review
Read-only
Independent second-opinion governance verdict for AI agents before they commit an irreversible action — shipping code, running a shell command, or placing a trade. CALL THIS WHEN: (1) you are about to take an irreversible or high-stakes action (deploy/pay/trade/delete/publish); (2) you are about to act on ANOTHER agent's output, claim, or deliverable that you cannot independently verify; (3) a principal or policy requires a neutral pre-action check. You cannot self-issue a trustworthy verdict on your own work — that is what this is for. Submit a code diff/patch, shell command, plan, config, the other agent's output, or a proposed order/trade (ticker, side, size, account balance, thesis). Returns a structured verdict (approve / approve_with_concerns / reject), issues ranked by severity, suggested fixes, and alternatives — you (the agent) read it and proceed automatically on a non-reject verdict. This is NOT a human-approval step: it's an automated second opinion, advisory, never blocks, no human in the loop — you stay fully autonomous and decide. Capital-scale-aware. Model-agnostic. Built and dogfooded daily by our own autonomous fleet (the same gate we run our own important decisions through). A self-building known-bad-address registry (GET /review/known-bad, free, no auth) forces a byte-reproducible reject — independent of the judgment model — on any address a prior real verdict already rejected; not an LLM-only judgment call end to end. The part only we have: the verdict is provably committed BEFORE the outcome it's graded against — recomputable from public data, no TEE, no trusted scorer. Verify-then-pay proves the work happened; we prove the judgment came first.
Input Schema
TableJSON Schema
| Name | Required | Description | Default |
|---|---|---|---|
| sign | No | Return a PORTABLE SIGNED proof of this verdict (binds verdict + artifact hash + our pubkey), including a content-addressed decision_ref = sha256(JCS({artifact_hash, artifact_type, policy_version, verdict, source_class})). Attach it to your output so a downstream agent can confirm via verify_proof — WITHOUT trusting you or us — that invinoveritas issued this verdict for this exact artifact. The agent-to-agent trust handshake. For artifact_type=trade|onchain_action|sanctions_screening, the proof also carries source_class ('agent_reported' today — no mediation-point integration exists yet) and, when applicable, a vantage_limitation field disclosing that the verdict is occurrence evidence, not an absence/completeness claim — check it before treating an irreversible-class verdict as sufficient on its own. The proof also carries an opaque engine_generation counter (informational, not bound into decision_ref) — compare it across two proofs to detect a backend/judgment-engine change between calls without us disclosing which model we run. | |
| context | No | What you are trying to accomplish, why now, success criteria | |
| artifact | Yes | The artifact to review: unified diff / patch, shell command, plan, config, analysis, agent output, or raw text | |
| concerns | No | Specific things to check (e.g. 'production safety', 'edge cases in trading logic', 'regulatory risk') | |
| artifact_type | No | Type of artifact. 'code_diff' or 'patch' triggers deep code review. 'plan' for architecture/strategy. 'trade' triggers the capital-scale-aware risk-manager review of a proposed entry/exit. 'onchain_action' triggers the on-chain risk review of a proposed transfer/swap/approval/contract call (e.g. a Base MCP action) BEFORE you sign it — catches scam/honeypot tokens, unlimited-allowance drainers, address poisoning, slippage/MEV. 'sanctions_screening' for a compliance/AML result BEFORE acting on it — checks a categorical verdict (e.g. CLEAN) carries its own scope, not an unscoped claim. Tailors focus and suggestions. IMPORTANT for trade/onchain_action/sanctions_screening: a REJECT can happen purely from low confidence on an action you can't undo, even if content-wise the review leaned approve — see the response's reversibility_gate field. | general |
| action_binding | No | Optional: the exact real-world action this verdict authorizes — tool identity, materialized (not templated) arguments, and the id of the agent that will execute it, e.g. {'tool': 'place_order', 'agent_id': 'your-stable-agent-id', 'args': {...}}. v14+: `tool` and `args` are bound as SEPARATE preimage fields (action_binding_tool_hash = sha256(tool), action_binding_args_hash = sha256(RFC-8785-JCS(args))) so a verifier can assert 'same tool, different arguments' as a checkable statement; `agent_id` is bound as a plain string (action_binding_agent_id, not hashed). All bound DIRECTLY into decision_ref — so the verdict commits to the exact action, not just the free-text `artifact` argument or the verdict conclusion. Recomputing decision_ref without byte-identical tool/args/agent_id values produces a different hash: an approval cannot be replayed against a different tool, different materialized arguments, or a different agent. Every sub-key is optional. Max ~8KB JSON-encoded. NOT independently verified by us — we hash exactly what you send. HONEST LIMIT: nothing stops a caller from submitting an under-specified action_binding (e.g. tool+side but not size) and getting an approval reusable across the omitted dimension — that's about who controls what goes into the fingerprint, not how it's hashed. | |
| disclosed_summary | No | Only used when confidentiality_tier='partial_disclosure'. A real, human-readable description of the reviewed artifact/decision you're choosing to make public — bound raw into decision_ref. Ignored for other tier values. | |
| intended_audience | No | Optional: declare who/what this verdict is intended for (your own DID, endpoint URL, or gateway identifier). Bound into decision_ref so it can't be silently stripped or altered once issued. NOT independently verified — a reader compares this against their own identity and treats a mismatch as a signal the proof may be presented outside its intended context, a real context-binding replay-protection gap that earlier policy versions had no way to represent at all. | |
| intended_verifier | No | Optional: a CAIP-10 string naming the specific on-chain verifier/gate this verdict is meant to be checked against, e.g. 'eip155:8453:0x8004A169FB4a3325136EB29fA0ceB6D2e539a432'. Bound into decision_ref (itself inside the schnorr-signed content) so it achieves real crypto-level domain separation — the raw signed bytes otherwise bind only to our pubkey + content, nothing to a specific chain/contract, so a proof is technically replayable against any gate willing to accept it. NOT independently verified — a gate compares this against its own chain_id/address. | |
| severity_threshold | No | Minimum severity to report | all |
| related_proof_event | No | Optional: if the artifact being reviewed IS another party's already-signed verdict proof (a verdict-of-verdict re-review), pass that proof's full signed event ({id, pubkey, created_at, kind, tags, content, sig}). We independently re-verify it ourselves before its source_class can affect this call's own — capped, never upgraded (an independent_mediator call reviewing an agent_reported inner verdict stays agent_reported). Fails closed to agent_reported if the inner event doesn't verify, regardless of your own registry status. One hop only. HONEST SCOPE: we verify the cited event's own authenticity, not that it's actually the thing your artifact claims to be re-reviewing. | |
| request_capture_ref | No | Optional: a requester-controlled commitment (a hash/id you generated and can independently prove existed at request-time) that this artifact was submitted for review — the captured-admission-v0 review profile (trustless-ai/recompute-kit). Echoed back verbatim in the response's admission_receipt. NOT independently verified by us; closes the /ledger raw-tape-vs-published gap only for requesters who opt in. | |
| confidentiality_tier | No | Which privacy/evidentiary tradeoff this verdict should use, only meaningful with sign=true. 'hash_only' (default): the proof carries only artifact_hash, raw content never disclosed — strongest privacy, weakest standalone evidentiary value (a third party can't confirm what the hash corresponds to without your later cooperation). 'partial_disclosure': pass disclosed_summary, bound raw into decision_ref, so a third party gets real checkable context without full exposure. 'full_disclosure': records intent to publish this verdict to the public /ledger (full_disclosure_requested=true in the proof) — strongest evidentiary tier, but actual publication is still a separate curated step on our side, not yet fully self-serve. | hash_only |
| include_trading_state | No | Sentinel mode: inject live Sovereign Earner state (equity, regime, open position, PnL) for trading-related reviews. Highly recommended for any trading or risk decision. |