Prism · agent harness · production

Agents that reason about science without making things up.

Prism is the 5-layer agent harness that orchestrates Sci-JEPA, DarkScan V2, and the broader SciRouter tool surface. Every scientific decision the platform makes is logged to a provenance ledger — model version, inputs, intermediate state, output, who asked.

The 5 layers

From request to provenance, no hidden state.

Typed triage

Inbound request is classified by intent (atlas query, neoantigen rank, vaccine design, comparative lookup). Schema-validated.

LLM planner

Frontier LLM proposes a plan as a typed graph of tool calls. The plan, not the prose, is what executes.

Typed-dispatch executor

Each step runs a real SciRouter tool with input/output JSON schemas. Failures are typed, retryable.

Reporter

Structured results assembled into a response object. Plain-English summary attached for human readers.

Provenance ledger

Every step + version + input hash + output hash + actor recorded. Internal-only in v1; hash-anchored externally in v2.

Why typed dispatch matters

LLMs hallucinate. Schemas don't.

Most agent frameworks let the LLM call tools by guessing parameters from prose. Prism only runs tool calls that conform to a JSON schema; if the planner can't produce a valid one, the call is rejected before any compute runs.

What Prism does NOT do

· Execute prose. The plan must be a typed graph.
· Trust LLM output as ground truth. Tool results are the truth.
· Hide intermediate state. Every step is in the ledger.
· Skip provenance. Even cancelled runs leave a row.

What Prism guarantees

· Reproducibility: same input + version = same output.
· Auditability: every decision traces to a tool call + inputs.
· Composability: any tool is callable from any agent step.
· Testability: planner + executor + reporter mock independently.

What this means for partners

Citable, regulatable, defensible.

For vet schools

Every Sci-JEPA result you cite in a manuscript ships with a ledger ID. Reviewers can re-run the exact query against the exact model version. We support replication.

For vet pharma

Companion-Dx claims need defensible evidence trails. Prism produces them by default — regulatory submission packets are downstream of the ledger, not built from scratch.

For practicing vets

Every report a vet downloads from the portal carries model versions, dataset snapshots, and a hash of the inputs. Defensible against second-opinion review.