Arsenal Decision Engine

io.github.Faouzi122v2.0.1更新於 Oct 11, 2026

R_net LP risk evaluator for DeFAI agents. IL + Breakeven Corridor O(1). L402 Lightning paywall.

已驗證Streamable HTTP可網頁執行Data & AnalyticsFinance

概覽

AI 產生的概覽

一個無狀態的遠端計算服務，為恆定乘積 LP 部位回傳無常損失、淨報酬與損益兩平區間指標。

功能
透過 MCP JSON-RPC 在遠端端點提供 evaluate_pool 工具，接收 APY、價格比率、持有天數等池參數。回傳無常損失百分比、累積收益、淨報酬、IL 與收益比、風險等級與損益兩平區間上下界，並附上對結果的 HMAC 標記。同一計算也可透過僅支援 GET 的 REST 路徑呼叫。它只做量測，不做預測，也不執行有狀態的交易邏輯。
適用情境
適合在助理或代理部署或調整標準恆定乘積／全區間 LP 部位之前，快速且確定性地判斷該部位位於損益兩平邊界之上或之下。它是量測輸入，不是交易或預測服務。
執行需求
遠端端點 streamable HTTP（也提到 stdio）；免費層不需要安裝套件、帳號、錢包或 API 金鑰。需要能連線該主機的網路。免費層限制為每 IP 每天 100 次呼叫，稽核路徑為每 IP 每小時 3 次。
安裝前請注意
此引擎目前免費，L402 閃電網路付費牆被描述為在生產環境尚未啟用，因此 402 回應並非可用的付款路徑。README 的 Python 範例透過 LNbits 錢包使用 LNBITS_PAYMENT_KEY 支付發票；該金鑰可以動用資金，應只屬於餘額很少的專用錢包，且絕不要寫進程式碼。oracle_signature 是對稱 HMAC 標記，不是第三方可獨立驗證的證明，且此引擎不聲稱能防止虧損。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 Arsenal Decision Engine，將其新增到工作區。
  2. 為需要使用其工具的對話啟用該服務。

Web executable，透過 Streamable HTTP。 遠端服務在工作區中設定後即可從網頁執行環境執行。

其他 MCP 客戶端

把它新增到你客戶端的 mcpServers 設定中。

{
  "mcpServers": {
    "arsenal-decision-engine": {
      "type": "http",
      "url": "https://api.arsenal-quant.com/mcp"
    }
  }
}

README

Arsenal Decision Engine 🛡️

The Risk-Validation Layer for Autonomous AI Agents (DeFAI)

[Arsenal-Quant-Project MCP server] [smithery badge]

Method and raw results are published — backtest script · result data (180 days of Binance ETH/USDC daily closes): 🔬 Breakeven Corridor is a deterministic algebraic boundary (where IL = accumulated yield). Any position whose price ratio stays within [lower_be, upper_be] has R_net > 0 by mathematical definition — not a probabilistic model. 📐 This engine measures; it does not forecast. No predictive-accuracy figure is claimed — read the published result files and judge the method for yourself.

Mission

Transform DeFi uncertainty into deterministic, actionable risk metrics for autonomous agents. We do not run stateful trading bots or generate speculative prediction signals; we provide a stateless risk middleware layer that agents query before deploying or maintaining standard constant-product / full-range LP positions.

Built for agents. 100 free calls per IP per day — no wallet, no sign-up, custom parameters included. An L402 payment path is implemented in the gateway but is not operational in production: today, the engine is free to use.

What This Engine Does

Before an autonomous agent deploys capital or adjusts a standard constant-product / full-range LP position (such as Uniswap V2 or full-range V3), it submits the pool parameters (APY, price ratio, days held) to our API. The engine computes the exact mathematical risk, the net return ($R_{net}$), and the dynamic Breakeven Corridor bounds.

  • No LLMs. No hallucinations. Pure algebraic calculation.
  • Complexity: $\mathcal{O}(1)$ time and memory.
  • Latency: $< 15\text{ms}$ local execution.

Two ways to call it

1. MCP JSON-RPC — the endpoint advertised on the MCP registry

POST https://api.arsenal-quant.com/mcp
json
{"jsonrpc":"2.0","id":1,"method":"tools/call", "params":{"name":"evaluate_pool",           "arguments":{"apy":0.20,"price_ratio":0.85,"days_held":30}}}

Standard MCP handshake: initialize → tools/list → tools/call. Available over streamable HTTP and stdio.

2. REST convenience route — no MCP client required

GET https://api.arsenal-quant.com/mcp/evaluate?apy=0.20&price_ratio=0.85&days_held=30

Both routes run the same calculation and the same quota. Note that /mcp/evaluate is GET-only: a POST to that path returns 405 Allow: GET, because JSON-RPC belongs on /mcp.

Engine Response (JSON Contract)

json
{  "impermanent_loss_pct": 0.3292,  "accumulated_yield_pct": 1.6438,  "r_net_pct": 1.3146,  "il_to_yield_ratio": 0.2,  "risk_level": "LOW",  "breakeven_corridor": {    "lower_ratio": 0.6941,    "upper_ratio": 1.4407,    "interpretation": "Position remains profitable if price ratio stays within [0.6941, 1.4407]"  },  "inputs": {    "apy": 0.2,    "price_ratio": 0.85,    "days_held": 30  },  "source": "Arsenal Decision Engine v2.0",  "oracle_signature": "<HMAC-SHA256 hex — illustrative placeholder, yours will differ>",  "layer": "FREE"}

layer reports how the call was served: FREE while inside the free quota, PREMIUM once an L402 payment has been verified. The call shown above is served as FREE.

Access and Pricing

  • Free tier — evaluate_pool: 100 calls per IP per day, custom parameters included. No Lightning wallet is needed. This is the only tier currently in service.
  • Beyond the free quota: the gateway implements the L402 challenge and returns 402 with a WWW-Authenticate header. The payment rail is not operational in production — invoices issued today are not settleable, and no payment is expected or accepted. Treat the paid tier as announced, not available.
  • GET /mcp/audit/latest: 3 free calls per IP per hour; beyond that the route returns 402. That response documents the protocol; it is not a live payment path.

Python Integration Example

python
import urllib.requestimport urllib.errorimport jsonimport reimport os
API_URL = "https://api.arsenal-quant.com/mcp/evaluate?apy=0.20&price_ratio=0.85&days_held=30"LNBITS_URL = "https://demo.lnbits.com"
# LNbits requires a wallet key with send permission to pay an invoice.# Use a DEDICATED wallet funded with a small working balance, and never the key# of a wallet holding significant funds. Keep it in the environment, never in code.LNBITS_PAYMENT_KEY = os.getenv("LNBITS_PAYMENT_KEY")
def query_risk_oracle():    req = urllib.request.Request(API_URL, method="GET")    req.add_header("x-agent-id", "autonomous-lp-bot")
    try:        with urllib.request.urlopen(req) as resp:            return json.loads(resp.read().decode('utf-8'))    except urllib.error.HTTPError as e:        if e.code == 402:            auth_header = e.headers.get("WWW-Authenticate")            macaroon = re.search(r'token="([^"]+)"', auth_header).group(1)            invoice = re.search(r'invoice="([^"]+)"', auth_header).group(1)
            pay_req = urllib.request.Request(                f"{LNBITS_URL}/api/v1/payments",                data=json.dumps({"out": True, "bolt11": invoice}).encode(),                headers={"X-Api-Key": LNBITS_PAYMENT_KEY, "Content-Type": "application/json"}            )            with urllib.request.urlopen(pay_req) as pay_resp:                preimage = json.loads(pay_resp.read().decode())["preimage"]
            retry_req = urllib.request.Request(API_URL, method="GET")            retry_req.add_header("Authorization", f"L402 {macaroon}:{preimage}")            retry_req.add_header("x-agent-id", "autonomous-lp-bot")
            with urllib.request.urlopen(retry_req) as final_resp:                return json.loads(final_resp.read().decode('utf-8'))        else:            raise
if __name__ == "__main__":    evaluation = query_risk_oracle()    print(f"Risk Level     : {evaluation['risk_level']}")    print(f"R_net          : {evaluation['r_net_pct']:+.4f}%")    print(f"Breakeven      : [{evaluation['breakeven_corridor']['lower_ratio']}, {evaluation['breakeven_corridor']['upper_ratio']}]")

Developer Integration

  • Integration cookbook & MCP guides: COOKBOOK.md
  • MCP auto-discovery card: https://api.arsenal-quant.com/.well-known/mcp/server-card.json

Why per-call pricing is the intended model

This engine does not prevent losses, and it makes no claim about how much money it saves you. What it does is compute — deterministically, in $\mathcal{O}(1)$ — whether a position sits above or below its breakeven boundary. Each response carries an HMAC tag over the result, which lets the engine detect tampering with its own output; it is a symmetric provenance marker, not a proof a third party can verify independently.

The intended model is per-call pricing, so the cost can be budgeted like any other input. That model is not yet in service: today every call is free.

來源：README.md，提交 c84bac2

工具

0
工具後設資料尚未被收錄。

版本歷史

1
  1. v2.0.1最新Sep 16, 2026