Content revised:
Detection methodology and data definitions
Understand what the data represents, what it cannot prove, and how to review it alongside timestamps, samples and your own tasks. See the detection guide for operating steps.
On this page
What a check can and cannot establish
Hvoy AI uses real requests and observations of consistency across several aspects of responses to flag service behavior worth reviewing. Checks cover different capabilities; read the actual report. A single field or a model's self-description is not conclusive.
Results are not official certification, exact identification of every model version, a compliance audit or a long-term quality guarantee. The detection guide explains operation and report examples; this page explains meaning and limitations.
Connectivity, uptime and latency
Connectivity indicates whether a request can complete, not model authenticity. Uptime comes from saved probe samples in the stated scope, not a promise of availability for your business. With no valid samples, unknown does not mean zero.
Public mean uptime first combines channels within a site by sample count, then averages independent sites with valid samples equally, so a site with more channels does not directly dominate. Latency is elapsed time for the relevant test request, not a substitute for your time to first token, long-task speed or local network experience.
Normal results, warnings and sample limits
A normal result describes the tested time, model and channel. A warning highlights an observed inconsistency for review, not a final judgment of the provider. Connectivity-only or quote-only records cannot establish authenticity, and an unfinished check is not a pass.
Sample size, task coverage, upstream updates, shared resources and channel settings all limit conclusions. Review under comparable conditions using a limited number of repeat observations at different times and sanitized real tasks. Do not generalize one normal or abnormal result to every user.
Displayed, qualified, monitored, priced and listed sites
Displayed sites are independent sites covered by public organic channels. Qualification counts sites meeting the relevant measurement criteria, not necessarily sites with fresh recommendations. Monitoring counts the union of sites probed in the recent 24 hours and publicly displayed sites; some identities appear only in aggregate counts.
Priced sites have publicly displayable quotes, while directory entries can include sites without current public prices. These sets need not match. Counts deduplicate independent site records: matching names or domains are not automatically merged, and multiple channels of one site are not multiple sites.
Statistical windows and four kinds of time
Rankings use saved 24-hour statistical snapshots, not a freshly calculated rolling window on every page load. Actual probe time says when a request ran; quote observation time says when a price was recorded. They can differ.
A data record's update time can also change because of pricing or other edits. Page generation time only says when content was assembled, not when a probe ran. Read the times shown on each page rather than assuming one site-wide refresh schedule; missing or unreliable timestamps must not become the current time.
Comparable prices and cost boundaries
A price is an observed public quote for the stated model, channel, currency, unit and tier, not a settlement guarantee for every account. Missing prices are not zero, and per-call, duration and token rates are not interchangeable. An input-price comparison does not include all output, cache or retry charges.
Keep platform reference units, the purchase cost of balance and group multipliers separate. Subscription permissions are not cash API credit. Reconcile with the charging provider's rules and actual debits; this page and model checks are not full billing audits or fixed-cost promises.
Advertising and organic results
Advertising placement and organic channel measurements are different information. Sponsored display rows do not enter organic statistics, but an advertiser's organic channel may qualify under the same criteria. Advertising is not evidence of passing a check.
These are measurement definitions, not a promise that payment has no effect on placement or exposure across every page. Neither rankings nor advertisements are official endorsements, custody of funds or refund guarantees. Consider labels, evidence and your own trial together.
Reviewing evidence and limiting exposure
Keep the model, group, time and sanitized errors, compare the same channel on comparable tasks, and send reproducible issues through the contact page. Share only necessary details, not full keys, personal payment credentials or production secrets. This page makes no unverified āno logsā promise.
Use a restricted test key, the smallest necessary top-up and an independent backup. Reconcile charges and revoke keys when appropriate. Normal model behavior does not ensure a provider will keep operating or replace your assessment of privacy policies, authorization, contracts and financial exposure.