# Understand the network

> Learn about the Gyblos design for peer-to-peer AI inference, USDC payments on Solana, model audits and application access.

Canonical: https://gyblos.tech/docs

What prices, service measurements, model seals and public identity actually mean.

## Model catalog & custom declarations

The [unified model catalog](https://gyblos.tech/models/catalog) combines Models.dev model-only metadata, Hugging Face repositories and local custom drafts in one table. Models.dev refreshes on demand with a six-hour cache. Hugging Face contributes up to 100 repositories per search or lab, ordered by downloads, with a ten-minute cache. Results paginate in groups of 20; this is not a complete mirror of Hugging Face. Lab filtering recognizes known publisher namespaces; unknown organizations remain labelled as publishers. Multiple selected labs form a union and persist in the URL. A single selected lab scopes the Hugging Face author search; multiple labs filter the shared window of up to 100 Hugging Face matches, alongside all matching Models.dev records. Third-party derivatives stay with their own publisher. Records group only through explicit references to a matching Hugging Face repository and compatible revision. Partial source failures are disclosed while other results remain available. Failed refreshes can retain metadata for up to 24 hours, labelled with the original successful check time.

Publisher namespaces are not verified laboratory identities. Hugging Face relationships, licenses and capabilities are source declarations. Repository commits and individual artifact hashes identify references, not the model actually served by a seller. Missing fields remain unknown. Neither source grants a Gyblos audit seal or establishes network availability.

Custom declarations are validated without publishing them, then stored only in this browser and pinned at the start of the same catalog. They remain in a separate Custom / local lab group and cannot inherit an official lab identity by using its namespace. You can export a draft or remove it. They are not synchronized across devices and do not register a node or create an offer. Signed seller publication belongs in the applications; no live discovery source is connected here. External metadata does not automatically become a comparable price reference.

Explicit source-scoped model identifiers connect selected catalog records to illustrative provider quotes. Similar names, derivatives, custom declarations and separately requested repository revisions do not inherit those quotes. Input and output minima are independent and may come from different providers; open the catalog record for each paired quote and its price comparison. Missing links mean no associated example, not proof of absent network supply. Savings use the existing dated tariff references and expiry checks, never a price inferred from catalog metadata. The provider-offers filter and highest input/output savings sorts apply to the complete bounded search snapshot before pagination. Models without current comparable references sort last; missing links are not treated as zero savings. These controls persist in the catalog URL and return context.

## Prices & offers

Input is text sent to a model; output is text it generates. Buyer prices are displayed with $ in USDC per million tokens. The initial P2P protocol fee is 4%, included in the buyer price. There is one market: model audits do not introduce a different fee tier.

Catalog minimum input and output prices can come from different sellers. The model comparison pairs both prices from the same offer. Real prices must be signed, unexpired and bound to the accepted session; a later price change only applies to new sessions.

The current network offers and provider activity are illustrative, not live supply or official vendor quotes. Context limits, precision and retention must come from each seller’s signed declaration; unknown fields are not inferred from a model name.

External reference tariffs are checked automatically from the named source, with a cache of up to one hour while this site is viewed. Input and output differences are calculated separately, assuming 1 USDC equals $1 USD. A failed update retains the last valid tariff and its original check time; comparisons expire after 48 hours. Hosting references are not creator list prices or proof of equivalent configuration. Plan conditions and other manual notes retain their separate capture date.

## Settled payments

Settled inference consumption means confirmed increments attributable to Gyblos inference. Deposits, top-ups, returned funds, stake, rewards and relay payments are excluded. Settlement and distribution are different stages of the same money, not two sales.

Vouchers are cumulative: $0.02 → $0.05 → $0.09 authorizes $0.09 in total, not $0.16. A channel can contain multiple requests. Model-level sales need the accepted offer, usage records and linked settlement; a blockchain transfer alone does not prove tokens or delivery.

## Service reputation

Reputation is an automatic service score, separate from model audits, identity badges and optional stake. Initial weights: honest charging 25%, complete delivery 20%, diverse payment history 20%, age and recency 15%, speed 10%, availability 10%. These require calibration and a published versioned formula.

No verified measurements means not rated, not zero. The current provider activity examples never generate a real reputation score. Latency needs model, region, period and sample size; p50/p95 are computed from observations, not averages of provider percentiles.

## Model audits

A verified-model seal belongs to a specific provider offer, model version, precision and configuration, with signed auditor evidence. Eligibility alone does not grant a seal. Public reference weights are a prerequisite, while logprobs, method support and successful checks must also be established.

API models without public reference weights are not eligible for this public-weight seal. That is a technical scope limit, not a judgement of service quality. A statistical result is not an identity check, a guarantee of every response or an automatic refund promise.

## Public identity & optional stake

8004 identifies a node through addresses and keys, not a verified person. A provider can supply an image and website. Domain control, a .sol name, address history and stake are independent public signals. There is no KYC/KYB requirement.

Stake does not raise the service score. No stake or rewards program is connected to this preview. Compensation rules, coverage, stake denomination and protocol parameters require implemented evidence; no guaranteed refund or token market is implied.

## Sources, coverage & privacy

| Source | Status | Coverage |
| --- | --- | --- |
| External model metadata | On-demand | Models.dev and Hugging Face catalog records; not signed seller offers or model audits. |
| Network offers | Illustrative | Local model and provider listings; signatures and expiry are not checked. |
| Payments | Not connected | No attributed settlement indexer. Deposits and channel counts are not inference revenue. |
| Service measurements | Not connected | No signed delivery, token, latency or reachability observations. |
| Audit evidence | Not connected | No signed audit records or verified-model seals. |

No data is different from zero. Live metrics must expose their time window, sample size, source and last observation. Registered, listed, observed and reachable nodes are different counts; a node is not necessarily a unique person or GPU.

Buyer activity should be aggregated with suitable privacy thresholds. Do not publish prompts, responses, precise home-node locations or a per-request wallet/model/payment feed. A missing indexer observation does not prove a node is down.

## Website privacy, cookies & measurement

Some explorer data requests use a signed, short-lived session cookie without requiring an account. Browser storage remembers your theme and privacy choice. Cloudflare delivers and protects this public website and processes request and security metadata. Cloudflare Web Analytics measures browser visits and performance independently of the optional Google Analytics choice below.

Google Analytics 4 is optional. If you allow analytics, Google receives public-page visits, recognized referral sources, approved campaign labels and clicks to Gyblos on GitHub and X, together with standard browser and approximate location measurements. Analytics cookies distinguish browsers over time; these measurements are not anonymous. Advertising personalization and Google Signals are disabled.

Our Google Analytics integration removes URL fragments, search and filter text, individual model and provider identifiers and unapproved campaign values. External referrers are limited to recognized public domains, without their paths or queries. It does not send form values, model drafts, prompts, responses or wallet addresses. Missing references, declined analytics and browser blockers limit attribution; a click does not establish a signup, purchase or AI citation.

Google Analytics starts off. The initial notice lets you allow or decline it and disappears once your choice is saved. You can change or withdraw that choice using the controls in the Website privacy section on the website. Withdrawing stops Google measurement and removes this stream's analytics cookies. Your choice is stored in this browser; clearing site storage removes that preference. External destinations such as GitHub and X apply their own privacy policies.

[How Google uses information from sites that use its services](https://policies.google.com/technologies/partner-sites)

## Explore here. Operate in the apps.

The website is public and read-only. Buying, selling, wallet operations, node configuration and reward claims belong in the applications. No account or wallet connection is needed to explore.

[Explore the network](https://gyblos.tech/network)

## Related

- [Gyblos — Peer-to-peer AI inference marketplace](https://gyblos.tech/)
- [Gyblos paper — P2P inference protocol and status](https://gyblos.tech/paper)
- [Gyblos — IA para todos, de todos](https://gyblos.tech/es)
- [Affordable AI: compare cost per task](https://gyblos.tech/affordable-ai)
- [IA económica: compara el coste por tarea](https://gyblos.tech/es/ia-economica)
- [Sell AI inference: requirements and costs](https://gyblos.tech/sell-inference)
- [Vender inferencia de IA: requisitos y costes](https://gyblos.tech/es/vender-inferencia)
- [Free AI: options, limits and who pays](https://gyblos.tech/free-ai)
- [IA gratis: opciones, límites y quién paga](https://gyblos.tech/es/ia-gratis)
- [Inference for AI agents: models, budgets and retries](https://gyblos.tech/inference-for-ai-agents)
- [Inferencia para agentes de IA: modelos, presupuesto y reintentos](https://gyblos.tech/es/inferencia-para-agentes)
- [P2P AI inference: how it works and its limits](https://gyblos.tech/p2p-inference)
- [Inferencia P2P: cómo funciona y qué límites tiene](https://gyblos.tech/es/inferencia-p2p)
- [Data & methodology](https://gyblos.tech/docs/data)
- [Follow Gyblos development](https://github.com/gyblos)