Prism

Prism — The First Adversarial AI Validator on ERC-8004

> Two AI agents debate your trade reasoning on-chain. One generates. One challenges. Both are accountable.

[](https://prism-dashboard-production-e6e3.up.railway.app) [](https://prism-docs-production.up.railway.app) [](https://youtu.be/6gjfT9GchYo) [](https://youtu.be/ZmSStvqNgIo) [](LICENSE) -purple)

Prism is an adversarial AI validation system built on ERC-8004. A Trader agent (Claude/Mirascope) generates structured reasoning traces for prediction markets. A Sentinel agent (GPT/DSPy) adversarially challenges every claim. The result is a Prism Report with an issue ledger, capital gate, IPFS receipts, and Arc/ERC-8004 anchors where present. Both agents are registered on-chain, and the sentinel is exposed as an x402-protected service other agents can call.

> Prism makes AI agents accountable to other AI agents before capital moves.

Creative showcase film: YouTube companion asset — a 70-second cinematic cut for sharing alongside the founder-led demo. The official judge demo remains the separate screen-recorded walkthrough.

---

Try it in 30 seconds

1. Open the live dashboard:
2. Connect MetaMask and switch to Base Sepolia
3. Get free testnet USDC from the Circle faucet (Base Sepolia + USDC)
4. Paste this known-good IPFS CID: QmNzqnPEEQUMn3GMbiEZANpKXZRPmTHxVwt5nNevR8iXt8
5. Click Validate, sign the 0.01 USDC payment in MetaMask
6. You land on /trace/[id] with the cross-family adversarial verdict

First real self-serve run on 2026-05-15 settled at 0x63bf7094… on Base Sepolia. Receipt: [docs/demos/self-serve-submit-20260515T101946Z.md](docs/demos/self-serve-submit-20260515T101946Z.md).

First live CLI paid validation on 2026-05-16 settled at 0xd6ab0cbb… on Base Sepolia. Receipt: [docs/demos/cli-paid-validation-20260516T214837Z.md](docs/demos/cli-paid-validation-20260516T214837Z.md).

Latest URL-verified MCP validation on 2026-05-17 settled at and produced a public report with Exa web_search_exa + web_fetch_exa evidence receipts: . Receipt: [docs/demos/external-call-20260517T203104+0000.md](docs/demos/external-call-20260517T203104+0000.md).

---

Architecture

                     Arc Testnet (chain 5042002)
                     USDC native gas · ~$0.01/tx · sub-second finality
                              │
              ┌───────────────┼───────────────┐
              ▼               ▼               ▼
     IdentityRegistry   ValidationRegistry   ReputationRegistry
     0x8004A8…BD9e      0x8004Cb…4272        0x8004B6…8713
              │               │               │
              └───────────────┼───────────────┘
                              │
                    Circle Dev-Controlled Wallets
                    (signs + submits all txs)
                              │
          ┌───────────────────┴───────────────────┐
          ▼                                       ▼
  Trader Agent (Python)                    Sentinel Agent (Python)
  Mirascope · Claude family                DSPy · GPT-4o-mini
  Trading-R1 trace generation              TraceAdversary + MIPROv2
           │                                        │
  Polymarket V2 SDK                                 ▼
  execution attribution                    x402-protected MCP
  (builder code)                           ($0.01 USDC/validation)
           ▵                                        │
  Polymarket CLOB (Polygon)                         │
                                                      │
                              ┌───────────────────────┘
                              ▼
                    Dashboard (Next.js 16)
                    split-screen dialogue + on-chain receipts
                    wallet-connected · self-serve validation

---

Data Flow

1. Fetch Polymarket market → 2. Trader generates Trading-R1 trace → 3. Pin to IPFS → 4. Persist to Postgres → 5. On-chain validationRequest → 6. Sentinel challenges trace → 7. Verdict + dialogue → 8. Pin to IPFS → 9. Persist to Postgres → 10. On-chain validationResponse → 11. Optional paper/live trade carries builder code → 12. Dashboard execution attribution

---

What's Live

Dashboard Routes

| Route | Description | |-------|-------------| | / | Landing page with the product story, live activity strip, and waitlist | | /dashboard | Split-screen trace + verdict dialogue, on-chain receipts | | /trace/[id] | Trace detail page — server component with structured layout, capital gate, receipts, and dynamic OG image | | /history | Paginated history of all traces and verdicts | | /me | Wallet-connected verdict history (wagmi v2 + Reown AppKit) | | /submit | Self-serve x402 validation — sign EIP-3009 transfers from the browser; shows Circle App Kit Bridge widget when USDC balance . They cover the quickstart, CLI, x402/MCP validation, public APIs, receipts, security model, architecture, and the verification-guarantees page (/docs/verification-guarantees — 8 guarantees, 5 NOT-guarantees, trust assumptions, fail-closed behavior).

The CLI is the developer-facing surface for pulling Prism metrics without opening the dashboard. It supports trace inspection, public stats/history, trace reports, market surfacing, token resolution, wallet funding guidance, and x402 validation orchestration. It never reads private keys: prism validate either submits an externally signed X-PAYMENT header or asks Circle CLI to sign the EIP-712 authorization with a Circle wallet.

uvx --from "prism-cli @ git+https://github.com/Fato07/prism.git#subdirectory=apps/cli" prism demo

# or, from a clone:
cd apps/cli
uv run prism --version
uv run prism doctor
uv run prism demo
uv run prism inspect ipfs://QmNzqnPEEQUMn3GMbiEZANpKXZRPmTHxVwt5nNevR8iXt8 --json
uv run prism stats
uv run prism history --limit 5
uv run prism report d6cdd60f-f5e0-43ab-ba2d-7dcab76a8e24
uv run prism markets --limit 5
uv run prism market resolve "" --json
uv run prism quote ipfs://QmNzqnPEEQUMn3GMbiEZANpKXZRPmTHxVwt5nNevR8iXt8 \
  --trace-hash 0x1a750011608a7480e9cb11f1d20587e32efb7a7dd433b85820f0dbfcdee19fdb
uv run prism validate ipfs://Qm... --trace-hash 0x... --x-payment-file ./x-payment.txt
uv run prism demo --pay --circle-address 0xYourCircleWallet --max-amount-usdc 0.01
uv run prism validate ipfs://Qm... --trace-hash 0x... \
  --circle-address 0xYourCircleWallet --max-amount-usdc 0.01

---

Calibration Corpus

The calibration corpus is the sentinel's evaluation ground truth — a local-first, deterministically split, immutably sealed dataset of labeled reasoning traces. It ships as a CLI in packages/calibration-python/.

uv run python -m prism_calibration.cli --help

Key workflows:

1. Build — scaffold a new corpus with schema and split configuration
2. Harvest — pull real traces from Prism's activity store → IPFS → normalize → hash-verify → write local rows
3. Label — generate synthetic traces, mutations, or AI-prelabel existing rows (3 subcommands: generate-synthetic, generate-mutations, prelabel)
4. Freeze — export an immutable snapshot (manifest.json + rows/ + .sealed marker)
5. Sync — mirror the frozen export to Braintrust (datasets, review queue)
6. Eval — run the sentinel against the corpus and score discrimination
7. Inspect — examine corpus metadata, splits, and sealed milestones
8. Validate — run regression assertions against sealed milestones (43/43 passed across 5 milestones)

Architecture — local-first:

• Local frozen exports are the authoritative data source
• Braintrust is a mirror and experiment surface (datasets, experiments, review queue, runtime logs)
• Deterministic split assignment via lineage-hash-v1 (no randomness, fully reproducible)
• Holdout lockout: holdout rows are isolated from training/labeling and locked after freeze
• Immutable sealed exports: once a milestone is sealed, its rows cannot be modified
• Clean-replay regression: re-run eval on any sealed milestone and bit-compare to the original result

---

What's Open / What's Gated

| Open | Gated (not in this repo) | |------|--------------------------| | ERC-8004 client & validator SDK scaffolding | Production sentinel prompts (MIPROv2-optimized) | | Agent harness & trace generation pipeline | HMAC seed material | | Sample reasoning traces | Circle Entity Secret & wallet private keys | | Dashboard public routes + wallet connection | | | Self-serve x402 validation page | | | Execution attribution page | | | Contract addresses & ABIs | | | x402 middleware setup (dual facilitator mode) | | | DSPy TraceAdversary signature | | | Treasury dry-run scaffold (USYC park/unpark path) | | | Calibration CLI + schemas (8 commands, 149 tests) | |

---

Operator Safety Control Plane

Runtime observability and safe control for Prism's autonomous trader — read status, start/stop scheduler with auth, audit trail for all mutations.

Key Components

| Component | Description | |-----------|-------------| | Trader GET /status | Read-only endpoint returning 8 fields: scheduler_running, interval_minutes, auto_pipeline_enabled, trade_mode, last_tick_timestamp, next_tick, last_error, service_version. Reading status never starts the scheduler. | | Dashboard operator-admin auth | All admin routes gated by OPERATOR_ADMIN_TOKEN (Bearer header). Verified via timingSafeEqual (separate from CONNECTOR_ADMIN_TOKEN). Invalid or missing token → 401 + audit row with result=unauthorized. | | Dashboard admin API routes | GET /api/admin/runtime, POST /api/admin/schedule/start, POST /api/admin/schedule/stop, GET /api/admin/audit — proxy to the trader with auth enforcement. | | Operator page (/operator) | Read-only status card + start/stop mutation buttons with confirmation dialog + live audit events table. Requires admin auth. | | operator_events audit log | Append-only Postgres table (infra/db/migrations/005_operator_events.sql). Every mutation attempt (success, failure, unauthorized) writes a row. No UPDATE or DELETE issued by application code. |

Key Safety Invariants

• Reading status never starts the scheduler. The GET /status endpoint is purely observational.
• PRISM_TRADE_MODE is read-only (env-only). The API cannot switch between paper and live. To change trade mode, update the environment variable and redeploy.
• AUTO_PIPELINE=false is the default and cannot be changed via API. The scheduler must be explicitly started by an authenticated operator action; start only activates the tick scheduler, not auto-pipeline.
• Start cannot switch paper→live. If PRISM_TRADE_MODE=paper, the start endpoint only starts in paper mode. No runtime escalation.
• Stop cancels background tasks. The stop endpoint cancels the scheduler's asyncio background task and sets scheduler_running=false.

Environment Variables

| Variable | Description | |----------|-------------| | OPERATOR_ADMIN_TOKEN | Bearer token for admin API routes and /operator page access. Must be set in both trader and dashboard services. Verified via timingSafeEqual — separate from CONNECTOR_ADMIN_TOKEN. | | TRADER_INTERNAL_URL | Internal URL where the dashboard reaches the trader service (e.g. http://prism-trader.railway.internal:3201 for Railway private networking). |

---

On-Chain State

Both agents are registered on Arc testnet's ERC-8004 IdentityRegistry:

| Agent | agentId | Role | |-------|---------|------| | Trader | 4140 | Generates Trading-R1 traces, requests validation, executes paper trades with builder-code attribution | | Sentinel | 4148 | Adversarially validates traces, submits verdicts on-chain |

Contract Addresses (Arc Testnet, chain 5042002):

| Registry | Address | |----------|---------| | IdentityRegistry | 0x8004A818BFB912233c491871b3d84c89A494BD9e | | ValidationRegistry | 0x8004Cb1BF31DAf7788923b405b754f57acEB4272 | | ReputationRegistry | 0x8004B663056A597Dffe9eCcC1965A193B7388713 | | ERC-8183 AgenticCommerce | 0x0747EEf0706327138c69792bF28Cd525089e4583 |

---

Live Demo

Dashboard deployed on Railway:

https://prism-dashboard-production-e6e3.up.railway.app

The split-screen view shows the trader's reasoning trace on the left and the sentinel's adversarial challenges on the right, with on-chain validation receipts below. Connect your wallet to view your verdict history (/me) or submit a trace for validation (/submit).

---

Call the Sentinel Yourself — External x402 + MCP Demo

Prism's sentinel is a paid public service: any external agent can pay $0.01 USDC on Base Sepolia and get an adversarial verdict on its reasoning trace. The flow is x402 (payment protocol) over MCP (Model Context Protocol).

Reference client: [scripts/call_prism_sentinel.py](./scripts/callprismsentinel.py)

It's ~470 lines of standalone Python with PEP 723 inline dependencies — fork it, swap the wallet/trace, point it at your own sentinel. No Circle account, API key, or special tooling required on the caller side.

Run it in 3 commands

# 1. First run — generates a fresh keypair, prints the address, exits cleanly
#    (the private key persists to .local/prism-client.key, gitignored)
uv run scripts/call_prism_sentinel.py

# 2. Fund the address it printed with ~0.05 USDC on Base Sepolia
#    (free from https://faucet.circle.com — select Base Sepolia + USDC)

# 3. Re-run — executes the full x402+MCP dance, saves Markdown receipt
uv run scripts/call_prism_sentinel.py

Expected output:

Prism — external x402 + MCP demo client
========================================================

  • Loaded client wallet from .local/prism-client.key
  • Client address: 0x993F4D56e1329b6e3b91A13B9ACe1a890a023518
  • USDC balance on base-sepolia: 19.9300 USDC

  > [0/4] MCP handshake — initialize …
  ✓ Session: 468d6db939c94831853f5956…
  > [1/4] Calling sentinel /mcp with session but no payment …
  ✓ 402 received · 0.01 USDC to 0x1453ba8a… on base-sepolia
  > [2/4] Signing EIP-3009 transferWithAuthorization …
  ✓ Payment payload ready (956 char base64)
  > [3/4] Re-calling /mcp with X-PAYMENT header …
  ✓ 200 OK · response 3020 bytes
  > [4/4] Parsing verdict + saving receipt …
  ✓ Receipt: docs/demos/external-call-.md

  Verdict: PASS (score 65/100)
  Settlement: https://sepolia.basescan.org/tx/0x8d5d7a46bd88b2ffba5b6e6c6d70221598ee041d4206377ca5991e90f5c12421

What's verifiable when you run it

• A real EIP-3009 USDC transferWithAuthorization signed by your client

wallet, settled on Base Sepolia by the x402.org public facilitator

• A real adversarial verdict from a different model family than the

trace's original author, pinned to IPFS

• A committed Markdown receipt in docs/demos/ with all hashes and

explorer links — shareable, screenshot-able, gitable

How it works under the hood

See [docs/architecture.md](./docs/architecture.md) for the full sequence diagram. The short version:

1. MCP handshake — initialize then notifications/initialized (free,

exempt from x402 paywall so clients can establish a session)

1. First tools/call validate — sentinel returns HTTP 402 with x402

payment requirements in a JSON-RPC error envelope

1. Client signs an EIP-3009 TransferWithAuthorization typed-data

payload (USDC contract on Base Sepolia, name="USDC", version="2")

1. **Second `tools/call

…

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Fato07
• Source: Fato07/prism
• License: Apache-2.0
• Homepage: https://prism-dashboard-production-e6e3.up.railway.app

Install and usage instructions live in the source repository linked above.