Concordia Colony
Auditable agent workflows for scientific evidence
A local-first AI system that traces generated claims through a durable workflow, bounded scientist agents, deterministic verification, and a content-addressed evidence graph.
Python
FastAPI
Qwen 3
Ollama
SQLite
DuckDB
PostgreSQL
RDKit
SHAP
Evo 2
Boltz-2
NVIDIA API
React
TypeScript
3Dmol.js
Docker
Hugging Face SpacesConcordia Colony turns genomic model explanations into traversable scientific records that connect every claim to its sequence context, model output, attribution, counterfactual result, and supporting source.
A seed scientist creates bounded workflow descendants that inherit versioned prompts, permitted tools, attribution settings, verification plans, and resource budgets.
Seven versioned, schema-validated tools run in child processes with path validation, deny-by-default network policy, and explicit resource budgets.
Every sequence, model output, attribution, mutation, paper, claim, and verification decision enters a content-addressed evidence graph with immutable manifests and explicit ancestry.
A deterministic evaluator traces claims to source artifacts and scores provenance completeness, counterfactual consistency, cross-method agreement, stability, and resource use.
A measured three-member Qwen 3 software run produced valid schemas, references, tool calls, reproducible lineage, and component-level fitness records. The public read-only workspace clearly separates fixture validation from scientific findings.