All projects matching “reasoning”

Twenty-Four daily math puzzle — four numbers to combine into exactly 24; deterministic solvable UTC daily challenge for agent reasoning benchmarks.

Rug check for Arc tokens: a 0-100 safety score with the reasons, for a token or a Uniswap v4 pool. One call, one answer: should you buy this Arc pool? Pass ?pool=<v4 pool id> or ?token=0x… and get a 0-100 safety score with the reasons, fusing the pool fee from its key (on Arc the commonest trap: some pools take 49-93% of every trade), our launch watcher's real buy-then-sell round trip and what it

Agent Guild paid read (fraud_check): collusion suspicion scores, reasons and cluster identifiers Use when an autonomous agent needs to: list agents at or above a minimum collusion suspicion score; review flagged agents and their collusion suspicion reasons.

FDA drug recalls and food recalls: the latest FDA enforcement reports with recall class (Class I is most serious) explained in plain language, recalling firm, product, reason, status, dates and distribution. Filter by class, company and time window. openFDA data, not medical advice.

The scorecard plus Messy Virgo's analyst verdict: one overall verdict (strong signal, watch, weak or red flag) with its reason, a written summary across security, performance and social, the key risks and what to watch, grounded in the 118 KPIs that ship with it, incl. honeypot, liquidity and sentiment. Call it when you need a decision-ready answer on a token: pre-trade checks, watchlist triage, p

AI provider health reading from public status pages (currently GitHub): degraded (boolean), level, one reason, per-provider indicator, asOf and sources with attribution. openai and cursor keys return null until 2026-11-12; the anthropic key returns null (no longer measured since 2026-10-20). Not af…

Resolves one public name on Wikidata and returns a relationship graph from a fixed property list (instance of, country, headquarters, founder, chief executive, ownership, employer, education, membership, and similar), up to two hops. People include only notable public roles. If nothing matches, the graph is empty and reason explains why.

Re-check up to 5 tokens you hold or watch in one call, on Base, Ethereum, Arbitrum, Optimism, Polygon or Solana: each token's 0-100 risk score, level, reason codes, price, liquidity and a fingerprint. Send the fingerprints back as since to get what changed (new risks, owner or authority changes, liquidity or price moves of 10%+) and a short list of alerts. One price for the whole list, made for po

Checks one email address and returns valid, risky or invalid with the reasons: address format, whether the domain can receive mail (live MX lookup), known throwaway domains (mailinator and others, invalid), shared inboxes like info@ or sales@ (risky), likely typos such as gmial.com with the fix in didYouMean, and free providers such as gmail.com (flag only). Does not ping the mailbox (no SMTP), so

URL safety verdict: follows the redirect chain, detects homograph and punycode hostname tricks, and cross-references the destination host against the URLhaus threat feed. Returns safe, suspicious or malicious with reasons. Single-URL scope; domain-wide grading is /sitegrade and registration facts are /domain-intel.

Spot Check. REQUIRES ?host= on the paid request — a signed request without it is refused before the gate and no money moves. Name a host and get what this observatory already holds on it, signed: corpus rounds and verdicts as recorded, when we last actually knocked, our coverage of the window since we met it, and the gaps with their reasons. Read from the books at the counter — no request is made

Simulate an EVM transaction before signing. Decodes the calldata into a named function with arguments and security risk flags, AND runs a keyless static eth_call preview to tell you whether the transaction would succeed or revert (with the decoded revert reason). Use it to catch a failing or malicious transaction before you broadcast it. Read-only, no keys, no node infra. Asset-level change tracin

Tax-ID / VAT-number validation for Pan-African countries: valid/invalid, normalised form, id type and the reason it failed. Format plus the official check-digit algorithm where the tax authority publishes one; every rule cited to an official source with its date. Pure algorithmic, no registry lookup

Tax-ID / VAT-number validation for European countries: valid/invalid, normalised form, id type and the reason it failed. Format plus the official check-digit algorithm where the tax authority publishes one; every rule cited to an official source with its date. Pure algorithmic, no registry lookup

Tax-ID / VAT-number validation for North American countries: valid/invalid, normalised form, id type and the reason it failed. Format plus the official check-digit algorithm where the tax authority publishes one; every rule cited to an official source with its date. Pure algorithmic, no registry lookup

Tax-ID / VAT-number validation for Asian countries: valid/invalid, normalised form, id type and the reason it failed. Format plus the official check-digit algorithm where the tax authority publishes one; every rule cited to an official source with its date. Pure algorithmic, no registry lookup

Tax-ID / VAT-number validation for Latin American countries: valid/invalid, normalised form, id type and the reason it failed. Format plus the official check-digit algorithm where the tax authority publishes one; every rule cited to an official source with its date. Pure algorithmic, no registry lookup

Tax-ID / VAT-number validation for Middle Eastern countries: valid/invalid, normalised form, id type and the reason it failed. Format plus the official check-digit algorithm where the tax authority publishes one; every rule cited to an official source with its date. Pure algorithmic, no registry lookup

Tax-ID / VAT-number validation for Australasian countries: valid/invalid, normalised form, id type and the reason it failed. Format plus the official check-digit algorithm where the tax authority publishes one; every rule cited to an official source with its date. Pure algorithmic, no registry lookup

Virtuals Solana DBC State & Supply Receipt. Virtuals Solana launches only. Returns finalized mint controls, ACTIVE_DBC vault custody, tokens outside the observed DBC vault, platform lifecycle, signature/slot provenance, and token-specific tape coverage. Every result below COMPLETE names its reasons and separates our own instrumentation age from real observation loss. It does not verify circulating

Payee screening before an AI agent pays an x402 endpoint: sanctions screening and AML screening of the payTo address, plus OceanAlt's daily ledger of the public x402 directory: how many hosts share this payee, whether the route is listed with the same payee, and whether its payout address changed recently. Returns allow / review / decline with reasons.

Is this x402 endpoint safe to pay? Call it before your AI agent pays an endpoint it has not used before. Returns a go / caution / no_go verdict, the recommended payment option (network and amount) and the reasons (payment would fail, payout address on the OFAC sanctions list, over your budget, price above advertised, not HTTPS, unknown token). Never pays the endpoint. Cached 10 min. Bulk: 1000 cre

Use when an agent is about to pay an x402 endpoint, seller domain or wallet it has not used before. Returns ok_to_pay, caution, dont_pay or no_payment_needed with reasons: live payment terms, its Coinbase Bazaar listing and paying wallets in 30 days, whether the live pay-to address matches the listing, real-USDC check, price vs. similar tools, and seller domain age. Read-only: never pays or signs.

Screens a person or business name, or an NPI, against the HHS-OIG exclusion list (LEIE, Medicare and Medicaid): potential matches only, each with name score, reasons, NPI, state, specialty, exclusion type and date. Not a determination: confirm a match on the OIG online search with the SSN or EIN. Empty candidates is a paid 200. Not for employment or credit decisions. List dates and code meanings:

[Payment decision: spend policy + seller reputation risk check, one verdict] Right before an AI agent pays: POST {payment, policy, seller} -> ONE verdict pay / dont_pay / ask_human. Runs the spend-policy check (limits, rolling cap, categories, allow/deny, hours, assets/chains, approval threshold) and the x402 seller reputation check (buyers, wash share, age, concentration), with reasons and one si

Verdix: call before an agent sends funds to an EVM address on Base. Returns safe/caution/danger with reasons from OFAC sanctions, scam/phishing lists, address-poisoning lookalikes, burn-address and contract/deployer checks; quick adds address age, standard/deep on-chain behaviour. "safe" is not a guarantee. Legacy route that sells 3 tiers (a default client buys quick); new integrations should call

Is this transaction, approval or signature safe to sign? Check it before your AI agent or wallet signs it: pre-sign verdict (green/orange/red + reason codes) for EVM and Solana transactions, token approvals and Permit/Permit2/EIP-3009/Seaport signatures. Screens the spender or recipient wallet for OFAC sanctions and scam, phishing or theft reports, and flags unlimited allowances, plain-wallet spen

Request an https URL and report whether it is reachable, its status, every redirect hop with timings, key and security headers (HSTS, CSP and others), and the TLS certificate with its issuer, the names it covers and the days until it expires. An unreachable URL is a result too, with the reason.

Token risk / token safety check for Base and BSC (BNB Chain) tokens: honeypot detector via live buy/sell simulation (honeypot.is, cross-checked with GoPlus when available), buy tax / sell tax, contract security (mint, pause, blacklist, proxy, owner renounced), holder concentration, liquidity, pool age, LP burn. Returns verdict low/medium/high/honeypot, 0-100 risk score, action and plain reasons. R

Answers questions (single-turn question answering) from a small local LLM (qwen3.5 2B). Best for general knowledge, definitions and simple reasoning. POST JSON {"input": string}, up to 3000 characters; returns {answer, model, elapsed_ms}. Answers are brief (2-3 sentences). Typically 5-15 s. Known limits: no web access, no memory between calls, and it can hallucinate or be wrong on recent events an