AI knowledge that must pass a seven-stage trial

What it does

ClearAI is a DeepSeek Harness (DSH) plugin that turns research sessions into structured ontology-building. It runs a seven-stage loop—frame, hypothesize, plan, observe, verify, evaluate, record—so that every knowledge entry carries evidence, boundary conditions, and a support level. The result is a domain-specific graph where contradictions surface automatically and even refuted hypotheses are kept as first-class results.

The interesting bit

The system is designed so the model cannot declare its own work complete: state is derived from the session record, and above a certain support threshold the agent that produced a result cannot be the one to judge it. This splits the doer from the verifier, whereas the default task loop lets whoever did it declare completion.

Key highlights

• Ships as a prebuilt npm plugin that adds a preset to DSH without modifying the engine.
• Contradictory assertions surface automatically in the graph (tinted amber) and retraction is a human decision.
• Failed hypotheses are preserved, not deleted or retried into oblivion.
• Graph projection is deterministic: the same ledger always yields the same ontology picture.
• Runtime dependency is only zod; the graph stack is bundled at build time.

Caveats

• The README explicitly lists known gaps and features not yet verified in a real browser.
• Most UI screenshots in the docs are in Chinese; English is supported but the visual documentation is not.
• Installing from source requires granting allowBuilds to pnpm, which the authors warn against unless you audit the code first.

Verdict

Scientists and researchers who need an audit trail for AI-assisted discovery should try this; if you just want fast answers without epistemic overhead, it will feel like paperwork.