Amawta CLI – turning hypotheses into auditable scientific workflows

Hi everyone,
I wanted to share a small project I built during the hackathon: Amawta CLI.

Amawta treats a hypothesis not as something to “discuss”, but as something to execute. The system converts a claim into a bounded scientific workflow that produces reproducible artifacts and explicit epistemic outcomes.

In practice it:

• Normalizes a hypothesis into a minimal schema (domain, entities, relations, observables)
• Generates a falsification plan with a small variant matrix
• Performs grounded literature search to avoid repeating existing work
• Runs a two-phase execution model
 – Toy run: fast sanity checks and quick falsifiers
 – Field run: dataset resolution and evidence-based execution when the toy stage doesn’t falsify
• Emits deterministic gate reports: PASS / FAIL / UNRESOLVED

All stages produce versioned JSON artifacts (normalization, literature search, falsification plan, datasets, runner logs, results, gate reports). The workflow is artifact-driven and resumable, rather than dependent on chat history.

The goal is to explore a different model for “AI science”:
less persuasive narrative, more auditable execution.

Compared with the impressive projects in the gallery this is a small experiment, but I’d love to hear thoughts from others interested in reproducibility, AI-assisted research workflows, or executable epistemology.

Thanks to everyone sharing such interesting work here.