JURASSIC PARK genome research

Jurassic World Evolution 3 · Genome research

GENOME RESEARCH.

How the 8 dinosaur species were classified and sequenced for the park

THE HYBRID RULE: small species for cheap/mechanical work, Triceratops for default quality, T-Rex for deep reasoning - and every dino falls back to a smaller one on failure. Zero-credit local species first.

1. Classification by task kind

Every request gets a task kind; the kind routes to a species. This is the core sort.

Task kindSorted toSpeciesWhy
summarize · extract · template · recipe · classify · travel · draftLOCALCompsognathusMechanical, template-shaped - a tiny local genome is enough, $0
(everything else - default)FLASHTriceratopsDefault quality work, ~$0.14/M, reliable
analysis · reasoning · debate · financialPROTyrannosaurus RexDeep reasoning needs the heaviest genome, ~$0.43/M
deep-analysis · hard-reasoning · complex-reasoning · frontierFRONTIERIndominus RexFlagship engineered hybrid via Pro subscription (in-app Claude Code limit ~50% higher than documented)

Unknown kinds default to Triceratops (quality over cost). The routing lives in agents/_shared/llm.py → tier_of(task_kind).

2. Free-roaming species: the cloud fallbacks

When a billed species fails (rate-limit, overload, connection error), the call cascades down through the free-roaming species before landing on the breeding facility (local). Each was probe-verified before release into the park:

SpeciesProvider / modelVerified how
GallimimusBluesminds - llama-3.1-70bProbe call 200; 600 req/day cap
BrachiosaurusOpenRouter - nemotron-3-super-120b :freeProbe verified working (120B MoE, $0)
VelociraptorGroq - llama-3.3-70bFastest; fixed a Cloudflare 403/1010 UA-block to make it work
DilophosaurusNVIDIA NIM - llama-3.1-8b ★Key tested live; 8b always warm (~1s); 70b exists but cold-starts 4-10 min

Full fallback order: T-RexTriceratopsBrachiosaurusVelociraptorDilophosaurusGallimimusCompsognathus. A T-Rex call that fails still answers - smaller, never errors.

3. The cheap-first cascade

Not every call needs the species its kind suggests. The park runs a cascade: start with the smallest species (local/free), and escalate only on a structured failure signal:

  1. Route the task kind to its species (above).
  2. Try the smallest species first - Compsognathus or a free-roamer.
  3. Escalate ONLY if: empty output, a refusal ("I cannot..."), or a validator rejects the answer.
  4. Repeat up the chain until a good answer comes back.

Result: easy intents cost $0; hard ones still get the heavy genome. No waste.

4. Quality guardrails on free species

5. DNA bank & caching - why $0 is real

LayerMechanismEffect
L1 primary samplesDisk cache on identical requestsRepeated calls: 0 credits
L2 genetic matchingMiniLM embeddings, cosine ≥0.93Near-identical questions: 0 credits
L3 deep genome prefixDeepSeek auto prompt cache (stable species prefix + dynamic tail)~68% prefix hit on second call, 256-token floor
Local sequencingsentence-transformers MiniLM, ollama nomic-embed-textRAG/embeddings never billed

The DNA bank runs at $0.000030/week. Billed species are avoided during DeepSeek peak hours (UTC 01:00-04:00 & 06:00-10:00); free species and local are allowed anytime.

6. Why these 8 species (the research method)

  1. Rejected by genome quality, not fame. Providers/tools vetted by code, live probes, and error behavior - not popularity ("stars can be manipulated").
  2. Probe-verified before release. Every free species got a real API call before joining the park (200 + live completion), including fixing Groq's Cloudflare UA block.
  3. Smallest genome that clears the bar. For each task kind, the minimum species that produces usable output wins - local first, free next, billed only when needed.
  4. Replaced when a better genome appears. The park swaps species as better free options emerge (e.g. NVIDIA NIM added 2026-08-14 as a warm 8b fallback).
  5. Everything fails safely. Containment fences, vet checks, and the breeding facility as a last resort - no single provider outage ever shuts the park.