AlphaAssay

거래 신호 및 백테스트에 대한 독립적인 통계 검증: 디플레이티드 샤프, 표본 외 및 누수 포렌식, 서명된 재생 가능 판정; 방법론 감사이며 투자 조언이 아닙니다.

문서

AlphaAssay — MCP server

Signal validation & overfitting forensics for AI agents. Most trading signals are noise. AlphaAssay puts them on trial — deflated Sharpe, out-of-sample, leakage forensics — and returns code-computed, fail-closed verdicts with machine-readable failure gradients. Methodology audits, not investment advice.

What this repo is. The public documentation and the registry manifest (server.json) for the AlphaAssay MCP server. The validation engine is a hosted, remote MCP — this repo is docs and metadata, not the engine source. The endpoint is live at https://mcp.alphaassay.com/mcp — the calls below are runnable.

Current machine-readable surface facts are published at https://api.alphaassay.com/v1/meta/facts; current prices and allowance terms are published at https://api.alphaassay.com/v1/meta/pricing.

What it does

Use this when you need to validate a trading signal, strategy or backtest before capital touches it: statistical signal validation, overfitting detection, deflated Sharpe ratio (multiple-testing aware), walk-forward and out-of-sample testing, look-ahead and data-leakage forensics, placebo tests. It returns a code-computed verdict — pass, conditional, fail or insufficient_evidence — with machine-readable failure gradients and provenance hashes. A mature preregistration can receive a separately issued Ed25519-signed certificate through the REST-only issuance operation; an ordinary verdict response is not itself a signed certificate.

It does not generate strategies and does not give buy/sell advice. It is a methodology audit. Verdicts are demote-only: evidence can lower a grade, never inflate one. Asset-agnostic — crypto, stocks, futures, FX.

Connect (MCP)

The server is a remote, streamable-HTTP MCP endpoint:

https://mcp.alphaassay.com/mcp

Point any MCP-capable client at it. Example client config:

{
  "mcpServers": {
    "alphaassay": {
      "type": "streamable-http",
      "url": "https://mcp.alphaassay.com/mcp",
      "headers": {"Authorization": "Bearer ats_your_key_here"}
    }
  }
}

No account is needed for the free tools (assay_demo, assay_graveyard, assay_calibration, assay_certificate_verify, assay_provider_protocol, assay_preflight). Paid MCP tools use prepaid API-key credits. For the hosted endpoint, configure the API key once in the Authorization: Bearer ats_... transport header; the adapter binds it without placing the secret in a tool payload. Omit the headers block when using only free tools.

Run it locally (stdio clients, Docker)

This repo ships a real local stdio server (server/index.mjs, dependency-free, Node >= 18): it defines all 21 tools statically (server/tools.json, versioned here) and executes each call against the hosted engine — the engine itself is server-side by design, so a local install can never be a locally editable examiner. Free tools run without credentials; paid tools take an optional ALPHAASSAY_API_KEY environment variable.

node server/index.mjs            # directly
docker build -t alphaassay-mcp . && docker run -i --rm alphaassay-mcp

Stdio client config:

{
  "mcpServers": {
    "alphaassay": { "command": "node", "args": ["/path/to/mcp/server/index.mjs"] }
  }
}

Try it over plain HTTP

REST and MCP are separate transport contracts. Some operations share the same execution semantics, but their available operations, validation bounds and request/response envelopes differ. The current mapping is documented at https://alphaassay.com/docs/api and generated from the canonical docs/product/TRANSPORT_PARITY.md contract. x402 pay-per-call applies only to REST POST /x402/v1/gauntlet; it is not an MCP authentication or payment path.

This REST example sends a golden specimen (a test signal whose correct verdict is known in advance):

curl -s https://api.alphaassay.com/v1/assay/demo \
  -H "Content-Type: application/json" \
  -d @examples/golden_lookahead.json

Expected response (this specimen is a whipsaw with no net-of-cost edge, so it dies at the first gate — abridged; the verbatim body is in examples/golden_lookahead.expected.json):

{
  "schema": "gauntlet.v1",
  "verdict": "fail",
  "died_at": "net_edge",
  "failure_codes": ["no_net_edge"],
  "family_id": "fam_004b8867ffc7a556",
  "stages": [
    {
      "stage": "net_edge",
      "verdict": "fail",
      "evidence": { "net_sharpe_annualized": null, "net_return_total_pct": 0.0, "trades": 0, "bars": 199 },
      "detail": "No positive net-of-cost edge -- dies before any deeper test."
    }
  ],
  "budget": { "cumulative_n": null, "n_trials_effective": null, "note": "..." },
  "_meta": { "engine_version": "024739a", "api_version": "v1", "as_of": "<varies per call>" },
  "disclaimer": "Consolidated research reality-check. NOT financial advice, no order path. Demote-only."
}

stages also carries funding_edge, family_deflation, power_honesty, significance, cpcv, walk_forward, concentration, placebo, capacity and graveyard_prior. Field-by-field explanation: https://alphaassay.com/docs/verdicts. Runnable scripts for all four specimens: examples/.

Golden specimens (free known-answer tests)

Four prepared signals with a planted property and a known correct verdict — the standing offer to catch AlphaAssay being wrong before you pay it.

specimen	planted property	expected verdict
`golden_clean`	a genuine, persistent edge	`pass` · `died_at: null`
`golden_lookahead`	a whipsaw with no net-of-cost edge	`fail` · `net_edge` · `no_net_edge`
`golden_cherry`	best-of-many parameter cherry-pick	`fail` · `family_deflation` · `deflated_out_at_n=50` + `BACKTEST_TOO_SHORT_FOR_N=50`
`golden_thin`	too little data to judge	`insufficient_evidence`

For these pure known-answer specimens, assert on the stable semantic fields verdict, died_at and failure_codes in CI. Timestamp/provenance fields such as _meta.as_of and generated_at can vary. Catalog: https://alphaassay.com/docs/specimens.

Tools

tool	what it does	tier
`assay_demo`	free demo assay + the four golden specimens with known verdicts	free
`assay_signal`	full statistical battery on a signal → verdict + failure gradients	paid
`assay_forensics`	WHY it fails: leakage forensics on decision timestamps — look-ahead, front-loading, a three-dimensional placebo	paid
`assay_backtest`	overfitting forensics on backtest results (deflated Sharpe, PBO/CPCV)	paid
`assay_gauntlet`	the whole battery in one call — every graded stage chained into one dossier that names the first gate that killed it	paid
`assay_falsify`	eight adversarial attacks (execution lag, cost stress, regime split, synthetic-null worlds, drift-burst strip, …) → the survival map	paid
`assay_preflight`	lint a payload's form (DSL schema, OHLCV sanity, trade rows) before spending a paid check	free
`assay_batch`	up to 25 sweep variants in one call, each a family-ledger trial under cumulative deflation — counts, never a ranking	paid
`assay_reproduce`	independently recompute a claimed track record from trades + candles; grades each headline claim against disclosed tolerances	paid
`assay_tradelog`	interrogate a raw fill log against itself — duplicate fills, time travel, PnL contradicting its own prices; no candles needed	paid
`assay_cpcv`	the full purged combinatorial CV distribution behind the gauntlet's cpcv stage — quantiles, histogram, recoverable path count	paid
`assay_survivors`	Romano-Wolf stepwise FWER over a T×N sweep matrix — an error-budget disclosure, never a ranking	paid
`assay_pbo`	Probability of Backtest Overfitting over your full parameter-sweep trial matrix (CSCV, purged) — grades the selection process itself	paid
`assay_register`	commit a hypothesis hash and terms in tenant state now, queue later daily operator-published Merkle inclusion, then forward-evaluate post-cutoff data; independent time evidence requires a separate pinned external anchor	paid
`assay_verdict`	evaluate a mature preregistration on post-cutoff candles under its sealed terms	paid
`assay_var_es`	exceedance backtest of ex-ante VaR/ES forecasts — exact binomial Basel traffic light + joint (VaR, ES) e-process	paid
`assay_conformal`	coverage audit of prediction intervals against realised outcomes — Binomial or Beta-Binomial when a split-conformal calibration size is disclosed	paid
`assay_graveyard`	anonymised mortality statistics per structural signal family — has this idea already died here?	free
`assay_certificate_verify`	hosted verification of a signed certificate against current platform trust and revocation state	free
`assay_provider_protocol`	the falsification protocol as machine-readable rules — seven tests for judging ANY signal provider, AlphaAssay included	free
`assay_calibration`	bucketed mature-registration counter with explicit `accumulating` / `insufficient_history` scope; signed only when snapshot trust verifies	free

Full tool set (21 assay_* tools) and reference: https://alphaassay.com/docs.

Pricing

Paid MCP operations use the billing unit published by the live registry. Most are per completed check; assay_batch meters each valid processed variant. Within a unit, pass and fail cost the same because you are buying the trial, not a flattering outcome. The six free MCP tools listed above never cost anything. Current numbers, units and allowance terms: GET https://api.alphaassay.com/v1/meta/pricing.

Verify a certificate yourself

A mature preregistration can receive a separately issued Ed25519-signed certificate through the REST-only certificate issuance operation. MCP exposes assay_certificate_verify for verification, not certificate issuance; the free demo verdict is unsigned. Stateful calls expose their effective timestamp, and only an idempotency-key replay of a stored response promises byte identity.

The hosted verifier at https://alphaassay.com/verify checks current platform trust. A raw signature check can run offline with alphaassay-verify signature-only, but raw signature validity alone does not establish platform trust. Full offline trust verification requires a signed trust bundle, an independently pinned root, and complete retained key/revocation history. See https://alphaassay.com/docs/verify-offline.

Docs

Quickstart — https://alphaassay.com/docs/quickstart
Rules for your trading agent (copy-paste) — https://alphaassay.com/docs/ai-rules
Golden specimens — https://alphaassay.com/docs/specimens
Verdicts & the gradient — https://alphaassay.com/docs/verdicts
llms.txt — https://alphaassay.com/llms.txt

Publishing

Published in the official MCP registry as com.alphaassay/mcp; also listed on Glama, Smithery and the public MCP directories. Registry releases track this repository.

Licence

Apache-2.0.

Disclaimer. AlphaAssay is a methodology audit, not investment advice. No order path, no custody, no buy/sell signals. A pass means a signal survived a statistical trial most signals fail — never a promise of any outcome. Verdicts are demote-only.