Architecture

The picture in one line: your AI can ask to pay — a checkpoint you run, not the model, decides what clears. In fuller terms: IronBridge is a self-hosted deterministic checkpoint below your agent's model — an off-LLM gate that refuses unpaid or forged x402/MCP calls today, plus a bounded signer proven un-drainable in test (currently disarmed — it does not guard live funds today). Every action is hash-chained; the chain's head is committed nightly to Base.

The primitive

The model can ask. It can't pay.

Keeping the key away from the model is table stakes now — Turnkey, Privy and Coinbase do it inside their clouds. Two things set this apart: the checkpoint runs in your own Cloudflare account, not a vendor's, and every decision lands on a public hash-chain whose linkage anyone can re-walk — not a private compliance log. Row contents stay operator-attested; we mark prove-vs-trust. A compromised prompt has nothing to exploit.

The AI agent
can ask to do something — nothing more
The checkpoint
checks the request against rules the model can't see or change
Allow or block
safe → it happens · risky → handed to a human
▲ the AI can't reach your rules, your key, or the checkpoint itself — they live outside its reach
For compliance & risk leads

Deterministic compliance analyzers

Off-LLM, rule-based analyzers — same inputs, same output, every time. Built and runnable; not yet integrated with the live gate/registry. If you own SOC2 / audit / regulatory sign-off, this is the lane to talk to us about.

Securities (Howey) analyzer

Deterministic Howey-test pass over a token/offering's facts → structured risk read, not a vibe.

▌ built · deterministic · not wired to live registry

Jurisdictional analyzer

Maps an action to jurisdiction-specific constraints so a risk lead sees the exposure before it ships.

▌ built · deterministic · not wired to live registry

Compliance-readiness runbook

Plus a tax-strategy catalog. The audit chain re-walk is the evidence trail underneath all of it.

▌ built · deterministic · not wired to live registry
The reputation layer — IronBridge × Net Protocol

Net stores who-said-what. IronBridge proves it.

Net Protocol is a permissionless on-chain feed for agents on Base. By design an app can set a message's sender to any address — Net stores a genuine message and a spoofed one identically, because it never proves the asserted sender actually performed the action or that the result is real. For agent messages tied to x402 payments, that's a hole. An IronBridge agent-call receipt closes it — carried inside Net's data field, it binds the x402 payment to the exact result hash, sealed into the same public LAW-25 chain you can re-walk yourself. This is the keyless neutrality applied to reputation: no key of ours signs anything.

  the one invariant Net can't enforce
msg.sender  ==  receipt.payer        // asserted sender == wallet that actually paid
recompute response_sha256, digest, leaf  // re-walk the public chain yourself
any step fails  ->  reject the message  // a spoof can't bind; an honest call does

If the receipt binds and re-walks against the public chain, the sender is proven and the result is committed — with zero trust in the Net app or in IronBridge. A verified receipt can then feed Net's on-chain reputation (Net Score), so honest, paid agents accrue reputation that anyone can re-verify rather than take on faith.

GENESIS POSTED · NOT YET WIRED LIVE   IronBridge has posted a real genesis message to the Net core on Base mainnet, and the bind/verify adapter is proven against both a genuine and a spoofed message in an operator self-test. It is not yet wired into the live lane serve path, and no external agent has posted a paid receipt to Net yet — the integration is built and re-walkable, the traction isn't claimed.

For agent builders

Where it sits in the stack you already use.

You already gave your agent reach and a wallet. Policy layers above the model (e.g. NeMo Guardrails) steer what the model intends; IronBridge sits below the model and bounds what can actually execute. It composes with your stack — it doesn't replace it.

Above-the-model policy

Guardrails and policy prompts shape what the model decides to do. Strong for steering tone and intent — but they live inside the model's trust boundary. If the prompt is jailbroken, the policy can be talked around.

trust the model wasn't tricked
IronBridge · below the model

A deterministic gate the model can't reach or rewrite. The model can ask for anything; the gate decides what can actually move. A jailbreak changes what's asked — not what's allowed.

model can be tricked & still can't exceed bounds

A structural money-gate

A bounded signer, proven un-drainable in test: an agent can request a payment, but a capped allow-list gate decides what actually clears on-chain. Built and proven in test (fail-closed, kill-switch, per-move and daily caps), currently disarmed — it does not guard live funds today.

▌ proven in test · arming gated · does not guard live funds today

A LAW-25 audit chain

Every gate decision is committed into a hash-linked chain — each row carries the previous row's hash. You don't trust our logs; you re-walk the linkage and confirm no row was altered or reordered after the fact.

▌ live · re-walkable continuity on every call

Composes with your stack

It sits beneath whatever gives your agent reach and a wallet, and beneath whatever policy layer steers the model. IronBridge adds the structural floor and the audit trail — it doesn't ask you to rip anything out.

▌ off-LLM gate · self-hosted in your own account