Project summary

Drytis proposes a research pilot to test whether independent checks and human review can identify unsafe behavior in AI coding agents that ordinary software tests miss. We would use controlled tasks with synthetic data, not live customer systems. The intended public benefit is practical evidence about where these safeguards help and where they fail.

What are this project's goals? How will you achieve them?

Our goal is to evaluate whether independently reviewed requirements, executable safety checks and targeted human review reduce the acceptance of unsafe AI-generated software. We would create tasks involving permissions, sensitive information and unauthorized changes. We would compare ordinary software testing against additional independent checks, both with and without human review, using predefined criteria for safe and unsafe outcomes.

How will you know it worked?

We would measure how often unsafe outputs are accepted, how often safe outputs are incorrectly rejected, task completion rates, and the additional time and cost of oversight.

A useful outcome would be a reproducible comparison showing whether the safeguards improve on ordinary testing. Negative results would also be valuable if they clearly identify limitations.

How will this funding be used?

Funding would support research design, engineering, independent review, model/API usage and evaluation infrastructure. It would support a defined research project rather than general company expansion.

The final budget, delivery schedule and staffing remain subject to confirmation. If other funders support overlapping work, we would disclose that support and avoid charging the same costs twice.

Who is on your team? What's your track record on similar projects?

Drytis is an AI-assisted software-development company combining software generation with human engineering support. Thomas Eide is the applicant contact.

The technical research lead, implementing engineer and independent reviewers have not yet been confirmed for this project. We are not presenting this proposal as completed or validated AI-safety research.

What are the most likely causes and outcomes if this project fails?

The safeguards may miss unsafe behavior, reject too many acceptable outputs, or require excessive human effort. Results may also depend heavily on the selected tasks or models and may not generalize to real-world development.

We would report these limitations and avoid making broader safety claims than the evidence supports. The planned evaluation would use a controlled environment to limit exposure of customers or sensitive information.

How much money have you raised in the last 12 months, and from where?

Alpha Labs has provided an estimated US$600,000–US$700,000 in funding since 2025, including approximately US$370,000 in equity. Drytis Inc. was incorporated in January 2026. We are reconciling the records to confirm the amount attributable to the company within the last 12 months and the treatment of pre-incorporation funding. These figures are provisional cumulative estimates, not a confirmed last-12-month total.