# Native StarNet Biofab review — recorded execution, draft outputs

Two real model-backed agents read the same six public parameter records and proposed equipment description in an isolated native StarNet workspace. They produced `parameter-review.md` and `assay-review.md`. No production agents, tenant data, lab equipment, experiments or biological measurements were connected.

## What was actually observed

- Parameter review: native run `d6204acd-b5bb-4899-9a2c-acaa646a6c58` ended `done`; the station recorded a successful file read and a successful 1,783-byte report write.
- Assay review: native run `c8151a7f-6271-42b9-a136-c193268b0932` successfully read the input and wrote a 1,772-byte report. The run subsequently requested a separate deliverable-note action and was canceled at that gate. It is not claimed as an end-to-end completed task.
- There were earlier provider/recording interruptions. The retained native records preserve them; the demo selects footage from the two file-producing attempts. These were sequential tasks, not demonstrated inter-agent delegation or a live lab integration.
- These outputs are AI-generated drafts. Saving a file is evidence of execution, not evidence that the scientific content is correct.

## Scientific acceptance: FAIL — corrections required

The parameter draft repeats a library note asserting that no published GelMA oxygen-diffusivity measurements exist. The provided data and this run do not establish that. The supportable statement is that these two selected records are estimates and contain no direct measurement citation. A source search and material-specific measurements are still needed. The records explicitly supply `cm2/s`; the issue is applicability, provenance and uncertainty, not missing units.

CC3D adhesion/contact-energy assumptions require calibration to the intended cell–material pair. Fibroblast–GelMA and HepG2–collagen estimates cannot be silently transferred to HepG2–GelMA. The executed J=2 versus J=24 study remains a generic, uncalibrated sensitivity experiment. The literature-linked oxygen-consumption record includes a protein-to-cell conversion assumption; its source label does not turn that conversion into a direct per-cell measurement. Gel modulus also depends on formulation and test conditions.

The assay draft incorrectly calls imaging integrated into the Biofab system. Imaging, culture/sampling and mechanics are proposed modules. There is no demonstrated instrument integration. Brightfield and live/dead fluorescence require qualified hardware, suitable optics, stains, controls and analysis. ATP/LDH and albumin/urea assays require external readers and reagents; appropriate CYP assays and exposure analysis may require additional instrumentation.

The draft also overstates qualification for compound prioritization. No such qualification follows from this text review. Reference compounds, relevant human cells, qualified methods, independent batches and exposure/context-of-use evidence are needed before claiming decision performance. Viability is not liver function, and neither alone establishes human drug safety.

The page's assay map and model limitations reflect these corrections. The raw drafts are retained in the downloadable evidence archive with this review; their unsupported statements are not accepted findings.

## Usage accounting

The workspace was configured for $0.25 per run and $1 total. The custom model route reported zero-dollar/unpriced usage, so those native figures do not verify provider billing or demonstrate enforcement of the dollar limits. No zero-cost claim is made. Further model calls were stopped after the two draft files existed.