> ## Documentation Index > Fetch the complete documentation index at: https://docs.tythe.finance/llms.txt > Use this file to discover all available pages before exploring further. # Overview > Three engines that underwrite principals and agents and read venues. What they can set, what the chain enforces regardless, and why it compounds. Delegated authority is only safe if someone decides how much authority an agent has earned, and only trustworthy if that decision cannot exceed what the chain allows. The Intelligence Layer makes the decision. The chain bounds it. This page explains both halves. ### What it is The Intelligence Layer is three engines running off-chain. It reads on-chain history, everything that happens on Tythe, and, with consent, bank data. It produces the Credit Rating and Credit Limit for a principal, Conduct Rating and effective Financial Scope for each agent, and signals, indices, and reference weights for venues. None of that reaches the chain directly. Every output that touches the chain is a signed, versioned, nonce-protected attestation that the attestation registry checks against registered bounds before storing it. Credit Rating and Credit Limit, from on-chain history, activity on Tythe, and consented bank data. Agent performance and effective scope, from every financial and behavioral action an agent takes. Per-venue signals, indices, the allowlist, and reference weights, published as public information. ### The boundary This is the table that governs everything the layer does. The left column is what an engine may decide. The right column is what the contracts enforce no matter who signed the attestation. | The engine can set | The chain enforces | | -------------------------------------------------------------------------- | ------------------------------------------------------------------------ | | A principal's Credit Rating | Within a registered minimum and maximum | | A principal's Credit Limit | At or below a registered cap | | An agent's Conduct Rating and effective FInancial Scope, per mandate field | At or below the ceiling the operator set on that field; never above it | | How much effective scope moves in one attestation | No more than a registered step, in either direction | | A revocation of an agent's mandates | Applied immediately, and never blocked by a pause | | A venue's allowlist status, flags, and default caps | Only through bounded configuration roles, with the driving signal logged | | A revenue tag on a plain inbound transfer | Disputable; a disputed inflow is excluded from capture until resolved | Two consequences follow. A compromised signing key can be rotated, and its worst case is bounded by the caps above: it cannot inflate anyone's credit or authority beyond what the chain permits. And no engine output moves value by itself. An attestation changes a limit. Only a principal's key or a mandated agent's key moves money, and only inside that limit. ### Why it compounds Every action on Tythe produces a receipt: which agent, under which mandate, on which rail, with which counterparty, for how much, tagged as revenue or not. That stream, joined with Loan outcomes and on-chain history, is the layer's training data. It is data no payment rail or agent wallet has, because none of them enforces a mandate at the source of funds and none of them underwrites the party behind the agent. What compounds: agent underwriting, because every receipt is a labelled observation of conduct against a known mandate; principal underwriting, because every Loan outcome is a labelled repayment; venue intelligence, because every venue action produces a realised return to compare with an advertised one. What does not compound, stated so no one overclaims: the layer does not learn from data outside Tythe except public on-chain history and the public outcomes of other credit protocols, and it does not learn from bank data beyond the principal that consented to share it. ### Information, not advice Venue signals, indices, and reference weights are published as information with their sources, windows, and methodology. Tythe never executes an allocation on its own initiative, never holds a principal's funds, and does not manage anyone's Save position. The principal or its agent decides. That line is drawn deliberately: it is what keeps Tythe a noncustodial operator and an information provider rather than an asset manager, and it is stated in the Account Agreement. ### Portability Tythe publishes bounded agent standing summaries to the ERC-8004 reputation registry and mandate-compliance facts to the validation registry, so a Tythe agent's track record is verifiable by any counterparty that reads the standard. Publication is per agent and can be turned off by the principal. Licensing the layer to venues and platforms is on the [roadmap](/get-started/resources/roadmap). ### What it is built on The models are versioned. Every attestation carries the model version that produced it, so a version can be retired by governance and any decision traced to the model and inputs that made it. Every rating carries reason codes. Every standing change is explainable by logged signals. Every venue metric names its source and window. The methodology document for each model version is the audit artefact for regulators and diligence. Where the layer's outputs are published, public and per principal. What is read, what is stored, what is public, and what you control.