Persona

Persona is a synthetic biomedical researcher built for the Claude Science hackathon at Gladstone: a system meant to stay “on” between sessions, hold a point of view, and act - not just answer questions about papers.

self vs. swarm

Most literature tools are either a search box or a one-shot summarizer - no memory, no accountability for what they believed yesterday. Persona splits identity from scale:

• The self is a small, durable belief graph: interests, beliefs, memory, and taste, persisted between runs. Every belief is provenance-typed - READ, INFERRED, HUMAN_CONFIRMED, or TESTED - so the system never confuses “the model thinks X” with “X is true.”
• The swarm is a large population of ephemeral, read-only agents spun up to read literature at scale (Europe PMC, Open Targets, SemMedDB, ClinicalTrials.gov, GEO) and report findings back. They carry no state of their own and can’t corrupt the self directly.

the acting loop

The point isn’t just tracking beliefs - it’s acting on them. When the swarm surfaces a contradiction against something the self believes, Persona runs a closed loop:

flagged contradiction
  -> falsifiable hypothesis (a claim with a metric, not a vague hunch)
  -> located public dataset (GEO / Open Targets / ClinicalTrials.gov)
  -> first-pass reanalysis
  -> belief-state updated, provenance-typed
  -> escalate to a human only when it needs wet-lab results or judgment

Humans are the resolver of last resort by design, not a limitation: anchoring their answers into the belief graph is what makes the whole system trustworthy under adversarial or noisy evidence.

membrane and poisoning resistance

A “membrane” gates what the swarm’s findings are allowed to do to the self’s beliefs - admitting support, rejecting noise, and routing genuine contradictions into the acting loop instead of silently overwriting confirmed beliefs. This was validated experimentally, not assumed: under correlated poisoning of the input stream, anchoring human-confirmed beliefs retained 100% of them, against 71% for a naive update rule that let new evidence simply overwrite old.

why the self stays small

Scale lives in the swarm; identity lives in the self. Keeping the self small and provenance-typed is what makes it auditable - a user can look at Persona’s living notebook and see what it currently believes, how confident it is, and why, the same way you’d read a colleague’s lab notebook rather than trust a black-box verdict.

why I built it

Existing tools (scite, Open Targets, PaperQA2, MedKGent, KARMA, AutoBioKG) are strong at retrieval and synthesis but stop short of the acting loop and a persistent identity. Persona exists to test whether a system that behaves more like a working scientist - holding beliefs, noticing contradictions, and following up - beats one that only answers the question it was just asked.