Plane 01

Provenance.
Know what can act.

You cannot govern what you cannot see or trust. Provenance is the backbone. Name everything that can reason or act, prove what each is allowed to do, and establish that the data feeding it is sound, sensitivity aware, and rightfully used.

01Provenance · The backbone

Can you name everything that acts, and prove what each is allowed to do?

Everything that can reason or act, in one register.

A model inventory records what can think. An agent register records what can act, as whom, and how far. The distance between the two is the distance between the world your controls were built for and the one you are entering.

The model inventory, table stakes

Models, owners, intended use, validation status, risk tier. The artifact your model risk function has kept for years, kept current automatically instead of by spreadsheet.

The agent register, the frontier

Intended use plus authority rung, standing access, escalation rights, and the named principal accountable for each agent. The layer few institutions have built, and the one autonomy makes urgent.

Authority as a first class field

Every entry declares its rung on the Authority Ladder. The register is not a list. It is where authority is declared and where the rest of the architecture reads it back.

Why it is the backbone

Every other control keys off the register. You cannot monitor, permission, or examine what is not in it.

Replaces

Agent and model registryAI GRC and policy registryModel risk system of record

Estate register · live

Munderwriting_score_v4validated
Awealth_rebalance_agentnamed owner
Akyc_review_agentnamed owner
Mclaims_triage_v2monitored

02Provenance · The deepest treatment

What entered this model, under what right, and what defect does it inherit?

Lineage for AI is a superset of lineage for reporting.

Reporting lineage answers where a figure came from. AI lineage also answers whether the model had the right to see its inputs, and whether an output carries a restriction forward. That is the superset an examiner will probe.

Provenance carrying data

Every record carries its lineage, quality score, and usage rights from the moment of ingestion. Provenance travels with the data. It is not reconstructed later under deadline.

The Quarantine Principle

Unclear lineage, expired consent, or an unverified license means the record is withheld before AI ever sees it. The default for doubt is exclusion.

Forward lineage

Trace any output back to every contributing input. When a source is found defective, name every decision downstream of it, for correction or recall.

Stakes for a financial institution

The plain question an examiner asks: prove this recommendation was based on data you were entitled to use. Reporting lineage cannot answer that. AI lineage must.

Replaces

End to end lineage engineSemantic layer for AICatalog and quality

Forward and backward lineage · a defective source

source flagged defective · two downstream decisions named

03Provenance · The first control

What is acting in your enterprise right now that is not in your register?

A register you fill by asking is a register of what people remembered to mention.

Discovery makes the register a record of what is actually there. Authority accumulates continuously: a SaaS update here, an escalated automation there, an inherited credential elsewhere. Point in time discovery ages the day it ends.

Find

Continuously surface models, agents, copilots, low code automations, and the non human identities they ride, across the estate.

Classify

By authority held and data reach. What can it do, and what can it touch. Shadow autonomy gets scored, not just listed.

Remediate

Move each find from found to registered, permissioned, and monitored. Or revoke it. Every find lands in the same register the rest of the platform enforces against.

The silent growth problem

A copilot that started writing back. A workflow that started deciding. A service account an agent quietly rides. Discovery is a standing capability, not an audit that is stale the day it ends.

Replaces

Shadow AI discoveryAI security postureNon human identity discovery

Estate radar · surfacing unregistered actors

found 37 · registered 29 · in review 8

See it on your own estate.

A 30 minute working session against a control, an agent, or a rule you already answer to.

Book a demo