Understand the platform
Provenience is a planned foundation for applications that need to know what happened on a project, who was involved, and where each claim came from. Its core path is Capture → Extract → Contextualize → Govern → Expose.
Projects arrange context
A project has a beginning, an end, channels, and a roster. A project-specific channel establishes the project. On the shared Site channel, a conversation with a customer goes to the customers' one shared active project, whoever else is on it; several shared active projects or none wait for an administrator. With no customer, every registered party must share the project: one shared active project files there, several go to the Review Queue, and none is company-level. Filing never comes from message content.
Sources preserve the record
The Deposit stores received originals for their authorized retention period. For voice, that original is the transcript returned by the provider, not audio held by Provenience. Audited deletion under retention rules leaves a tombstone. SMS/MMS Extract can use qualifying images and short PDFs with their message; each attachment is also cataloged as a file. Other files can be cataloged without making a claim about their contents.
Artifacts make communication useful
The Catalog contains discrete records derived from sources. An artifact carries its category, project context, source citation, and governance information. Certainty describes whether the claim was firm or tentative when made. Standing describes its relationship to later claims: current, refined, or superseded.
Reprocessing produces a new extraction generation; it does not rewrite the original source or change a claim's standing. A confirmed close-out sweep is what changes standing.
Signals record expectations and outcomes
A signal is an artifact with a kind, value, dimension, and subject. “We expect to spend $30–60k” is a stated budget signal. A recorded final amount is an actual. Human-confirmed reconciliations link a chain to its outcome at close-out.
Reference data supplies another kind of evidence
A connected accounting or job-management system supplies reference rows rather than artifacts. Its identifiers map to projects during setup. Integrations are read-only in Phase 1; there is no general write-back feature or fixed supported-vendor list.
MCP exposes governed knowledge
Any AI assistant that can add a remote MCP server and follows the MCP authorization standard can connect to it. The administrator's steps are in Connecting an AI assistant; an operator turns it on with Turn MCP on. Its seven read tools list projects and typology, read a project and its roster, page through or read one governed artifact, and list or read an open Review Queue card. Successful record results return authorized, whole extracted artifacts and links into the Site rather than raw message text. The server checks returned rows again as the caller before it renders them, and every result ends with a Review Queue attention line.
Artifact lists are newest first, hold up to 500 records per page (200 by default), cut near 60,000 characters, and include a continuation cursor when another page may remain; a full final page can be followed by an empty page. Phase 1 has one administrative role within each customer business; more granular and external roles are deferred.
MCP tool names and application SDKs are not published contracts yet. The reference describes implemented interfaces only. Start with the workflow to see how the pieces fit together.