Agentic Finance Graph

Quality 72 / 100needs work

25 tools, about 5,820 tokens to load them all · tool set 11726cd6956a · tools last changed 2026-10-16 · checked 2026-10-19 22:14 UTC

Tool changes

• 2026-10-16canary_statusadded

  Every canary you registered, with the watcher's last reading (nonce, code, balance, checks) and the newest 20 events per canary (trigger, severity, evidence, whether you declared it a test, the reading in plain words and what to do). Needs Authorization: Bearer <Desk session>. Labels are your own text and come back in the untrusted envelope.

• 2026-10-16canary_registeradded

  Register the public part of a decoy you planted: kind address|key|url|apikey|tool|x402 with exactly one target (address, url_sha256, key_sha256, tool_name or resource_path), a label, the triggers to watch (nonce, code, permit, siwe, inbound_transfer, http_hit, api_use, tool_call, x402_request) and planted_at. Needs Authorization: Bearer <Desk session>; the session's wallet is the operator. Private material is refused before validation; a bait key must be fresh (nonce 0, no code). A canary is a decoy you plant where only an intruder or a misled agent would read it: a bait key that never holds a balance (its first use is the signal), an address planted as a treasury, a URL, an API key, a tool description or a paid x402 resource. Register the PUBLIC part only (address, hash, tool name, path); we refuse anything that looks like a key, a URL or an API key in clear. We read the chain every 2 minutes (nonce, code, gas funding, inbound transfers, permits) and your own channels when you push them (canary ingest). A fire is a pattern for a human to read, written to your canary log; nothing is published about a plant. Shadow period: at most 25 live canaries per operator, no e-mail yet.

• 2026-10-16blast_radiusadded

  For an agent (L8 or above) or one of its bound wallets: priced stablecoin and token balances and ETH in USD at one block, the EIP-7702 delegate if any with its code hash and known sweeper family, and the open ERC-20 allowances the wallet granted, confirmed on chain and capped at the balance per token (agent.blast_radius.v1). Read every six hours from free public nodes. A measurement of exposure at one block, not a prediction; unpriced tokens are listed by symbol.

• 2026-10-16agent_passadded

  What the ledger measured about one registration at its last evaluation under agent.pass.v2, valid three hours: published level L7+, no open harm-kind incident, a counted payment in the last 30 days, its newest payment in character, no issuer freeze on a bound wallet, no unlimited approval to a plain key, no harm-kind incident in 90 days, and a month without a payment out of character. Answers holds or not, which checks failed, which could not be taken (listed, never written as 0), when it was checked and when it expires, with the issuer's public key. In shadow mode nothing is signed yet. A Pass says what was checked and when; it never says safe, verified or trusted, and says nothing about the next three hours.

• 2026-10-16gate_checkadded

  One call before an agent pays: the payee facts, the agent's own context from our ledger (what it paid in the period, how often, whether and when it first paid this payee, how many new payees today), the out-of-character score of this payment as the agent's next, and a decision under a policy written in the AP2 mandate vocabulary (allowed_payees, amount_range, budget {max, period: daily|rolling_24h|rolling_7d}, agent_recurrence, execution_date) plus ours: new_payee {max, window_hours, max_new_per_day, step_up_ratio, step_up_days}, refuse_conditions {fact: block|warn|log|ignore} and signal {v2: {notable, major}}. Anything the evaluator does not understand fails closed. Named policies: log-only (default), balanced, recommended, recommended-v3, strict (their measured cost on 500 real agents is in the policy document). The answer is block / warn / log / allow with every reason, and is recorded in log-only mode: we never block anything; your client decides, and decides its own fail mode when we are unreachable.

• 2026-10-16payee_factsadded

  What the chain and our ledger say about an address an agent is about to pay (payee.facts.v1): its code kind (plain key, contract, EIP-7702 delegated) and whether the delegate matches a known sweeper or the tiny-forwarder shape; whether it forwards most of what it receives within minutes (relay shape); whether its stablecoin outflow goes to swap venues; who funds it and where it exits (7 days of Transfer logs); whether a contract exposes a claim-on-proof selector; an issuer freeze; how many registered agents pay it; and, given the paying agent and the amount, whether it is a never-paid destination at a size the agent never sent and whether it is a look-alike of an address in the agent's own history. Each fact carries value, evidence, method and measurement time; an unavailable channel is said, never guessed. Facts describe an address; they do not judge it.

• 2026-10-16agent_patternadded

  The out-of-character score of an agent's newest counted payment against its own earlier payments (behaviour.out_of_character.v2: amount out of its distribution, a new payee at a new size, a never-paid plain key, a probe then a large payment, a look-alike address, an approval change, off its rhythm), with the reasons that fired, how many of its last 30 days of payments scored notable or major, and the market's rates. Levels are none / notable / major; a level is a pattern for a human to read, never a verdict, and in shadow mode nothing is blocked from it. Every figure carries its measurement time and definition id.

• 2026-10-16ranked_agentschangedoutput schema
• 2026-10-14verify_sum_logadded

  Trust nothing you did not recompute. Verifies the Ed25519 signature on the current sum-log head (the Merkle-sum root over every counted payment and the total it commits to), optionally climbs one agent's proof from its leaf to the root and the whole total, reports how many of the previous head's leaves the current head declares unchanged, changed, removed and added, and verifies the statement log's signed tree head and, given the head you kept (tree_size, root_hash), the consistency proof that our statement history was not rewritten. Keep the heads it returns and pass them next time. Definitions sumlog.v1 and agent.statement.log.v1.

• 2026-10-14ladder_dynamicsadded

  The whole-cohort document: cumulative incidence of the four competing exits (drained, abandoned, dormant, none) at 30, 60, 90 and 180 days with the numbers at risk and the restricted mean time without an exit (ladder.exit_incidence.v1); the Standardised Drain Ratio per minting address with 30 or more ranked agents, with 95% intervals (platform.sdr.v1); the day-to-day transition matrix between L7, L8, L9 and out (ladder.transition.v1); and the landmark re-check of the published survival figures (survival.landmark.v1).

Showing the latest 10. Full history: list_changes · trend: quality JSON.

Checks run on what our credential-free, read-only probes observe; tools are never called and no code audit is performed. How it is computed.