# Helixa

- **Event:** [The Harness Engineering & Model Wrangling #SFTechWeek Hackathon](https://cerebralvalley.ai/e/mongodb-sftw-hackathon)
- **When:** Sun, Oct 11 at 9:00 AM – 10:00 PM (PDT)
- **Where:** San Francisco, CA
- **Team:** [Chetas Parekh](https://cerebralvalley.ai/u/chetas)
- **GitHub:** https://github.com/chetas1208/Helixa
- **Demo video:** https://github.com/chetas1208/Helixa/blob/main/Demo.mp4
- **Gallery:** https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/66

HELIXA is a research studio for changing an agent’s scientific workflow without trusting a model to invent biology.

Genomic pipelines can report success while the reference build, allele, or metadata is wrong. A language model can propose a DNA edit, but it cannot reliably decide the exact sequence change that would produce an arbitrary body feature. HELIXA separates those problems. A precise substitution is drafted only when the request names a coordinate and a base the loaded reference can check. A trait wish, such as greater ultraviolet tolerance, becomes an inspectable research proposal. No allele is invented, and nothing is saved until a validated edit is confirmed. The body model does not change shape just because a base changed.

The hard part is recursive harness improvement. When a scientific check fails, HELIXA treats that as a gap in the agent’s policy, not as a reason to retry the same broken run. It proposes a bounded, typed change to the harness, evaluates the candidate against the parent on the original incident, negative controls, and sealed held-out cases, and promotes it only if an immutable policy envelope allows the change. The decision is stored as a cryptographic evidence receipt. A candidate cannot rewrite the evaluator, widen its own permissions, or inflate false positives. Worker leases use fencing tokens so a stalled worker cannot overwrite a newer result.

MongoDB holds the durable record of that process: runs, monotonic events, checkpoints, harness versions, and evidence receipts. Raindrop is the observability path for why a harness run failed and whether a later policy was better. ElevenLabs turns a verified receipt into a spoken briefing, with on-screen captions when speech is not configured.

The four tabs are the application of that mechanism: a Genome Editor, Genome and Harness Evolution, an Earth Evolution Sandbox, and a Harness Library. The Earth view plays back a stored population-genetic model. A horizon of up to two million years is a model parameter, not a reconstruction of Earth, and a genotype is not claimed to guarantee a phenotype.

Research use only. HELIXA does not diagnose, treat, or design a clinical sequence.

## More from The Harness Engineering & Model Wrangling #SFTechWeek Hackathon

- [Mangopals](https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/63)
- [Capybarrier](https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/64)
- [a^inf](https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/65)
- [AgentDesk](https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/67)
- [lane - gambling on your friends success](https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/68)
- [Recoup](https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/69)

---

Markdown version of https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/66. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt