aurora-agent AGENTS.md

AURORA Agent

AURORA Agent is a context harness for biological research agents. Its one idea is that context assembly is a compiler pass, not a retrieval heuristic: a typed decision query compiles into the smallest decision-sufficient evidence region, delivered as a Decision Section and accompanied by a Context Certificate stating exactly what was omitted and whether the omission could have changed the decision.

Everything else in the workspace is downstream of that. Storage exists to make compilation output-sensitive. Evaluation exists to test whether the compiled context was any good. Mutation exists to generate more decisions to test against. The MCP server exists so the compiled context reaches something that can act on it.

The engine crate is bioprism-fiber; the workspace is prefixed bioprism- throughout, from the BioPRISM blueprint it implements.

What we can never compromise on

1. Honest labelling is the product

Every other system can tell you what it included. This one tells you what it left out, and whether that could have mattered. If we lose that, we are a slower retrieval library.

Concretely, and these are non-negotiable:

• Zero influence is not unknown influence. "Provably cannot matter" and "nobody checked" are different states and must never share a representation. A single unknown-influence group voids a sufficiency claim.
• Unmeasured is not zero. A capability with no evidence is Unmeasured, categorically distinct from measured-and-poor. There is no score_or_zero.
• A right answer from an incomplete basis is not a pass. Protected closure is mandatory before any relevance step, so a strategy cannot be credited for guessing correctly from evidence it never had.
• Instance count is not benchmark count. Report independent equivalence classes. A million paraphrases are a robustness check, not a million benchmarks.

2. Invariants belong in the type system, not in comments

Where a rule can be made unrepresentable, make it unrepresentable. Existing examples to match: