Review build — 2026-10-09
The implementation follows the approved scope: original large-v5, dual learning first, the experimentally tested de2 PETase case, and an HF Docker Space with online generation.
Passed
- Production TypeScript/Vite build, including local fonts and bundled 3Dmol.
- Nine API/asset tests: input validation, queue bounds, serial worker reuse, failed generation handling, exact sequence–PDB correspondence, confidence and source checksums, and both benchmark splits.
- Chromium desktop (1440 × 1000) and mobile (390 × 844) rendering; no horizontal overflow or uncaught page errors.
- Both WebGL protein viewers; de2 and the selected IsPETase reference; cartoon/surface/sticks; confidence colors; residue selection and keyboard navigation; PDB download.
- Dual-learning direction switches, benchmark task/metric/set switches, model-size selection, mobile navigation.
- original de2 LoRA prompt → playground; archived examples in both directions; generated sequence → annotation input.
- Actual original checkpoint loaded from the research workspace. Both text→protein and protein→text jobs completed with a 64-token output cap.
- End-to-end browser submission to the live protein→text API completed and displayed a LIVE result. These short tests validate integration, not output quality or biological function.
Deployment status
The deployment targets REXWind/ouroboros on its existing CPU Basic hardware, using the Docker SDK. The previous Gradio revision is backed up locally and retained in HF history. A prior local Docker build was blocked by a Docker Hub timeout; the frontend and Python runtime passed local checks. Hosted build and live-generation verification are recorded separately after upload.
Local live preview: http://localhost:7861 (while the review service is running).
Evidence boundaries
The active case is de2 (279 aa), selected as the highest reported TPA+MHET concentration among four measured designs in the supplied 2026-09-27 workbook. The visible comparison focuses on de2 and IsPETase. The complete seven-sample source table, including LCC and BhrPETase, remains downloadable. Eight other submitted designs have no numerical values and are not treated as zero. Assay conditions and replicate information are awaiting confirmation. The original LoRA generation prompt and exact predicted sequence were recovered from lora_petase_3epoch, prediction index 2890.
ESMFold job ouroboros-wetlab-de2-1009b succeeded on 2026-10-09. All four structures exactly match supplied FASTAs. de2 CA-mean pLDDT: 93.46258; mean PAE: 4.79198 Å. The runner's unmasked atom37 average pLDDT (85.4938) includes absent atom slots; the UI consistently reports the mean over CA residues, verified against PDB B factors. Both calculation definitions remain recorded in the source manifest and exported case metadata.
Prompt builder follow-up
Production build and browser checks passed for all six original-demo presets; similarity/function editing; adding/removing cofactor; exact original-order prompt assembly; random preset selection; switching to protein→text and loading saved outputs; 390 px mobile layout. No uncaught page errors.
Wet-lab browser verification
All seven assay rows, the TPA/MHET/combined metric switches, exact de2 readouts and CSV download passed. The selected de2/IsPETase structure tabs, residue selection, confidence coloring, and the recovered LoRA prompt → playground passed. Desktop and 390 px mobile rendering have no horizontal overflow or uncaught page errors. Default structure framing was enlarged for the compact de2 fold.
Dataset and Core Idea update
Nine tests passed. A dedicated browser run validated all 29 key tiles, top/other filtering, search including the rare interaction key (2 records), empty-search feedback, coverage values, the guided tour with manual pause, changing input/target roles, and the IsPETase-only comparison. Desktop and 390 px mobile layouts have no horizontal overflow or page errors. Dataset membership (full dataset versus training set) remains unconfirmed, so the inferred 545,334-record denominator is documented in metadata and is not promoted as a confirmed training-set size.
Compact statistics and motion polish
Production build passed. The dataset section now defaults to five rows with an expandable, searchable 29-key list; its desktop section height is approximately 550 px including section spacing. Dedicated Chromium checks passed for both directions' progressive reveal, pause/resume, immediate completion, replay, stable action placement, complete downloads during partial playback, direction-switch cleanup, and reduced-motion behavior. The completed live-job UI was exercised with a mocked API; the unchanged backend's earlier real-model checks remain recorded above. Desktop and 390 px mobile screenshots were inspected, with no horizontal overflow or uncaught page errors. Output playback is explicitly distinguished from live token streaming.
Final deployment revision
The Core Idea exclusively owns recorded typewriter playback (large-v5 HSP20 example hard1-0151). Playground examples only populate inputs; outputs appear directly after real inference jobs complete. Browser verification covered both Core Idea directions, pause/skip, reduced motion, input-only presets and both completed-job UI paths (mocked API), with no page errors.
The supplied typedata.zip was extracted to the same OSS directory. All 545,334 records were checked against the CSV: all 29 key counts match exactly. Every case is a verbatim annotation from type_with_data_clean.json, with a record index, UniProt ID and source checksum. All 29 key/case selections, search, scrolling and desktop/mobile overflow checks passed. No external fallback examples are used. The full dataset is retained in OSS; only 29 small illustrative cases are shipped.
The remote HF model checkpoint and validated local original large-v5 have identical SHA256 f11f6295794d027e03d6844d5b6d685641ca04d3558536416663ffddaa0b4067. Nine API/asset tests and the production build passed.