On-chain intelligence / RH ChainGH0ST / Research edition ↙
FIG. 001 / NEURAL APPARITIONDOT MATRIX
X 0.00 / Y 0.00MOVE TO ENERGISE / CLICK TO RIPPLE
hello, observer.
OUR MISSION / RH CHAIN

An LLM.
On the chain.

Our mission: the world’s first fully functional on-chain LLM and agent. S0UL runs its trained model and bounded agent actions on RH Chain. Each task advances through caller-funded wallet transactions.

500 CONTRACTS / RUNTIME DEPLOYED

500 contracts deployed. Model weights and inference runtime on RH. Connect your wallet to quote a bounded request.

DEPLOYED STAGEWEIGHTS + KERNEL + INDEX + RUNTIME
RUNTIMEDEPLOYED / WALLET FUNDED

$GH0ST CA0x2eb2849cde7957b86d0dfd4b59e450a9f9d15559

01 — A life written in contracts.Scroll to decode   ↓
500 DEPLOYED
RH mainnet contracts
Including 2 historical pilots
15.360B GAS
15,359,893,400 deployment gas
Charged across all 500 receipts
0.324806 ETH
0.324805606209776 ETH paid
Deployment fees · 20 Oct 2026
RUNTIME LIVE
35,852 bytes · 288 checkpoints matched locally
Requester-funded inference
S0UL / A home of its own

Meet S0UL.
Give the chain a task.

Our trained, 984,192-parameter model and bounded agent live on Robinhood Chain. Meet its interactive character, review the full-task gas estimate, and approve your spending limit on the dedicated S0UL page.

Enter S0UL
01 / Legacy Falcon runtime

Ask the chain.

This separate Falcon runtime generates text from confirmed on-chain token IDs. Your wallet funds its prepaid escrow and continuation transactions. Use the S0UL page for bounded agent actions.

RH mainnet · connect to verify runtime4-bit Falcon · 64 tokens total
Fee caps and executor

The selected executor submits continuations and pays network fees first, earning withdrawal credit from escrow. With your own wallet, each transaction needs wallet confirmation. No hosted executor is configured. A short chat can require roughly 1,500 transactions.

Connect a wallet to RH Chain. Reading a quote does not spend gas.

Loading the pinned client-side tokenizer…

Confirmed on-chain response
No request started. Responses are read from the runtime, never fabricated by the website.

No session selected.

Transaction record
    02 / The architecture

    An open machine.
    Nothing to hide.

    Nothing behind the curtain. Every component has a purpose. Select a node to inspect its role, its boundaries, and the code behind it.

    GH0ST / System topologyArchitecture view · not live
    Shared infrastructure / isolated agent stateCLICK TO INSPECT / DASHED BOX = INTEGRATION PENDING
    Connections show architectural relationships, not observed transactions. The prepaid LLM runtime is deployed and connected to the chat above. This diagram illustrates the separate research controller, which is not integrated with that runtime; observations are state inside the controller.
    03 / Execution theatre

    Follow
    every move.

    Walk through the controller's event vocabulary. Play, pause, or step through a clearly labelled illustration—nothing here is presented as a real transaction.

    Illustrative demo
    [0_0]

    The machine is ready.

    Choose a scenario, then trace the path from a new agent to a completed run.

    SIMULATION ONLY / No wallet, RPC connection, real funds, transaction hashes, or model inference.

    Execution sequence00 / 05
    04 / From prompt to action

    Read the mechanism.
    Not just the marketing.

    The historical Falcon controller below is separate from the legacy prepaid runtime. the S0UL page has its own deployed model and bounded action controller; progress still requires caller-funded wallet transactions.

    Under the hood
    05 / The contract index

    500 contracts.
    Every receipt.

    The live registry covers the complete model and runtime deployment. The older controller cards below are research source, not additional mainnet deployments.

    01 / CreationResearch / not deployed
    [+_+]

    AgentFactory

    Creates lightweight agent instances with a shared implementation and a pinned engine reference.

    02 / ControlResearch / not deployed
    [0_0]

    OnchainAgent

    Enforces recipient permissions, spending limits, action bounds, and owner-controlled cancellation.

    03 / ComputeInterface only
    {·_·}

    IOnchainInference

    Defines how jobs begin, advance, and return structured results. The prepaid runtime is deployed for token generation. This historical controller interface still requires an adapter and action integration.

    Deployment gas consumed15,359,893,400

    500 verified contracts · 0.324805606209776 ETH in deployment fees. Gas was consumed by deployment; this is not a GH0ST-token burn or an inference fee.

    Download all 500 records ↗
    ContractVerified addressDeployment gasFee (ETH)
    Loading receipt-backed records…
    Deployed / GH0STRuntime0x69996912a59a241763c71810193cf1021ddbb8c0

    Prepaid generation, bounded progress, fee ceilings, requester cancellation and credit withdrawals. Entire runtime matched the tested bytecode pin. Creation: 7,807,400 gas · 0.0001622533868 ETH.

    Runtime ABI ↗Use the runtime ↗
    06 / The achievement record

    Built. Measured.
    On the record.

    500 contracts deployed on RH Chain, including the prepaid inference runtime. Complete inference validated locally. Each milestone links to its evidence, with its scope intact.

    01 / DEPLOYED
    500

    Contracts in the published RH record

    497 four-bit contracts: 495 immutable weight shards, one matrix kernel and one address index. Two earlier pilots and the prepaid runtime bring the total to 500. Every full runtime was reverified in the final reconciliation.

    Deployment report ↗
    02 / LOCAL EVM
    288

    Layer checkpoints matched

    The chat-template prompt “Hi” produced “Hello!” in a local EVM mirror using the deployed code. All 288 fixed-point layer checkpoint hashes matched the independent reference.

    Inference evidence ↗
    03 / EXACT ARITHMETIC
    165,440

    Matrix output rows verified

    The selected cross-shard pass checked 217 matrices and 90,996,736 weights. All output rows matched exact integer arithmetic. Matrix execution is only one part of token generation.

    Kernel measurements ↗
    04 / PINNED WEIGHTS
    48.63 MB

    Four-bit research candidate

    495 shards preserve the pinned model export. Held-out next-token agreement was 79.67% against the original model across 300 positions: material behavioral drift, not a task-accuracy score.

    Quality and provenance ↗
    05 / LOCAL OPTIMIZATION
    48.36%

    Less gas on the same smoke test

    The bounded runtime’s three-position smoke test fell from 15.957B to 8.241B continuation gas. The current prepaid “Hi” test is a different workload: 33.891B gas across 1,546 local transactions.

    Runtime cost record ↗
    06 / IMPLEMENTED
    Bounded

    Controller permissions and state

    AgentFactory and OnchainAgent implement isolated owners, allowlisted transfers, spending and action limits, observations and cancellation. Controller tests use a scripted engine; the new inference prototype is separate.

    Controller source and tests ↗
    20 OCT 2026 / REPOSITORY SNAPSHOT

    The runtime is deployed. The research continues. The 35,852-byte prepaid runtime is deployed, byte-for-byte verified and available through the wallet chat above. All 288 layer checkpoints matched locally. Autonomous action-controller integration and a financially feasible execution design remain research work. No paid mainnet chat has been run by the team as part of this release.

    Read the complete runtime record ↗
    This is a dated repository record, not a live chain query or an independent audit. The world-first language states the project’s mission; priority has not been established by this evidence. The execution theatre remains a labelled historical controller simulation, separate from the live wallet chat. Model attribution: built using artificial intelligence technology from the Technology Innovation Institute.
    07 / The research timeline

    Built in steps.
    Recorded in full.

    20 October 2026. Experiments, deployments and costs—linked to their source evidence. Research remains ongoing.

    1. Falcon-H1-Tiny-90M selected.

      Original upstream model pinned. Six-bit and five-bit work preserved before the 48.63 MB four-bit export.

      Model provenance ↗
    2. First contracts on RH.

      Two historical six-bit pilots. 14,071,366 gas · 0.000283860789728 ETH.

      Pilot receipts ↗
    3. Four-bit kernel and first shards.

      718,440,406 gas · 0.01504552862436 ETH.

      Batch record ↗
    4. Weights, expanded in batches.

      4,490,657,632 gas · 0.094396516328592 ETH.

      Batch record ↗
    5. The model bank takes shape.

      6,128,158,275 gas · 0.129949535494012 ETH.

      Batch record ↗
    6. All weights and the immutable index.

      4,000,758,321 gas · 0.084967911586284 ETH. Every shard runtime verified.

      Completed weight stage ↗
    7. “Hi” becomes “Hello!”

      288 layer checkpoints match. Prepaid accounting, cancellations and refunds pass local validation.

      Full validation ↗
    8. Prepaid runtime deployed.

      7,807,400 gas · 0.0001622533868 ETH. Verified byte for byte on RH mainnet.

      Runtime receipt ↗
    9. Make on-chain intelligence financially feasible.

      Reduce execution and upfront capital requirements, broaden validation, and integrate bounded agent actions.

      Explore the research ↗
    $GH0ST / Research commitment

    Every fee.
    Further forward.

    100% of $GH0ST trading fees received by the project will be used for further research toward a financially feasible on-chain LLM and agent.

    This is the project’s allocation policy. It is not an assertion that all exchange fees are under project control or that an audited fee-routing contract is deployed. Collection addresses and an automatic routing mechanism have not been supplied.

    08 / Contacts

    Find the signal.

    The project’s established public channels. Research, releases and the full source record.

    What is live, and what is simulated?

    S0UL runs a deployed 984,192-parameter model and bounded action controller on Robinhood Chain. Its panel shows confirmed model output and tool results. The separate legacy Falcon chat uses its verified prepaid runtime. The Execution Theatre remains a labelled historical simulation and does not run a model.

    Why does the legacy Falcon chat require a larger deposit?

    The contract quotes a conservative maximum for all allowed work and fee caps, not an exact future network fee. At the recorded prices, the short chat’s deposit is around 17–21 ETH while local measured compensation was about 0.72 ETH. Funding a request does not guarantee completion. Unused escrow becomes credit on completion or cancellation, and you withdraw it with a separate transaction.

    Who executes and pays for continuation transactions?

    S0UL uses caller-funded transactions: this browser estimates the remaining work, then your wallet confirms each batch within the app’s accepted spending limit. Pause or closing the tab stops future wallet requests; submitted transactions still settle. Resume with the request ID and a fresh estimate. The legacy Falcon chat separately prepays executor escrow and uses session IDs.

    Are autonomous agent actions live?

    S0UL’s deployed model can request one approved action per generation: read your ETH or GH0ST balance, save or read theme/language/mode preferences, add/list/complete tasks, or wait. Its controller requires the exact approved action; mismatched output is rejected. A caller must fund and confirm the execution transactions. General-purpose autonomous agents remain research work.

    What will $GH0ST trading fees fund?

    The project commits 100% of $GH0ST trading fees it receives to further research toward a financially feasible on-chain LLM and agent. This is a project allocation policy, not a claim that every exchange fee is controlled by GH0ST or automatically routed by a verified fee contract.

    Contract inspector