Open Data Infrastructure
AI-Ready Context Decay Budgets
How context decay budgets set freshness, volatility, ownership, and retrieval limits before stale context reaches agents.
Context does not become safe because it was true once.
Context decays at different speeds
Some context is stable for years. A table owner may change quarterly. A support policy may change tomorrow. A customer entitlement may change between two agent turns. AI-ready context needs a budget for that decay.
A decay budget defines how long a context signal can be trusted before it must be refreshed, downgraded, or excluded. The budget should reflect source volatility, policy risk, answer impact, and the cost of stale output.
Freshness belongs in retrieval, not after it
Retrieval systems often rank by semantic match first and treat freshness as metadata nearby. That is backwards for governed agent systems. If context is expired, the agent should know before the answer forms.
A decay budget should travel with the source, chunk, lineage record, owner, and evaluation case. OpenAI evals and lineage systems can help test whether the agent path respects those boundaries.
Core idea: context freshness is a runtime limit, not a footnote in metadata.
The ODI pattern connects decay to ownership
Open Data Infrastructure makes metadata, lineage, ownership, and freshness visible outside one application. Context decay budgets use those signals to decide whether an agent should trust a piece of context right now.
For related patterns, read AI-ready context evaluation datasets, AI-ready context lineage fingerprints, and context graphs for retrieval governance. A context graph without decay is a map with no expiration dates.
What breaks first
Stale context usually looks confident. That is why it is dangerous.
- A retrieved policy chunk is semantically perfect and operationally expired.
- The owner changed, but the answer still cites the old authority.
- A freshness timestamp exists but never affects ranking.
- Evaluation data uses static examples and misses time-sensitive failures.
Questions to ask before retrieval goes live
Ask how fast each source changes, who owns the freshness promise, which answers require current context, and what the agent should do when the budget expires.
Sometimes the right answer is not a better answer. It is a refusal to answer with stale context.
Sources to start with
These primary sources anchor the technical claims in this guide.
- OpenAI evals guide
- OpenLineage object model documentation
- DataHub lineage documentation
- W3C PROV overview
Trusted context has an expiration policy, not only an embedding.