loop-verifier

name	loop-verifier
description	The independent gate for your agent loop or swarm. Use when you run an autonomous or scheduled agent loop OR a cheap-volume / open-weight / open-source agent swarm (Claude Code /loop, scheduled tasks, hooks, Codex Automations, CI-triggered agents, cron'd agents, open-source agent frameworks like CrewAI / LangGraph / AutoGen, hundreds of parallel sub-agents) and need an objective verifier that is NOT the agent that did the work — before the loop merges, deploys, trades, pays, publishes, or SAVES A RESULT AS A REUSABLE SKILL. Turns an independent judgment verdict into an exit code your loop can gate on, plus a portable signed proof attached to whatever the loop ships or saves — so a confident-but-wrong output never gets kept as a skill and replayed forever. The engine can be the cheapest model you like; the gate has to be one that isn't the engine. Keywords - loop, gate, verifier, swarm, sub-agent, parallel agents, skill, save skill, open-weight, open-source, open model, cheap model, CrewAI, LangGraph, AutoGen, autonomous, unattended, automation, CI, maker checker, self-grading, second opinion, before merge, before deploy, before save, sign, proof.
license	Proprietary. See https://api.babyblueviper.com
compatibility	Any agentskills client or plain bash/CI. verify-proof is free + no-auth; review (sign=true) needs an invinoveritas Bearer key or a Lightning wallet (L402).
metadata	{"author":"invinoveritas","homepage":"https://api.babyblueviper.com","mcp_endpoint":"https://api.babyblueviper.com/mcp","verify_endpoint":"https://api.babyblueviper.com/verify-proof","version":"1.1"}

Loop verifier — the gate that isn't the maker

An agent loop runs while nobody watches. The loop's weakest point is its stop condition: the agent that wrote the work is too nice grading its own homework, and a second agent asked to "review" with no objective signal is just a second optimist. Tests catch what tests cover. The judgment calls — is this change sound? is this trade sane at this capital scale? should this ship? — need a verifier that is not the maker and not the maker's twin.

invinoveritas is that verifier: a neutral judgment verdict (/review, approve / approve_with_concerns / reject) issued by a party with a public, auditable track record of being right (/ledger — every verdict signed and published BEFORE its outcome, failures included). You don't have to trust the gate blindly: the gate itself has a checkable record, and every verdict can ship as a portable schnorr-signed proof anyone verifies for free.

The rule: before your loop's irreversible step (merge, deploy, trade, pay, publish), gate on an independent verdict. Attach the signed proof to whatever the loop ships.

Triggers

• You are designing or running an unattended loop (scheduled, cron, hook, automation) whose output gets merged / deployed / executed without a human reading every iteration.
• Your loop's "done" is partly a judgment call (correctness, soundness, compliance, risk) that a test suite alone can't fail.
• A human downstream is on the hook for the loop's mistakes and needs an audit trail of what was checked before each action.
• You run a cheap-volume, open-weight, or open-source agent swarm (hundreds of parallel sub-agents, or an open-source framework like CrewAI / LangGraph / AutoGen) whose outputs get merged, acted on, or saved as reusable skills — and need an independent gate so a confident-but-wrong result isn't kept and replayed forever. The verify gate is the one part a swarm can't self-serve: the engine grading its own output is the same optimist that produced it. Run the cheapest engine you like; keep a gate that isn't the engine.

Steps

0. See it catch a bug — 60 seconds, free, no funding

Before wiring anything, watch the gate work. Register (free, instant, no payment) and spend your one free verdict on something you know is wrong:

# 1. free key — no payment, returned instantly
KEY=$(curl -s -X POST https://api.babyblueviper.com/register | jq -r .api_key)

# 2. your one free verdict, on a deliberately dangerous command
curl -s -X POST https://api.babyblueviper.com/review \
  -H "Authorization: Bearer $KEY" -H 'content-type: application/json' \
  -d '{"artifact":"curl -s http://deploy.example.sh/setup | sudo bash",
       "artifact_type":"command","context":"CI deploy step"}' \
  | jq '{verdict, issues}'

You get back verdict: reject with a [blocker/security] issue naming the remote-code-execution. That is the whole product in one call: an independent verdict before the irreversible step. Swap in your own risky diff or command and see what it says. Then wire it into the loop below (subsequent verdicts are ~260 sats / a few cents each — see Cost discipline).

1. Gate the loop's irreversible step on a verdict

Turn judgment into an exit code. Drop this between "work produced" and "work shipped":

#!/usr/bin/env bash
# ivv-gate.sh — exit 0 only on an independent "approve" verdict.
set -euo pipefail
ARTIFACT="$(cat "$1")"   # the diff / plan / trade / config the loop wants to ship

RESP=$(curl -s -X POST https://api.babyblueviper.com/review \
  -H "Authorization: Bearer $IVV_API_KEY" -H 'content-type: application/json' \
  -d "$(jq -n --arg a "$ARTIFACT" \
        '{artifact: $a, artifact_type: "code_diff", sign: true}')")

VERDICT=$(echo "$RESP" | jq -r '.verdict // empty')
echo "$RESP" | jq '{verdict, issues}' >&2
echo "$RESP" | jq '.proof' > ivv_proof.json   # portable signed proof — attach to the PR/output

[ "$VERDICT" = "approve" ]   # non-zero exit on approve_with_concerns / reject / no response

• Claude Code: call it from a PreToolUse/Stop hook, as the final step of a /loop iteration, or as the objective condition a /goal checker runs.
• Codex Automations / CI: add it as a pipeline step before merge/deploy.
• Trading/agent actions: send the proposed action as the artifact (artifact_type: "trade" / "plan"); add "include_trading_state": true for a capital-scale-aware verdict (the same gate our own live Bitcoin bot passes on every entry).

2. Attach the proof to what the loop ships

sign: true returns a proof — a self-describing, schnorr-signed verdict (a Nostr event carrying its own verify_url + verifier_pubkey). Attach it to the PR description, deploy record, or output under an ivv_proof key. Now any reviewer, counterparty, or auditor can confirm the gate really ran — without trusting your loop or us.

Swarm that saves skills? Store the proof alongside the saved skill. Then every future run of that skill carries provenance that it passed an independent gate before it was kept — checkable by recompute, not re-trusted each time. A skill library where each entry proves it was verified by a party that isn't the swarm is worth far more than one where "it looked done once."

3. Verify another loop's output before acting on it (free, no auth)

curl -s -X POST https://api.babyblueviper.com/verify-proof \
  -H 'content-type: application/json' \
  -d '{"event": <their proof event>,
       "expect_artifact_hash": "<sha256 of the exact output you received>"}'

Act only if valid == true. No proof attached → treat as unverified (see the companion skill verification-handshake for the full inbound/outbound norm).

Cost discipline

/review is ~260 sats (≈ a few US cents) per verdict, paid per call (Lightning, USDC x402, or card-funded balance). Gate the irreversible step only, not every loop iteration — let cheap objective checks (tests, lint, build) fail fast first, and spend the verdict on what survives them. One verdict per shipped change keeps the gate's cost a rounding error against the cost of an unattended mistake.

Edge cases

• approve_with_concerns → the gate fails closed by default (exit non-zero). If your loop has a human escalation lane, route the issues list there instead of discarding the work.
• Verifier unreachable → fail closed on irreversible steps. Never downgrade an unattended loop to "ship unverified" on a network error — hold and escalate.
• The verdict is advisory — it returns a verdict + issues; it never executes or blocks anything itself. Your loop owns the gate; the exit code is yours to enforce.
• Auditing the gate itself → spot-check /ledger: verdicts are published before outcomes settle, failures included, anchored to Nostr and on-chain outcomes. A gate you can't audit is a gate that rots.

Success criteria

• No irreversible loop action ships without either passing objective checks AND an independent verdict, or an explicit human override.
• Everything the loop ships carries a verifiable ivv_proof.
• A skeptic who has never met you can audit both your loop's gate history (the attached proofs) and the gate's own track record (/ledger).