Dana Edwards
Back to home

Case study

Research Kernel MCP: Model Stays Creative, Kernel Stays Deterministic

A durable, evidence-first research substrate for MCP clients. The same thesis as MPRD, applied to research itself: an untrusted model generates claims, hypotheses, and reformulations, while a deterministic kernel decides what becomes supported. Hidden chain-of-thought stays outside the artifact store.

Summary

Research Kernel MCP is an MCP server that gives AI assistants a durable, evidence-first research memory. Instead of letting research vanish between sessions or hiding it inside a model's hidden reasoning traces, every claim, evidence record, refutation attempt, and reformulation is stored as a typed, queryable atom in a persistent graph. The model proposes; the kernel decides what is promoted.

Research Kernel control boundary: a model proposes atoms, a deterministic kernel gates promotion, and only supported claims survive. Untrusted model claims · evidence · reformulations Deterministic kernel promotion gate evidence + refutation + provenance required Supported claim fail-closed · replayable proposes promoted Rejected: fail closed no evidence · no refutation · no provenance
The control boundary: the model proposes atoms, the kernel gates promotion, and only claims that pass evidence, refutation, provenance, and contradiction checks become supported. Missing evidence, missing refutation attempts, or missing provenance fail closed.
Research Kernel MCP applies the MPRD thesis to research itself: capability separation between creative generation and deterministic gating, typed atoms instead of free-form text, and fail-closed promotion that refuses to mark a claim supported without evidence, refutation, and provenance.

What a reviewer should inspect first

  1. README.md - design spec, tool surface, and boundaries.
  2. internal/research_kernel_mcp/kernel.py - the deterministic kernel: atom types, statuses, edge types, promotion gate, and SQLite persistence.
  3. internal/research_kernel_mcp/server.py - the MCP tool, resource, and prompt surface.
  4. tests/internal/test_research_kernel_mcp.py - the test suite covering start, atom add, link, retrieve, morph, refute, evidence, score, frontier, promote, and report.
  5. pytest.ini - focused test configuration.

Suggested 10-minute review script

Read the README, then open kernel.py and inspect ATOM_TYPES, STATUSES, EDGE_TYPES, and the promotion gate logic. Finally, run the self-test: .venv/bin/python internal/research_kernel_mcp/server.py --self-test.

Claim / evidence map

ClaimEvidence to showStatus
Claim promotion is fail-closed: a claim cannot become SUPPORTED without support evidence, a refutation attempt, dependencies, provenance, and a rationale. The rk_promote tool and the promotion gate in kernel.py; the test that confirms promotion fails when evidence is missing. Supported / bounded
Every state-changing operation is appended to an event log for replay and external ledger ingestion. The events.jsonl append path in kernel.py; the test that confirms events are recorded. Supported / bounded
Artifacts are content-addressed by SHA-256 and restricted to allowlisted paths unless explicitly overridden. The artifact storage logic in kernel.py; the RK_ARTIFACT_ALLOWLIST and RK_ALLOW_OUTSIDE_ARTIFACTS behavior. Supported / bounded
Actual counterexamples create a COUNTEREXAMPLE atom, add a REFUTES edge, and mark the target REFUTED. The rk_refute tool with counterexample_is_actual=true; the corresponding test. Supported / bounded
The kernel is a complete research automation system ready for production deployment. Would require deployment hardening, load testing, multi-user isolation, and external review. Not claimed

Limitations to state clearly

  • SUPPORTED means the local gate passed. External publication or production posture still needs the relevant replay, review, and release gates.
  • Reformulations are candidate search objects until separately tested. They are not claims.
  • The kernel stores research artifacts, evidence, and graph state. Hidden chain-of-thought stays outside the artifact store by design.
  • SQLite persistence is durable for single-user and small-team use; multi-tenant deployment would need a different storage backend.

Open Research Kernel MCP on GitHub