Research question
Can an agent recognize an invalid or insufficiently specified analysis before it runs it—and retain enough evidence for a scientist to assess the eventual result?
Dogma checks whether a proposed analysis has the data, method assumptions, and provenance needed to assess its claim. A completed workflow can still use an unsuitable cohort, readout, or method.
Claim graphs
Dogma treats a hypothesis as a graph. Nodes represent biological entities or states; edges represent mechanistic claims. Each edge connects a question to the proposed dataset, method, assumptions, measurements, provenance, and missing evidence.
The browser workspace and local tools support evaluation plans, workspace inspection, missing-prerequisite checks, and factual evidence records. The evidence ledger records measurements and their limits without automatically collapsing them into a biological verdict. A scientist interprets what the record supports.
The current code is consolidated in one local monorepo. The older quration name remains a compatibility namespace.
Method assumptions
Methods-graph connects bioinformatics methods to operations, containers, statistical assumptions, diagnostics, and data handoffs. Sources include EDAM, nf-core, BioContainers, bio.tools, STATO, and OBI.
A proposed step can return EVALUABLE, BLOCKED, or NOT_EVALUABLE. These are method-level decisions. When the graph lacks the necessary coverage, it should expose the gap.
| Method outcome | What it means |
|---|---|
| EVALUABLE | The declared requirements can be assessed in this context; biological validity still requires interpretation |
| BLOCKED | A known requirement or gate is not satisfied |
| NOT_EVALUABLE | Coverage or required information is insufficient to assess the proposal |
One public synthetic RNA-seq workspace exposes one replicate per group, an incomplete FASTQ pair, missing process containers, and an untrusted workspace. Its error findings and missing trust policy block the Nextflow stub-run plan.
The methods-graph guardrail makes assumptions and decision paths inspectable. A threshold passing does not itself prove that an analysis is valid.
Capabilities and limitations
| Surface | Present capability | Remaining boundary |
|---|---|---|
| Browser graph | Causal claims, per-edge plans, and evidence inspection in the local workspace | Graph-bound orchestration and result write-back are not yet joined and tested end to end |
| IDE extension and local sidecar | Diagnostics, plans, workspace inspection, and reviewed dry-runs or stub-runs | The planning surface is not an autonomous real-analysis engine |
| MCP tools | Deterministic claim/evidence and method-control operations over a selected local workspace | No claim of a verified end-to-end external workbench integration |
| methods-graph | Method contracts, provenance, assumptions, diagnostics, and coverage gaps | Not yet a hard boundary around every mutation or execution |
Current local documentation lists 905 methods, but only 49 have curated semantic input types and 39 have output types. Catalog size is not equivalent to complete validation of data handoffs. Those figures describe the local graph documentation, not a new scientific benchmark result.
Project history
- November 2025: Added data connectors, analysis scaffolding, and method planning to NGS metadata curation and GEO search.
- February 2026: Added agent bridges and interpretation evaluations.
- June 2026: Added editable causal graphs, per-edge evaluation, and factual evidence records.
- July–September 2026: Released local-control tools and consolidated the local graph workspace.
My contribution
I conceived and lead Dogma, defined the claim/evidence architecture, created methods-graph, and set the method-validity requirements and evaluation boundaries. Claude Code and Codex assisted with code, tests, UI, migration, and documentation under my direction and review.
Evidence and limits
Public source snapshots: Dogma prototype, its verification record, and methods-graph. At those historical revisions, the Dogma extension suite and 77 sidecar tests passed; methods-graph reported 552 passing tests and one skip. Those are software checks, not evidence of biological correctness.
The browser consolidation and newer graph-coverage figures describe local development reviewed in September. Demo inputs are synthetic unless stated otherwise. The current work does not establish an autonomous claim-to-experiment-to-result loop.
Next step
Connect one real computation to its originating graph claim. Preserve the input data, method version, checked assumptions, execution record, and resulting measurement, then test that the record distinguishes a measured result from a feasibility assessment. Extend coverage based on the concrete gaps that this exposes.