Here's the problem. An agent researches a market overnight and returns a confident, well-written brief. It reads great. Is it right? Without provenance, you have exactly two options: trust it blind, or redo the work yourself. The first is dangerous; the second erases the point of delegating.
Provenance turns output into evidence
Provenance means every claim in a delegated result can be traced to its source: the document, the feed item, the meeting record, the memory entry. In practice it changes the review conversation from "does this feel right?" to "these three claims carry the argument — let me check their sources."
That's the difference between plausible and reviewable. A plausible brief is a liability wearing good prose. A reviewable brief is leverage.
The source of a recommendation should never disappear
IEVA's prototype enforces this the simple way: knowledge is harvested only from completed, shipped work, memory lives in readable files, and Git history plus source links make every piece of context traceable. If a recommendation cites something, you can open the something.
The product direction keeps that rule as delegation scales to specialist agents:
- Delegated tasks run with bounded access, so the possible sources are known in advance.
- Results return with source-level citations attached, not summarized away.
- Provenance survives into memory — when a delegated finding becomes remembered context, it keeps its lineage.
Provenance is also a governance primitive
Audit history, approval gates, and memory correction all depend on provenance. You can't meaningfully approve an action whose basis you can't inspect, and you can't correct a memory whose origin you can't find. Which is why provenance in IEVA isn't a feature of the research agent — it's a property of the operating layer, applied to everything that moves through it.
Delegation should end in a clean handoff. A clean handoff starts with "here's what I found — and here's where every piece came from."