Alexa+ × ChipLoop: Evidence-Labeled Experiment Assistant

Inspiration

Researchers need a compact, reviewable way to turn a design-grid question into a next experiment while keeping assumptions and uncertainty visible. This project explores an Alexa+ style conversation for that task, using the hackathon rules' simulated-experience path.

What it does

Text requests such as “What should I test next?”, “What is the risk?”, and “Explain this recommendation” route to bounded local tools: recommendation, risk check, explanation, simulated observation, and session status. A deterministic ChipLoop-inspired scorer ranks a synthetic dose/time/flow grid. The browser shows the selected parameters, uncertainty estimates, risk flag, and session audit trail. Optional browser speech synthesis can read the response aloud.

In the initial state, C104 ranks first: dose 8.111308 uM, time 36 h, flow 4 uL/min, predicted response about 0.576. Its nominal toxicity upper estimate is 0.351170635, above the 0.35 reference cap, so the app flags it for human review. The cap is a soft scoring penalty, not a hard safety constraint. A simulated C042 observation updates the observation count from 8 to 9 and available candidates from 136 to 135.

How it was built

A new static HTML/CSS/JavaScript browser implementation adapts my earlier BM02 ChipLoop research prototype. It runs locally without a cloud service, external LLM, Amazon SDK, Alexa device or live Alexa+ API. The intent router and synthetic scorer are deterministic and inspectable. The repository includes source, MIT license, setup instructions, file manifest, and a 12-check local self-test.

The public 72-second silent video is assembled from six actual browser screenshots in order, each held for 12 seconds: initial state, recommendation, risk, explanation, simulated C042 observation, and session status. It is a still-frame walkthrough rather than continuous screen recording.

What I learned and challenges

A conversational prototype needs to expose underlying values and evidence state. Separating intent routing from scoring keeps the workflow reviewable. Cross-checking generated descriptions against actual outputs was necessary: the corrected walkthrough shows the real C104 parameters and the above-cap warning. Local software checks establish reproducibility, not scientific validity.

Evidence boundaries

All data and outputs are synthetic and nonclinical. Nominal intervals are assumption-dependent and uncalibrated. Physical and clinical validation are NOT_RUN. No real experiment, patient data, safety guarantee, experimental approval or live Amazon-service integration is claimed. Human review remains necessary.

AI assistance, provenance and license

OpenAI Codex assisted with implementation and wording. The author reviewed the package and remains responsible for it. The README identifies the same-author BM02 source provenance and distinguishes prior MIT code, Apache 2.0 Kaggle notebook and CC BY 4.0 write-up licenses; the prior notebook/write-up are not redistributed. This submitted browser package is MIT-licensed and contains no credentials, patient data or private competition data.