Proofline MCP
Enables AI agents to verify World Cup match evidence, check settlement readiness, purchase proof packets, and anchor results on Injective EVM testnet.
README
Proofline
Don’t trust the score. Re-run the proof.
Proofline is a conflict-aware match-verification and settlement layer for World Cup products and AI agents. It turns attributed sports observations into a reproducible evidence packet, holds automation when sources disagree, and anchors the final canonical hash on Injective EVM testnet.
The product is built around VARA — the Verifiable Autonomous Referee Agent. Unlike a score widget with a blockchain button, VARA exposes the full decision path: provenance, independence groups, disagreement, Evidence Score inputs, canonical event JSON, payment boundary, and on-chain commitment.
Hosted demo: proofline.axiqo.xyz ·
GitHub: a252937166/proofline ·
Immutable release: global-cup-final
Public Injective testnet proof
The complete paid-proof path was executed on Injective EVM testnet
(eip155:1439) on 11 July 2026:
| Evidence | Public result |
|---|---|
| Registry v3 | 0x380D75d068dec45D8145ef89B7A40a6201Ac1ef1 — source fully verified, Solidity 0.8.35 |
| Deploy | 0xdf71…9523 |
| Grant anchorer role | 0x5858…acb4 |
Anchor WC-2026-M97-FRA-MAR final result |
0x24cd…7344 |
x402 settlement, 0.01 test USDC |
0x2923…842e |
This 2026 purchase moved the payer from 19.99 to 19.98 test USDC and the
payee from 20.01 to 20.02. The complete public packet, issuer signature,
anchor, purchase-binding hash, and receipt are available as the
no-wallet featured sample.
The judge moment
The default case is the real 2026 delayed result France 2–0 Morocco. Its
ESPN and FIFA observations converge into evidence root
0xe048…61b3, which is anchored and sold through the real testnet path above.
Then run the secondary historical replay of Wales 0–2 IR Iran (25 Nov 2022). A
synthetic lagging feed changes Wayne Hennessey’s 86th-minute red card to yellow.
Proofline immediately pulls the Evidence Score back, marks the event contested, and
keeps settlement held. When the official corroboration arrives and the bad
claim is retracted, the same deterministic verifier recovers. Late goals,
full-time status, and a matching anchor finally open the settlement gate.
The fault is labelled synthetic everywhere. The football facts are sourced from a fixed CC0 OpenFootball commit and an independently linked FIFA match review; the replay is permanently labelled Historical Replay · Not Live.
The product also exposes two honest 2026 data modes. WC-2026-M97-FRA-MAR is
a delayed, post-match France 2–0 Morocco snapshot with ESPN and FIFA provenance.
WC-2026-M99-NOR-ENG and WC-2026-M100-ARG-SUI are scheduled fixtures with
score=null; they are never presented as live matches. The deterministic 2022
replay remains the reproducible conflict-control path, while the 2026 result is
the primary proof, anchor, x402, and fresh-verification path.
End-to-end product
flowchart LR
A["World Cup sources"] --> B["Normalized observations"]
B --> C["VARA evidence graph"]
C -->|conflict| D["Quarantine settlement"]
C -->|corroborated| E["Canonical event hash"]
E --> F["Injective EVM anchor"]
C --> G["Free match API"]
C --> H["x402 proof packet"]
G --> I["Proofline MCP"]
H --> I
F --> I
I --> J["Agent Skill policy"]
K["CCTP future funding path"] -.-> H
Competition fit
| Requirement | Proofline implementation | Judge proof |
|---|---|---|
| Injective | Append-only MatchProofRegistry on Injective EVM testnet |
Fully verified contract, deployment, role grant, and real anchor linked above |
| x402 | 0.01 native testnet USDC proof resource with a policy-capped Agent flow |
Real successful settlement plus payer/payee balance deltas linked above |
| USDC / CCTP | Native testnet USDC payment is executed; CCTP is future work | No burn, attestation, or mint transaction is claimed |
| MCP Server | Ten domain tools for match lookup, event verification, settlement readiness, proof purchase, packet and anchor checks | Stdio runtime plus committed official Injective MCP execution evidence |
| Agent Skill | Explicit source, settlement, spending, replay, and CCTP safety rules | Agent refuses low-score or non-final settlement |
| AI-native product | Machine-verifiable evidence and deterministic decisions, not a chat wrapper | Packet integrity, trusted issuer signature, and latest on-chain commitment verify independently |
Quick start
Requirements: Node.js 20+ and npm.
cp .env.example .env
npm install
npm run check
npm run dev
Open http://localhost:5173. The API listens on
http://localhost:8787. The page opens on the 2026 France–Morocco delayed
snapshot and exposes its two source lanes without claiming a live feed. The
default configuration is intentionally a no-key, no-wallet sandbox: the
secondary 2022 conflict-control flow emits a real HTTP 402 negotiation shape
and an explicitly labelled deterministic demo receipt. It never claims that
sandbox payment or anchoring happened on-chain.
Conflict replay script
- Reset the historical replay.
- Run it and watch the red-card event move from
observedtocontested. - Inspect the exact conflicting field and source independence groups.
- Let the replay reach full time, verification, and anchoring.
- Request the premium proof: the first response is
402 Payment Required. - Use the labelled sandbox signature to retrieve the packet locally.
- Inspect its three verification layers, then modify one field and observe packet integrity fail.
For the real testnet path, follow the guarded commands in the testnet runbook. The deployment command fills the registry address automatically; anchor and x402 payment commands remain non-broadcasting until their action-specific flag and acknowledgement are both present.
Commands
npm run dev # API + web control room
npm run check # typecheck + tests + production builds
npm run compile:contract # compile Solidity registry
npm run wallets:create:testnet # create isolated testnet-only wallets in .env
npm run deploy:contract # deploy registry and atomically fill .env (broadcasts)
npm run verify:contract:testnet # publish and verify exact compiler input
npm run testnet:preflight # read-only registry, role, balance and x402 checks
npm run testnet:api # real anchor + official inline facilitator; no startup tx
npm run anchor:testnet # prepare replay anchor; no tx unless explicitly approved
npm run buy:proof # official Agent quote/sign-only; no payment by default
npm run evidence:publish-sample # publish a settled packet without its payment signature
npm run mcp # start the stdio MCP server
npm run smoke # exercise replay, 402, packet and tamper rejection
Verification model
96.49/100 is an Evidence Score, not a probability that an event is true.
It is a deterministic policy score built from source reliability, independent
quorum, agreement, freshness, and conflict penalties. Only one strongest
representative per independenceGroup can affect candidate ranking and score;
duplicating one upstream feed 100 times adds no voting weight.
Every premium packet is checked at three independent layers:
- Packet integrity: recompute canonical event JSON,
eventHash,evidenceRoot, policy result, andpacketHash. - Trusted issuer: recover the EIP-712 signer and require it to match the configured trusted issuer, not merely any cryptographically valid signer.
- Latest on-chain commitment: read the match-wide latest registry revision and match its event hash, evidence root, score, and valid state. A later dispute or rejection invalidates an older proof for settlement.
Run npm run check for type checks, unit/integration tests, real-EVM contract
tests, browser tests, production builds, and Solidity compilation. npm run smoke exercises conflict, recovery, proof negotiation, verification, and
tamper rejection locally.
Repository map
apps/web/ React control room and replay/proof experience
apps/api/ Replay engine, evidence API, anchoring and x402 boundary
packages/core/ Canonicalization, VARA verifier and portable proof packet
packages/mcp/ AI Agent-facing MCP tools and hard spending policy
contracts/ Injective EVM append-only proof registry and deploy script
skills/ Project-owned Agent Skill
data/replays/ Attributed, deterministic historical replay fixture
docs/ Architecture, trust, design and judging evidence
Trust boundaries
- One source family never counts as multiple independent votes.
- A single source cannot reach
verified. - Any active incompatible claim forces
contestedand holds settlement. - Each browser/MCP replay has an isolated session, so judges cannot reset one another's evidence state.
- Every 402 quote freezes a packet hash for five minutes; the paid retry is rejected before settlement unless its signed requirement carries that quote ID, so replay progress cannot swap the report after review.
- Signed payments are journaled by session, packet, payer, and EIP-3009 nonce. Concurrent duplicates are locked, settled retries are served from cache, and uncertain outcomes remain pending across restarts instead of charging again.
- A second
ProofPurchaseEIP-712 signature bindspacketHash, payer, payee, amount, deadline, USDC nonce, and browser session. The UI requests it from a separate, explicit page click so OKX surfaces confirmation2/2; signature1/2remains in memory and no payment header is submitted between the two. - The browser persists only a high-entropy, non-spendable recovery capability,
never
PAYMENT-SIGNATURE. A settled report can be restored after reload or a Chrome restart without opening a wallet or calling the facilitator. - Legacy packets affected by the historical replay-clock issuance bug may be reissued without repayment only when the original integrity/current-issuer signature passes, the issuer-time failure is the sole failed check, payment chronology is valid, and a fresh Injective registry lookup matches. The paid packet hash and replacement packet hash remain explicitly separate.
- Frozen packet JSON and entitlement state are atomically persisted with mode
0600, so a restart cannot regenerate a different packet for a settled quote. - Registry v3 has one concurrency-aware writer,
appendRevision; the shortcut writer is removed and aFinaldecision is fully immutable. - A final result requires a finished match, threshold Evidence Score, no conflict, and a confirmed anchor for the same canonical hash.
- Real testnet verification checks registry identity and the match-wide latest revision. Demo receipts never pass this public-chain check.
- Provider credentials and raw licensed payloads stay server-side.
- The chain proves a commitment and ordering, not sporting truth by itself.
- Demo and testnet paths are distinct machine-readable modes.
- CCTP is future work; no burn, attestation, mint, or balance recheck is claimed.
See the product specification, architecture, trust model, data provenance, visual system, and the judge guide for the complete reasoning.
Production release assets for proofline.axiqo.xyz are documented in the
CentOS deployment guide. The deployment uses a
loopback-only API, nginx, systemd, root-only environment file, checksum-verified
immutable releases, health-gated atomic switching, and rollback.
Testnet configuration
Proofline targets Injective EVM testnet chain ID 1439 (eip155:1439) and
native testnet USDC. Copy .env.example, deploy the registry, and set:
CHAIN_MODE=injective-testnet
PROOF_REGISTRY_ADDRESS=0x...
ANCHOR_PRIVATE_KEY=0x...
X402_MODE=injective-testnet
X402_PAY_TO=0x...
X402_FACILITATOR_PRIVATE_KEY=0x...
Use funded testnet-only wallets. Never commit .env.
After completing the replay, npm run buy:proof uses the official Injective
x402 client to quote, policy-check, and sign an EIP-3009 authorization with the
dedicated X402_AGENT_PRIVATE_KEY, but keeps it in memory by default. A real
settlement additionally requires --pay and the exact ephemeral acknowledgement
documented in the testnet runbook. It refuses any origin, redirect, chain,
asset, payee, frozen packet hash, or price outside the allowlist in .env.
Agent execution evidence
Proofline's stdio MCP exposes ten narrow tools, from list_matches and
verify_event through purchase_match_proof and verify_onchain_anchor. The
repository also contains a reproducible, sanitized execution transcript from
the official Injective MCP at pinned commit
f5af39367975872a85b5447cefc9a197f2e635ea: it listed 37 tools and successfully
executed address_normalize, usdc_native_info, and account_balances on
testnet; the balance call reports the payer's real testnet holdings. See
evidence/agent/official-injective-mcp.json
and the capture script. A second
transcript records five real calls through Proofline's own ten-tool MCP,
including a fresh revision-1 verify_onchain_anchor read:
evidence/agent/proofline-mcp-testnet.json.
License
Project code is MIT licensed. Replay data attribution and third-party terms are documented separately in data provenance.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。