Gene agents that hold their own policy

Four properties, each checked on every run

1. No controller decides who acts. Evidence items arrive in the order CIViC's curators reviewed them, and every item goes to every agent. Each agent decides for itself whether an item concerns it.
2. States you can check. An agent is Idle, Reasoning, Publishing or Cooldown, and moves only along its state machine. Every transition is a record with the state before and after, so the color on screen is the state the agent was in, not a guess from traffic.
3. Policy is a file the agent holds. Each agent loads its own policy file once and cites its id, version and sha256 on every decision. At the default floor, 140 of 269 EGFR items are refused for therapy claims; the finding stands, what is refused is using it to say a treatment will work.
4. One way out. The agents contain no network or backend code. A capability call goes to a gateway that holds its own policy file: it refuses a backend that is not on its allow-list, serves an exact repeat from cache, or runs the call.

policies/tp53.policy.json
{
  "policy": "gene-agent/tp53",
  "version": 1,
  "evidence": { "therapeutic_floor": "B", "allow_flagged": false },
  "capabilities": [ "assemble_contig" ],
  "cooldown_steps": 2,
  "on_cooldown": { "evidence": "defer", "request": "refuse" }
}

A real QAOA run, reported with its margin

TP53 may request one capability: de novo assembly of a short contig as QAOA, following the published QuASeR method (Sarkar, Al-Ars and Bertels, PLoS ONE, 2021). Three reads, nine qubits, on PennyLane's default.qubit simulator. It returns GCGCA, checked against the brute-force optimum.

The result is shown with its margin. The six valid orderings come out nearly even: the correct one has probability 0.08385 and the runner-up 0.08335, a margin of 0.0005. Told apart by sampling, that would take about six million shots on a perfect device. The demo does not run on quantum hardware: under a Heron-class device's noise model the correct ordering ranks second. Asked for ibm_hardware, the gateway refuses before anything is submitted.

The assembly code is from the public quantum-biology demos (doi:10.5281/zenodo.22741061).

The limits, stated

• No biology in the quantum step. The contig is a toy three-read instance, not sequence from any gene; the connection to the gene agents is how the call is governed.
• Simulator only, exact probabilities. No quantum advantage is claimed.
• The evidence refusals are this demo's policy, not CIViC's. CIViC's download carries accepted evidence only.
• No clinical claim. Nothing here is decision support.
• Reasoning is rule-based, so a run replays deterministically. A model call, if added, would go through the gateway like any other.

Evidence: CIViC (CC0), nightly clinical evidence summaries retrieved 2026-09-20. The Decision Trace view is the published @agent-scope-ca/graph-canvas component.