17.13 Toward an integrated open-domain benchmark

AGENTIC-CORPUS–0 uses a temporally cut, closed GLP-1 corpus with disjoint development and confirmation sources. It withholds a context-dependent mediator, the evidence-reconciliation workflow needed to represent it, and the query policy that efficiently locates the decisive sources. A locked evaluator contains later documents or structured evidence not available during construction.

The component factorial tests whether CLIC alone proposes unsupported latent variables, whether OPTIC alone improves extraction without producing a theory, and whether RELIC alone retrieves more documents without resolving their types. The full system must construct a source-bearing theory extension and a discriminating external experiment. Locked documents confer only corpus-relative credit; causal identification requires appropriate experimental, clinical, or simulator evidence. Otherwise the correctly calibrated output is a proposal with preserved alternatives.

Experiment: AGENTIC-CORPUS–0. Epistemic status: proposed integrated benchmark, not a completed AGENTIC result.
World: temporally cut scientific corpus with heterogeneous evidence, contextual disagreement, and locked confirmation sources.
Joint construction: source-bearing theory extension, typed evidence workflow, and adaptive acquisition or experimental plan.
Admission: provenance and reproducibility first; then locked experimental, clinical, or simulator evidence where identification is claimed.
Controls: component factorial, flat retrieval-and-generation agent, admission-record ablation, and source, mechanism, and query-policy oracles.
Boundary: coherent text synthesis and citation agreement do not identify a causal mechanism.

. Can a system convert heterogeneous scientific appearances into a theory whose missing variables and experimental consequences are explicit?

The following visual testbed changes the medium but preserves the structural problem. A generated image is a model of a prompt-conditioned visual theory; violations of cardinality, role, layout, or identity reveal gaps between that model and the declared sketch. This makes structural diagnosis directly visible before the book asks the harder question of inventing a reusable visual technique.