Evidence-backed Context
Context becomes operationally trustworthy when every important fact carries provenance, scope, freshness, and confidence.
Context is information selected for a task. Evidence-backed context preserves why that information should be trusted.
The missing dimensions
A useful fact needs more than a value:
{
"fact": "api depends on billing-contract",
"scope": "project:api",
"evidence": ["package manifest", "import observation"],
"observedAt": "revision identifier",
"confidence": "verified",
"freshness": "current"
}The exact schema can differ. The dimensions should not disappear.
Why plain context fails
A retrieved document can be relevant and stale. A source file can be current and incomplete. A model-generated summary can be useful and inferred. Without labels, consumers flatten these different trust levels into the same prompt.
Evidence-backed context allows a consumer to distinguish:
- verified facts from inference;
- current observations from stale reports;
- facts in scope from adjacent information;
- known gaps from accidental omissions;
- primary evidence from explanatory summaries.
Context is a projection
The canonical system model should not be copied wholesale into every agent prompt. A context producer selects the smallest relevant view and retains evidence references and unknowns.
That makes context reviewable, reproducible, and safer to cache. It also lets another consumer reconstruct why an answer was available at a specific point in time.