🧠 VeriModel: Decentralized AI Model Benchmark & Verifiable Evaluation Escrow Protocol

VeriModel is a decentralized, intelligent milestone evaluation and bounty escrow protocol built natively on GenLayer. It solves the reproducibility and benchmark contamination crisis in open-weight AI research by conditioning grant/prize disbursements on decentralized multi-validator neural consensus over live authority leaderboard telemetry (HuggingFace Open LLM Leaderboard API, LMSYS Chatbot Arena, OpenRouter, and Weights & Biases).

---

🎯 The Web3 & AI Problem: Benchmark Contamination & Fake Evals

As DAOs, research foundations, and decentralized compute networks distribute millions in grants and bounties for open-weight AI models, they face a critical dilemma:

• Benchmark Contamination & Overfitting: AI models frequently overfit to public test sets or report fabricated benchmark metrics (HumanEval, MMLU, GSM8K) in non-reproducible environments.
• Unbonded AI Grant Claims: Grant sponsors disburse funding upfront, leaving zero recourse if the delivered open-weights model fails independent evaluation.
• Traditional Oracles Cannot Parse AI Evals: Scalar price oracles (Chainlink) cannot parse unstructured HuggingFace eval trees, LMSYS Arena ELO rankings, or JSON benchmark harnesses.

💡 Why GenLayer is Central to VeriModel

GenLayer provides the only execution layer capable of evaluating off-chain AI benchmark reproducibility:

1. Live Authority Leaderboard Telemetry Grounding (gl.nondet.web.get): Validators independently retrieve real-time eval metrics from committed endpoints (huggingface.co/api, lmarena.ai, openrouter.ai).
2. Multi-Validator Neural Consensus (gl.vm.run_nondet_unsafe): Validators analyze benchmark scores against contracted specifications under the Equivalence Principle with strict canonical action decisions.
3. Deterministic Slashing & Exact Payout Preservation: Slashes developer collateral only upon verified affirmative proof of falsification/failure (confidence >= 80) or awards the bounty prize upon verified reproducibility with zero financial drift.
4. 404 & Network Failure Retry Protection: HTTP 404s, missing endpoints, or server errors are strictly treated as retry outcomes (EXTEND_EVAL_WINDOW), NEVER as grounds for slashing.

---

🏛️ Exact Payout Preservation & Canonical Action Consensus

To eliminate numeric drift and ambiguous threshold crossings, VeriModel enforces Canonical Action Decisions:

Canonical Action Decision	Validation Criteria	On-Chain Execution
RELEASE_BOUNTY	benchmark_achieved == True AND confidence_score >= 80	Releases bounty prize + stake refund to Developer (emit_transfer(bounty + stake))
SLASH_CHALLENGE	Verified affirmative benchmark falsification / contamination AND confidence_score >= 80	Slashes developer stake + refunds bounty to Sponsor (emit_transfer(bounty + stake))
EXTEND_EVAL_WINDOW	In-progress eval runs, HTTP 404s, network errors, or sub-threshold confidence (confidence < 80)	Challenge remains active for retry; zero funds released

Key Security & Solvency Invariants:

1. 🗓️ Fail-Closed Runtime Block Timing: Timestamps are strictly derived from enforceable GenLayer runtime block state (_get_runtime_timestamp()). Unavailable timestamps strictly fail closed.
2. 🌐 Strict Authority Host Whitelist (SSRF Hardened): Exact hostname extraction neutralizes subdomain, query, and path spoofing (e.g., huggingface.co.attacker.com is strictly rejected).
3. 🔒 Committed Source Adjudication Binding: Adjudication is strictly bound to the target leaderboard URL committed on-chain during challenge creation. Callers cannot substitute uncommitted URLs.
4. 📡 Fail-Closed HTTP 200-299 Status Validation: Telemetry responses missing explicit status or returning non-2xx status codes fail closed immediately.
5. 🏦 100% Solvency Invariant: Tracks total active liabilities (total_active_liabilities) and prevents over-allocation.

---

🏛️ System Architecture

sequenceDiagram
    autonumber
    actor Sponsor as 🏛️ AI Grant Sponsor / DAO
    participant VeriModel as 🧠 VeriModel (GenVM)
    actor Dev as 🧑‍💻 AI Model Developer
    participant Validators as ⚖️ GenLayer Validators (Optimistic Democracy)
    participant Web as 🌐 Authority Leaderboard (HuggingFace / LMSYS)

    Sponsor->>VeriModel: create_challenge(dev, "CODING_HUMANEVAL", spec, committed_url, stake) + deposit 100 GEN
    Note over VeriModel: Locks 100 GEN bounty & commits leaderboard endpoint with strict host validation
    Dev->>VeriModel: stake_and_enter_challenge(chal_id) + deposit 30 GEN stake
    Note over VeriModel: Challenge status = ACTIVE (Total 130 GEN locked)
    
    Dev->>VeriModel: adjudicate_benchmark(chal_id, notes, committed_url)
    
    rect rgb(15, 23, 42)
        Note over VeriModel,Validators: Non-Deterministic Multi-Validator Consensus
        Validators->>Web: gl.nondet.web.get(committed_url)
        Validators->>Validators: gl.nondet.exec_prompt(Evaluate benchmark scores & reproducibility)
        Validators->>Validators: Equivalence Principle Check (Canonical Action Match & Non-Crossing Threshold)
    end

    alt Benchmark Verified (Confidence >= 80)
        VeriModel->>Dev: emit_transfer(130 GEN) [Bounty Prize + Stake Refund]
    else Falsified Evals / Failed Thresholds
        VeriModel->>Sponsor: emit_transfer(130 GEN) [Bounty Refund + Slashed Developer Stake]
    else In-Progress Evaluation
        VeriModel->>VeriModel: status = ACTIVE (Evaluation window extended for retry)
    end

Loading

---

📁 Repository Structure

verimodel-protocol/
├── contracts/
│   └── verimodel.py           # Core Intelligent Contract on GenVM
├── frontend/
│   ├── index.html             # Glassmorphic DApp UI with live genlayer-js client
│   └── client.ts              # TypeScript GenLayer client integration SDK
├── tests/
│   ├── direct/
│   │   └── test_verimodel.py  # 100% Passing in-memory direct VM test suite (7 scenarios)
│   └── integration/
│       └── test_verimodel_integration.py # StudioNet / RPC deployment integration tests
├── pytest.ini                 # Pytest direct suite collection configuration
├── gltest.config.yaml         # GenLayer Testnet/StudioNet network configuration
├── package.json               # genlayer-js & development dependencies
├── requirements.txt           # Python dependencies (genlayer, pytest)
└── README.md                  # Complete architectural & technical documentation

---

💻 Frontend & GenLayer Client Integration

The included interactive DApp (frontend/index.html) is connected to the real genlayer-js@1.2.0 client, enabling full on-chain lifecycle management:

1. Wallet / Account Management: Auto-generates testnet keypairs or imports custom private keys.
2. Multi-Network Support: Switch seamlessly between GenLayer Bradbury Testnet (4221), StudioNet (4222), and LocalNet.
3. Bounty Challenge Deployment: Create challenges, define verifiable score targets, and commit to authority endpoints (create_challenge).
4. Developer Staking: Deposit collateral bonds to activate challenges (stake_and_enter_challenge).
5. Live Neural Adjudication: Trigger multi-validator consensus over live authority leaderboards (adjudicate_benchmark).
6. Live Contract State Queries: Dynamically reads get_challenge and get_protocol_stats with explorer links.

TypeScript Client Example (frontend/client.ts):

import { getGenLayerClient, createChallenge, stakeAndEnterChallenge, adjudicateBenchmark, getChallenge } from './frontend/client';

const client = getGenLayerClient('0xYourPrivateKey...');
const contractAddress = '0xB8e1c3559B66B1b1d7d0823FBEB5A967732e999';

// 1. Sponsor creates 7-day Coding Benchmark Challenge (100 GEN bounty, 30 GEN required developer stake)
const tx1 = await createChallenge(
  client,
  contractAddress,
  '0xModelDeveloper...',
  'CODING_HUMANEVAL',
  'HumanEval pass@1 >= 80.0%, MBPP >= 75.0%',
  'https://huggingface.co/api/models/open-llm-leaderboard/evals/deep-coder-v2',
  30, // Required stake
  86400 * 7,
  100 // Bounty deposit
);

// 2. Developer stakes 30 GEN collateral to activate challenge
const tx2 = await stakeAndEnterChallenge(client, contractAddress, 0, 30);

// 3. Trigger Benchmark Adjudication on committed leaderboard telemetry
const tx3 = await adjudicateBenchmark(client, contractAddress, 0, 'Official HF run completed (84.6% HumanEval)', 'https://huggingface.co/api/models/open-llm-leaderboard/evals/deep-coder-v2');

// 4. Query Final On-Chain State
const chal = await getChallenge(client, contractAddress, 0);
console.log(`Status: ${chal.status}, Verdict: ${chal.adjudication_verdict}, Confidence: ${chal.adjudication_confidence}%`);

---

🧪 Test Suite & Verification

Run the complete direct test suite:

pytest
# or
pytest tests/direct/ -v

Verified Test Scenarios (9/9 Tests Passing):

1. test_benchmark_success_and_bounty_release:
   • Sponsor creates challenge with 100 GEN bounty. Developer deposits 30 GEN stake.
   • Live telemetry proves 84.6% HumanEval (>80% target).
   • Consensus on RELEASE_BOUNTY (conf: 96) -> 130 GEN released to Developer.
2. test_repeated_404_responses_leave_challenge_active_and_move_no_funds:
   • Multiple consecutive HTTP 404s and server errors strictly return EXTEND_EVAL_WINDOW.
   • Leaves challenge status ACTIVE, is_finalized False, and moves exactly 0 funds.
3. test_sub_threshold_confidence_cannot_slash_and_moves_no_funds:
   • Sub-threshold confidence (<80%) is strictly normalized to EXTEND_EVAL_WINDOW.
   • Prevents premature slashing on ambiguous evidence, moving 0 funds.
4. test_benchmark_falsification_slashing_with_high_confidence (Adversarial):
   • Developer submits falsified/contaminated benchmark (52.4% vs 80% target).
   • Consensus on SLASH_CHALLENGE with high confidence (conf: 98 >= 80) -> 130 GEN awarded to Sponsor.
5. test_in_progress_eval_grace_period_extension:
   • Queued or in-progress runs trigger EXTEND_EVAL_WINDOW without releasing funds.
6. test_mismatched_leaderboard_url_reverts (Adversarial):
   • Reverts when caller attempts to submit an uncommitted URL during adjudication.
7. test_untrusted_domain_spoofing_reverts (Adversarial):
   • Rejects hostname substring spoofing attempts (e.g. huggingface.co.attacker.com).
8. test_unauthorized_early_release_reverts (Fail-Closed):
   • Verifies active challenges cannot be released before expiration.
9. test_non_developer_stake_reverts (Access Control):
   • Enforces that only the designated developer can deposit collateral.

---

📄 License

MIT © Lesnak1 & GenLayer Community