DDN

Safe Harbor: real genomic evidence investigations with durable memory and evaluated harness adaptation From docs/safe-harbor/SPECIFICATION.md: What it is: a system of AI agents that uses real genome data to look into genomic safe harbors. These are candidate spots in the human genome where DNA could be inserted without disturbing nearby genes. It also tests changes to its own workflow (the "harness") to see whether they make those investigations better. Scope: Genome and cells: the human GRCh38 reference genome, in H1 embryonic stem cells. Candidates: three sites from a published study (Autio et al., eLife 2022): Pansio-1, Olônne-18 and Keppel-19. It is not a new whole-genome search. Question each investigation asks: "Do these measured expression changes justify excluding this candidate region?" How it works: Data: agents use approved tools on the real source data: the paper's supplementary tables, GENCODE v36 annotations and the actual reference DNA. Execution: one coordinator runs at most two workers, with at most 24 tasks and 3 replans per run, and hard token, tool and cost budgets. Record keeping: every result is committed to MongoDB in a transaction (now moving to Atlas), so runs survive crashes and can be replayed and exported. Harness changes: the workflow can change itself only through saved, executable structural patches. These are compared fairly against two fixed baselines before a change is promoted. UI: coordinate-accurate genome zoom, an investigation tree of the agents, the evidence behind each result, and a timeline replay. Ground rules: It never labels a site globally "safe". Its three statuses (screen result, experimental evidence, freshness) are reported separately. Missing evidence counts as "incomplete", not as a negative result. It shows clearly whether a run was a mock, deterministic operational, a real model run, or a recorded replay. It makes no claims of new discoveries, biological safety or improvements it hasn't measured.