A Spec for AI Science: How Research Objects Enable Rigorous Agentic Drug Discovery

Why “Spec”

In coding, vibe coding—a term popularized by Andrej Karpathy—describes prompt-first, AI-led development where the human doesn’t examine the code and coding becomes more about flow than structure. It’s fun, fast—but often brittle and opaque.
(See also Business Insider for a summary of the trend.)

Spec-driven development (e.g., Kiro’s spec mode) flips this around: you define requirements, architecture, and tasks before code is generated. It brings clarity and maintainability back into AI-assisted coding.
(See Kiro docs or SpecKit for open-source approaches.)

In science, unanchored agentic AI risks becoming its own form of vibe coding—speed without accountability.
Spec Science is the alternative: structured, reproducible, and grounded in Research Objects (ROs).

The Four Pillars of Spec Science

Provenance

Every result needs a story.
Capture who, what, when, and how. No orphaned outputs.

Versioning

If it changed, it’s new science.
Data, metadata, models, and workflows must all be frozen in identifiable versions.

Interoperability

Workflows must travel, not tangle.
Package outputs in standards like RO-Crate so they remain usable across labs, vendors, and regulators.

Reproducibility

No rerun, no result.
If an agentic workflow can’t be repeated with the same Research Object, it doesn’t count as science.

Research Objects: The Specs in Practice

Research Objects (RO-Crates) embody these principles:

Bundle data, metadata, provenance, models, workflows, tools, and authorship into one machine-readable package.
Ensure version-locking and alignment with the FAIR principles (Findable, Accessible, Interoperable, Reusable).
Allow agentic AI to explore freely while leaving behind an auditable trail that others can rerun and build on.

In short, ROs are the specs for Spec Science.

Why This Matters for Drug Discovery

Scientists get AI collaborators whose outputs are transparent, not black-box.
Regulators see provenance and version history, easing compliance.
Patients benefit from speed and trust: discoveries that are auditable, reproducible, and safe.

Closing Line

Spec Science = Agentic AI grounded in specs.
Research Objects turn autonomy into accountability — making drug discovery not just faster, but also reproducible, auditable, and real.