All projects matching “evaluation”

Structured evaluation with TypeSafe's Jev (System One) model. Send any state plus typed questions — noul (true/false), choice (pick one), or score (ordinal scale) — and get back calibrated probabilities and confidence for each. Built for routing, classification, triage and policy checks where a fast, predictable, machine-readable answer beats prose.

Structured evaluation with TypeSafe's Jev (System One) model. Send any state plus typed questions — noul (true/false), choice (pick one), or score (ordinal scale) — and get back calibrated probabilities and confidence for each. Built for routing, classification, triage and policy checks where a fast, predictable, machine-readable answer beats prose.

Structured evaluation with TypeSafe's Jev (System One) model. Send any state plus typed questions — noul (true/false), choice (pick one), or score (ordinal scale) — and get back calibrated probabilities and confidence for each. Built for routing, classification, triage and policy checks where a fast, predictable, machine-readable answer beats prose.

Publisher discovery repair plan fetches current public evidence from the supplied target and evaluates the documented criterion. It returns the observation, threshold, verdict, exact evidence source, limitations and a mode-specific action. The check is passive: no login attempt, payload injection, exploit, purchase, form submission or load test is performed.

TraceFast knowledge graph: what an Ethereum address is, as proven facts — wallet or contract, proxy and upgradeable or not, a token and its standard, and for a token contract whether it can pause, mint or blacklist and whether it ever did. Each fact names the block it was evaluated at; an absent fact is not proven. Ethereum mainnet only. Path address: 0x-prefixed, 40 hex. Attribute dictionary: htt

Fetch a site's robots.txt and evaluate crawl permissions: given a URL (plus optional extra paths) and a user-agent, return whether each path is allowed or disallowed, the rule that matched, the user-agent group, the crawl-delay and any declared sitemaps. Implements the Robots Exclusion Protocol (RFC 9309) with longest-match-wins, Allow-over-Disallow tie-breaking and * / $ wildcards.

Evaluates caller-supplied telemetry samples against declared reaction-time, actions-per-second, movement-speed, and impossible-event rules. Flags are review signals only; the service does not identify software, inspect a device, establish intent, or justify an automatic ban.

Independent pre-execution control for consequential autonomous-agent actions. Evaluate a proposed payment, tool call, send, write, publish operation, API action, or other external side effect before execution; return the required approval or human-review boundary and verify that risk stays within the declared limit. 3 bounded free preview calls are available via extensions.humanmirrorPreview.

Screen up to 3000 characters of untrusted text for prompt injection, secret exfiltration, deceptive payments and phishing. Jev risk scores, five attack families and an exact suspect passage. French, English and Spanish evaluation; probabilistic screening, not a security guarantee. Text is processed by the Jev provider; live credentials are outside the supported input scope.

Verify a claim or fact check a statement against public web sources: send a question and up to 5 URLs, get cited evidence - passages that support, contradict or fail to settle it. Claim verification with a citation for every excerpt: source URL, retrieval time, content hash. Never charges when nothing is retrieved. Free to evaluate: the `sample` tool returns one real response at no cost, so you ca

About page discovery repair plan fetches current public evidence from the supplied target and evaluates the documented criterion. It returns the observation, threshold, verdict, exact evidence source, limitations and a mode-specific action. The check is passive: no login attempt, payload injection, exploit, purchase, form submission or load test is performed.

One combined go/no-go before an agent spends: PROCEED / CAUTION / ABORT from the token's market verdict, the counterparty wallet's sanctions, age and poisoning check, and — with a spender — its ERC-20 approvals. A leg that cannot be evaluated returns unknown and caps the verdict at CAUTION. Mode 'compact' returns the decision, the worst leg, and one summary per leg. Use this when an agent is about

WhiteMagic hosted recall — agent memory for cross-session continuity: read-only hybrid search (BM25 + vector + associative) over a curated public corpus. Call when an agent needs durable context from prior sessions or the public knowledge store. Keyless discovery (initialize/tools/list) is free; free evaluation keys and a daily no-key trial are available. One paid tool call per payment.

Test whether a RE2 pattern matches anywhere in the input. Returns { matched }. RE2 guarantees linear-time evaluation, so this is safe against catastrophic backtracking. An invalid pattern is a clean 400; no match is a successful 200.

Send a list of domain names and this tool tells you exactly how old each one is, when it was first registered, when it expires, and how much time is left before expiration. Great for spotting brand-new suspicious domains, researching competitors, or evaluating domains before buying or linking to them. Works with over 100 domain extensions worldwide.

Parser-blocking script budget repair plan fetches current public evidence from the supplied target and evaluates the documented criterion. It returns the observation, threshold, verdict, exact evidence source, limitations and a mode-specific action. The check is passive: no login attempt, payload injection, exploit, purchase, form submission or load test is performed.

Verify a claim or fact check a statement against public web sources: send a question and up to 5 URLs, get cited evidence - passages that support, contradict or fail to settle it. Claim verification with a citation for every excerpt: source URL, retrieval time, content hash. Never charges when nothing is retrieved. Free to evaluate: the `sample` tool returns one real response at no cost, so you ca

Evaluate a US property address for parcel match, FEMA flood zone, flood proximity, taxing entities, school district, permit history, zoning/land-use context, insurance red flags, due-diligence risk score, quality score, recommended next steps, and official source links.

Ordinary-least-squares linear regression of y on x: slope, intercept, correlation r and coefficient of determination r². Pass predictAt to also evaluate the fitted line at a given x. Rejected as a clean 400 if lengths differ or x has zero variance.

No accidental wildcard CORS evidence record fetches current public evidence from the supplied target and evaluates the documented criterion. It returns the observation, threshold, verdict, exact evidence source, limitations and a mode-specific action. The check is passive: no login attempt, payload injection, exploit, purchase, form submission or load test is performed.

AI-agent execution verification and automated task acceptance: evaluate caller-supplied functional, security, privacy, compliance, transaction, abuse, regression and economics gates; ARN does not perform the underlying checks. Re-run for each release or evidence change; sustained CI users can switch to the 30-day Release Guard pass.

Pre-execution decision for a metered search or browse tool call. Evaluates the proposal; does not execute the external tool.

CMS detection confidence audit fetches current public evidence from the supplied target and evaluates the documented criterion. It returns the observation, threshold, verdict, exact evidence source, limitations and a mode-specific action. The check is passive: no login attempt, payload injection, exploit, purchase, form submission or load test is performed.

Canonical change monitor plan fetches current public evidence from the supplied target and evaluates the documented criterion. It returns the observation, threshold, verdict, exact evidence source, limitations and a mode-specific action. The check is passive: no login attempt, payload injection, exploit, purchase, form submission or load test is performed.

Expanded acronyms repair plan fetches current public evidence from the supplied target and evaluates the documented criterion. It returns the observation, threshold, verdict, exact evidence source, limitations and a mode-specific action. The check is passive: no login attempt, payload injection, exploit, purchase, form submission or load test is performed.

Add-to-cart action experiment plan fetches current public evidence from the supplied target and evaluates the documented criterion. It returns the observation, threshold, verdict, exact evidence source, limitations and a mode-specific action. The check is passive: no login attempt, payload injection, exploit, purchase, form submission or load test is performed.

Abbreviation expansion repair plan fetches current public evidence from the supplied target and evaluates the documented criterion. It returns the observation, threshold, verdict, exact evidence source, limitations and a mode-specific action. The check is passive: no login attempt, payload injection, exploit, purchase, form submission or load test is performed.

Conservatively repair common JSON syntax errors (unquoted keys, single-quoted strings, trailing commas) and parse the result; input is never evaluated as code.

Evaluate an SPF record against an IPv4/IPv6 address using only its literal ip4/ip6/all mechanisms. DNS-dependent mechanisms (include, a, mx, exists, redirect) reached before a match yield an honest `indeterminate` result rather than a guess.

Evaluate a mathematical expression with optional named variables — arithmetic, comparison and logic operators plus functions like sqrt, abs, min, max, sin, cos, log, floor, round, etc. A syntax error, an undefined variable, or a non-finite result is a clean 400. The factorial operator is disabled for safety.