Retrieval Context Provenance Auditor

The problem

Which retrieved information influenced this decision, and how trustworthy was its supply chain?

When a retrieval-augmented agent makes a decision an incident responder disagrees with, the hard question is not what the model said — it is which retrieved material produced that answer, where that material came from, and whether the source had already started behaving badly. Most stacks cannot answer any of the three after the fact.

What it does

It records the provenance of every context assembly as digests and declared metadata only — never document text — and keeps those records in a hash-linked chain a later reader can verify. From those records it traces a decision backwards through context, transformations, chunks and sources, ranks which sources actually influenced it, scores source trust and freshness with an explicit basis, and flags retrieval schema drift.

How it works

A zero-dependency stdio MCP server written fresh from a public behaviour specification. Influence attribution always reports the share it cannot explain as an unexplained residual, and missing lineage edges are reported as missing rather than inferred. Source trust is advisory, bounded per observation, and the policy in force is published by the server itself.

Tool surface

MCP over stdio. Compatible with Claude Code, Claude Desktop and Cursor, or usable directly as a TypeScript library. No API key, account or Tenable product is required.

• index_sources

  Register retrievable sources with kind, first-seen time and optional declared schema.

• record_retrieval

  Record one decision's context assembly as digests plus declared metadata; never accepts document text.

• record_source_observation

  Record a corroboration, contradiction or retrieval-failure observation about a source.

• trace_decision

  Traverse the source → chunk → transformation → context → decision lineage graph.

• score_source

  Advisory source trust, freshness and retirement recommendation, with its basis.

• report_influence

  Ranked influence attribution with a mandatory unexplained residual.

• check_retrieval_drift

  Compare an observed source schema against its declared baseline.

• describe_scoring_policy

  Return the trust policy identifier and its fixed public maxObservationStep.

• export_audit_records

  Export the payload-free hash-linked audit records produced so far.

What it deliberately does not do

• ·It never performs retrieval and never ranks documents for an agent at query time.
• ·It never blocks a decision — it records and reports.
• ·Document text is never accepted or stored; provenance is digests and declared metadata only.