What is Epistemic Provenance?

How epistemic provenance works

Epistemic provenance is a typed metadata object that rides alongside every retrieved piece of context. For a search result, it records the source file the content came from, where inside that source the chunk sits (position and total length), when it was ingested, and any typed edges that connect it to other entries, such as whether another chunk elaborates on it or contradicts it.

The distinguishing property is that an agent reads it at inference time and acts on it. Before issuing the next call, the agent can see that this match has 32 sibling chunks in the same source, that it carries an elaborates edge to one other entry and a contradicts edge to another, and that it was ingested yesterday rather than two years ago. Those signals reshape the next action. The agent can traverse to adjacent chunks instead of issuing a new search, weight a newer source over an older one, or surface both sides of a contradiction instead of flattening them.

This is a different target than traditional data lineage, which is an offline record designed for human auditors. Lineage answers “how did this value get here?” after the fact. Epistemic provenance answers “how should the agent act on this next?” right now.

Why it matters

Most retrieval APIs return an id, a similarity score, and a content blob. That is enough for a single-shot RAG answer, but it leaves the agent blind on every decision that comes after the first match. Without position data, the agent cannot tell the difference between “the whole source” and “a fragment.” Without ingestion time, it cannot tell fresh from stale. Without typed edges, it cannot distinguish content that backs up a claim from content that refutes it.

The consequence is a class of failures that look like hallucination but are actually attribution failures: the agent produces a claim that contradicts retrieved evidence, cites a source that does not say what it implies, or ignores a newer contradicting document because it arrived in the result list alongside older corroborating ones. Structured context with typed provenance closes that gap at the contract level.

There is also a cost dimension. When an agent can traverse from a good match to adjacent chunks or related entries, each step is cheaper than issuing another full semantic retrieval. Wire’s April 2026 agent-efficiency benchmark measured a 20% reduction in tokens consumed per question when the retrieval surface exposed navigable provenance, because the agent stopped re-searching for context it had already landed near.

Common misconceptions

• “Provenance is a security feature.” Security teams care about provenance for audit and poisoning defense, and those uses are real, but treating provenance as a security layer alone misses what it does for retrieval quality. It is primarily a grounding and planning primitive that happens to have security benefits.
• “Provenance is the same as citations.” A citation is one surface produced from provenance. Provenance is the underlying typed metadata; citations, traversal hints, and reranking signals all derive from it.
• “A trust score in the result is better than raw provenance.” A trust score collapses a judgment the agent should be making in context. It also becomes an attack surface. Neutral structural metadata preserves the agent’s ability to reason.
• “Provenance only matters for compliance-heavy domains.” Any multi-step agent benefits. Cross-document synthesis, long-running memory, and correction-aware retrieval all require it.

Epistemic provenance and Wire

Every result returned by wire_search and wire_navigate carries a typed provenance object: source file and id, chunk index and total chunks, ingestion timestamp, tags, section headers, and a _meta.wire.navigate block that surfaces sibling counts and typed relationship edges. Relationship types include elaborates, corroborates, contradicts, and supersedes, each with an edge direction so an agent can tell “X contradicts my chunk” from “my chunk contradicts X.” Wire does not attach a trust score. The agent decides what to trust; the container tells it where things came from and how they connect.