The calculator, not the model, is the knowledge base

Frontier LLMs write confident-but-wrong numbers. Labwright changes the contract: the model only proposes raw inputs (geometry, flows, seeding), and everything derived — shear stress, volumes, cell counts, dosing — is computed by deterministic physics calculators and re-proved by a verifier before a design is ever submitted.

Verified numbers, reverse-verified protocols

• No fabricated numbers. Every derived value is recomputed from its own raw inputs — an internally inconsistent design is rejected, not shipped.
• Reverse-verify published protocols. Paste a paper's geometry and flow; the verifier recomputes its shear/volumes and flags numbers that don't follow from the paper's own inputs.
• Domain-aware sanity. Physiological bands and unit aliases (µm↔mm, µL↔mL, dyn/cm²↔Pa) catch order-of-magnitude and unit-misread errors.
• Cross-domain. Microfluidics (OOC), plate culture, and 3D spheroid/organoid working volumes — one verified-design path.

Bare LLM vs Labwright

Two frontier models (deepseek-v4-flash / deepseek-v4-pro), five systems, six gold sets. A design is usable only if it is internally consistent and recovers every target within ±5 %.

Percentages are seed-pooled usable rates (reading / new-domains ×5 seeds, the rest ×3; Wilson 95 % CI in the repo). Bare-LLM numbers are the single run — the gap is so wide that no Labwright CI overlaps a memory-system CI on any set.

The seven post-v1 domains (barrier, oxygen, pumpless, breathing, pulsatile, scaling, gradient) were benchmarked end-to-end with the Labwright system on the 14 new-domain goals (no bare-memory baseline was run on this set). Flash submits 13/14, pro 11/14; among submitted designs hallucination is 0.000 on both. The honest boundary: the gradient-fgf8-pattern goal ends in silence on both models, and Labwright's fast-path front-end — the fixed local fine-tuned extractor, trained across all eleven domains on ~61k synthetic goals — answers 4/14 here (v4 scored 0/14), 5/14 with schema repair: the v6 retrain adds natural-register variants for these domains, closing part of the phrasing gap; the failures are silence-dominated rather than fabrication: six of the nine still-unanswered goals end in silence, the rest in wrong-target or calculation errors (base hallucination mean 0.512).

Memory-only systems (bare, soft-gate, self-verify) rarely produce a usable design. Labwright's derived numbers are architecturally consistent — the honest headline is the blind-set drop: the gate stops fabricated numbers, but it cannot supply domain knowledge the model does not have. That boundary is the point. The gap is stable across seeds (95 % Wilson intervals never overlap the memory systems), and an iterating fix-and-resubmit agent (labwright_iter) repairs all 41 verifier-fired entries yet is a wash on usable rate — iteration is a correctness loop, not a domain-knowledge loop.

The same verifier runs over the literature: a large-scale audit (21,094 SciRecipe protocol summaries, titles resolved to Crossref DOIs) is under methodology review — its lessons are shaping the evaluation framework. The verifier itself ships today.

Run it yourself

The CLI runs the exact same pipeline as the Space. Reverse verification is pure deterministic physics — no API key needed; design needs a DEEPSEEK_API_KEY.

• pip install -e . then labwright verify-protocol examples/verify_protocol.json — recomputes a paper's claimed shear/volume/Re from its own reported geometry and flow, flagging numbers that don't follow.
• labwright audit "liver-chip model of DILI at sinusoidal shear" — the offline fast path: a fixed local fine-tuned extractor proposes raw inputs and the same verifier re-proves every derived number. No API cost.
• labwright design "…" — the LLM agent loop (needs an API key). labwright ui launches the same two tabs as the Gradio Space.
• Full benchmark protocol, per-entry records, and honest caveats: the repo's eval/ and evaluation/ directories.