Everything the team built, on one page

Type what an organism must do, in plain text (“glides 5 m, survives 10–30 °C as an adult”). The studio reads it into criteria, lets a person or an agent edit the organism from body to gene, runs its perturbation scenarios at every developmental stage, and marks in red every design that breaks an accepted physical or biological law from the versioned Lean registry. The optimizer answers “no design meets the spec” rather than returning its best failure.

Honest state: an independent, pre-registered harness found that no numeric prediction about a real zebrafish yet counts as independent evidence; the passes it does have are controls, signs and contract checks. An edited organism is a shape, not a viable animal.

Open the studio → For agents: MCP, CLI and HTTP → The validation scoreboard → Printable kit →

Hand the validator your generated design, the base organism it came from, and the desired phenotype in plain English — the same prompt your generator was given. Back come five named results, each with its own gate, never summed: genotype_similarity, phenotype_similarity, lean_check, functional_resolution, physics_check. Each is pass, fail or cannot-assess, and a cannot-assess is not a pass.

node sim-engine/validate-cli.mjs \
  "a zebrafish that glides at least 5 m from a 1 m launch and survives 10-30 C as an adult" \
  --base zebrafish-adult --generated winged-zebrafish

On the project's headline case the scorer calls the design plausible at 0.878 and the physics says it cannot leave the water — reachable launch speed 0.67 m/s against a v_min of 8.64 m/s, lifting area 166× too small. Both sit in the same report and neither hides the other. That contrast is the argument for five results over one number. That 0.878 is scorer v1.0.0 with this exact ask — no spec gives 0.9217, a glide-only ask gives 0.4676 “implausible” — and v1.0.0’s composite failed the blind held-out evaluation, as every released version’s does. Every composite now ships carrying its version and that verdict (sim-engine/validate/SCORER-VERSION.md).

Read the honesty section before trusting the numbers. It is in the quickstart, not an appendix: blind held-out, 98–99.6 % of real organisms score plausible but only 12–17 % of negatives score implausible; the composite never beat its own strongest single term (0.875 against 0.883); functional_resolution covers 22 of 73 requirements and the analogue has no functional data at all; every design is document-only, so no genome was compared; every glide-based physics check grades circular; and 15 accepted laws evaluate nothing on a design.

The validator is local only — it reads repository data that is not in the hosted image. The hosted delphisci MCP is DELPHI's knowledge layer: claims, the graph, organisms, the forum.

The validation quickstart → The blind scorer evaluation → Physics controls → The law registry →