Recoil Sentinel

Inspiration

Everyone is shipping AI agents right now — coding agents, support agents, trading agents. But there's a quiet problem underneath all of it: an agent that works today can silently start being confidently wrong tomorrow. Someone tweaks a prompt, a model updates, the data shifts — and nobody notices until a customer does, or until a wrong number costs real money.

The industry's current answer is basically hope, plus a Notion doc of past mistakes. We wanted something stronger: a hard gate that sits between an AI agent and the real world and refuses to let it publish or act on anything it can't prove is true.

What it does

Recoil Sentinel is a verification gate for AI agents. Before an agent is allowed to publish a report or take an action, Recoil checks every factual claim it made against live ground truth — and blocks it if anything doesn't hold up.

We proved it works across three very different domains on one engine:

1. Crypto market intelligence — pulls live data, writes a brief with Claude, and verifies every number against the source before publishing a cited report to the public cited.md platform.
2. On-chain transaction integrity — reads a wallet's real state from the blockchain and blocks any claim (a balance, an amount) that doesn't match the chain. An agent that can't lie about money.
3. Agent/code regression prevention — the original Recoil CI gate: every fixed failure becomes a permanent test, and a regression blocks the deploy (exit code 1).

When a claim fails, the report is refused, nothing is acted on, and the failure is frozen into a permanent regression suite so the same mistake can never recur. It runs autonomously every six hours, is monetized with the x402 payment standard, and surfaces the whole verification workflow in a live dashboard.

How we built it

• Engine (Python): a provider-agnostic claim verifier, a deterministic ground-truth grader, a freeze-and-replay regression gate, and a Typer CLI.
• Ground truth: live CoinGecko + DefiLlama (market) and keyless Base JSON-RPC (on-chain), with Airbyte as a ground-truth control plane.
• Agent + judge: Anthropic Claude (Sonnet for generation, Opus for judging) with structured output and hard anti-hallucination rules.
• Publishing: verified reports go live on cited.md via the Senso content engine (we reverse-engineered the real API from the Senso CLI).
• Actions & payments: Composio opens GitHub issues autonomously; x402 (Coinbase) paywalls the premium report endpoint in USDC on Base.
• Data + serving: ClickHouse for real-time run analytics, FastAPI + SSE, a Next.js + Tailwind dashboard, deployed on Render with a built-in autonomous scheduler.

Challenges we ran into

• Making verification visible. Early on every run passed, so you couldn't see the gate doing anything. We built a fault-injection mode (--tamper) that plants a false claim so judges watch the system catch a lie and refuse to publish.
• The cited.md integration was a real investigation. The public docs had the wrong API host and endpoints. We validated our key with the Senso CLI, inspected the CLI's source to find the real API (apiv2.senso.ai, /org/content-engine/publish), then created a GEO prompt and the Cited.md destination to publish a genuinely live article.
• Proving generality. We deliberately made the same engine work for market data, on-chain transactions, and code regressions — different ground truth, same gate.

What we learned

The durable value of an AI system isn't the generation — that gets commoditized as models improve. It's the independent, deterministic verification around it, the audit trail, and the memory of past failures. As models get more capable, we trust them with bigger decisions, and bigger decisions need stronger guardrails — not weaker ones. Trust but verify scales up with capability.

What's next

A general "bring your own agent" verification gateway: point Recoil at any agent's output and any source of truth, define a policy, and it guarantees the agent can't publish or act on something false — and never repeats a mistake it's already made

Built With

• airbyte
• anthropic-claude
• base
• cited.md
• clickhouse
• coinbase-cdp
• composio
• fastapi
• next.js
• python
• render
• senso
• sqlite
• tailwind
• typescript
• x402