Can research papers preserve the experiments that failed?

Publishing compresses a branching, iterative research process into a linear narrative, and ARA (Agent-Native Research Artifact) names the two costs of that compression precisely. The Storytelling Tax is the systematic erasure of process knowledge — failed experiments, rejected hypotheses, the backtracking that explains why the final approach was chosen — to fit human-readable convention. The Engineering Tax is the gap between reviewer-sufficient prose and agent-sufficient specification: critical implementation details left unwritten because a human referee never needed them. Both taxes were tolerable when every consumer was human. They become critical when AI agents must reproduce and extend the work, because an agent cannot interpolate the tacit knowledge a human author assumed. ARA's response is structural: a four-layer package — scientific logic, executable code with full specs, an exploration graph that preserves the discarded failures, and evidence grounding every claim in raw outputs.

What makes this more than a better file format is the framing of failure-knowledge as a first-class published object rather than an editing casualty. The exploration graph treats the rejected branches — the part that explains the result — as the deliverable, not the offcut. This is the supply-side complement to the demand-side problem in Can AI verify research outputs as fast as it generates them?: if agents generate faster than anyone verifies, then publishing evidence-grounded, machine-checkable artifacts is exactly the verification substrate that pace requires. It also reframes Where does AI assistance become unreliable in research? — much of what makes the "novel experiment / judgment" stage hard for agents may be the missing tacit specification ARA aims to externalize, not an irreducible capability gap.

The skeptical reading: ARA presumes the value of a paper is its reproducible specification, but a paper's narrative also does persuasive and sense-making work — it argues why the result matters, which an exploration graph does not supply. There is also an incentive problem the protocol cannot solve by structure alone: authors are rewarded for clean stories, and publishing one's failed branches exposes the messiness that the Storytelling Tax was designed to hide. The format can make failure-knowledge expressible; it cannot by itself make researchers want to expose it.