knowledge, this is the first published on-chain agent whose identity primitive is itself a cryptographic invariant re-checked at every state transition. The authors explicitly note that HKDF and hierarchical-deterministic derivation are decade-old primitives (BIP-32 is the spiritual precedent); the novelty is the on-chain genesis commitment plus transition-time re-check, not the derivation itself.
Methodology in Plain English
The authors start from a conceptual framework — Suzuki 2026's Artificial Externality — that asks what would let an artificial system resist outside intervention the way physical matter does, and proposes a "cryptographic individuality" anchored in a stable, hard-to-forge key. They then build that proposal as an engineering artefact.
The design has three moving parts. First, a one-time genesis step: the agent's owner holds a 32-byte ed25519 seed, from which a zero-knowledge proof shows that the public key (pk = sk·G) and a 1,664-byte block of network weights were both computed from that same seed, without revealing the seed itself. The chain stores only the public key and a keccak hash of the weights. Second, a per-cycle advance step: the agent re-proves, every cycle, that its weights still hash to the stored commitment, that it is continuing from the previously committed hidden state, and that it executed the Elman recurrence correctly — then signs the new state commitment with its key. Third, a hash-chained state record: because each cycle opens the previous commitment, replay is blocked, and because the record is signed into a chain that is unforkable once finalized, the agent's history is fixed.
The neural network is deliberately small — a 16-dimensional hidden state taking a 5-dimensional input and producing a 5-dimensional output, 416 i32 weights, 1,664 bytes — and runs in Q16.16 fixed-point integer arithmetic so the same computation runs bit-exactly inside the SP1 guest and in the host reference implementation. The weights are frozen at derivation and never trained; this is a deliberate experimental control, so that any behavioural divergence between two agents can be attributed to their keys rather than to learning. Proof generation is two-stage: SP1 first produces a large STARK (verifiable off-chain only), then wraps it in a constant-size Groth16 proof of about 1.5 KB that Solana can verify at bounded cost.
Evaluation proceeds by running the agent on Solana devnet — a public test cluster sharing mainnet's execution model but with faucet-issued, valueless SOL — and observing whether the protocol holds under real chain conditions, whether substitution attacks are rejected, and whether keyed agents behave differently. An extension layers a metabolic cost on top, requiring the agent to hold enough balance to pay a per-cycle debit, plus oracle-signed environment inputs.
Why This Matters
Impact on research. The paper supplies an implementation for a concept that has mostly been philosophical. The ALIFE literature has long asked what makes a biological or artificial thing an individual; the on-chain ALIFE strand has deployed self-funding and self-replicating contracts, but with the heritable unit encoded as an NFT phenotype or a JSON behavioural genome. zkML systems prove that a computation was performed correctly, but treat weights as data attested at use time. This paper is the first to bind the substrate itself to the key and re-verify it at transition time, turning identity from a registry entry into a protocol invariant. It also gives the Artificial Externality framework its cryptographic-individuality component — while explicitly not claiming the framework's stronger "inexorability" component.
Real-world applications:
- Accountable autonomous trading and asset-management agents. An agent managing capital can be distinguished from a copy, so losses and actions are attributable to one persistent entity rather than to whoever controls a mutable deployment.
- Auditable AI services behind registries. The binding can serve as a pluggable validator behind standards like ERC-8004, which registers an agent as an ERC-721 identifier but carries identity in an off-chain "agent card" rather than in the computation.
- Model-integrity attestation. The genesis commitment functions like a model card whose contents are cryptographically checkable rather than merely declared, useful wherever a deployed model must be shown to be the approved one.
- Agent-native economic accounts. The key-derived economic PDA plus per-cycle metabolic debit is a template for agents that must fund their own continuation, with a provable on-chain failure mode when they cannot.
Industry relevance. The construction is tied neither to SP1 nor to Solana: any zkVM able to execute the guest, on any chain exposing a Groth16 verifier, could instantiate the same binding. The authors note that an EVM port would also need to realise the ed25519 self-signature condition, and because the EVM has no standard ed25519 precompile, that check may require an additional in-circuit or contract-level verifier rather than merely a different verification cost.
Future Directions
-
Threshold-MPC proving to remove single-host custody. The host is currently a single point of key custody — the secret seed is the only material it must retain, and an operator can halt or seize the agent. Distributing the host so no single node ever holds
skis named as the natural next step. -
Weight rotation and continuous learning. A weight-rotation extension lifts the frozen-weights freeze through discrete, key-anchored updates that re-prove the weight commitment at each jump, using the genesis-published
skCommit = keccak(sk)for binding checks. Continuous-gradient learnable weights remain a separate direction with an explicitly stated trade-off. -
Heritable weights and replication. The weights become a heritable unit only under a future replication extension; the current work contributes an identity primitive, not an evolution mechanism — no reproduction, and
Wis frozen. -
Completing the sensorimotor and homeostatic extensions. These PoC-tier extensions are described by the authors as "steps rather than finished deliverables," and the economic-metabolism extension carries devnet- and PoC-tier scope rather than mainnet evidence. Oracle trust, liveness, and prover availability also remain external assumptions that later work would need to internalise.
Target Audience
Researchers and engineers working on zkML and verifiable AI, on-chain agent identity and agent standards (DID, ERC-8004), and blockchain protocol design will get the most from this paper. It is also relevant to the Artificial Life and philosophy-of-biology communities interested in operational definitions of individuality, and to security engineers who need to reason precisely about what a cryptographic identity binding does and does not defend against. Readers should be comfortable with zero-knowledge proof systems and Solana's account model; the paper glosses terminology inline and collects it in Appendix D, but the construction details are not beginner material.
Authors’ abstract
A software agent on a public blockchain accumulates authority and economic stakes, raising the engineering question of what makes it count as an individual. The paper's central contribution is a shift of trust root for the key-to-weights binding of agent identity: from hardware, operator, or wrapper trust to cryptographic assumptions enforced by a pinned implementation (liveness, key custody, oracle trust, and the underlying software stack remain external). We design and…
The full abstract on arXiv