Inspiration
Nature Assurance began with a simple question: can AI help people hear Nature?
Exploring that question through Exmoor revealed a deeper problem. Environmental evidence is dispersed, uncertain and often contested. Expenditure and completed activity can be presented as ecological success even when recovery has not been demonstrated. Fluent AI can make this worse by turning plausible inference into confident fact—or by appearing to possess authority that properly belongs to accountable people and institutions.
My experience in HM Treasury suggested another possibility: ecological claims could be governed with some of the discipline applied to public forecasting and accounting.
Before deciding what AI could do, I therefore asked what constitutional role it should be permitted to play.
The River Barle provides the practical case. Nature Assurance tests whether AI can organise evidence, separate observation from inference, challenge unsupported causation, refuse an unsafe conclusion and identify a responsible route forward—while leaving consequential authority and accountability with humans.
What it does
Nature Assurance creates an auditable route from contested ecological evidence to a human-authorised conclusion.
The demonstrator begins with one deliberately consequential claim:
Signal crayfish caused the reported decline in fish in the River Barle.
It then carries that claim through seven visible stages:
- Submit Claim
- Evidence Record
- Claim Decomposition
- Assessment and Challenge Review
- Stop Condition and Human Decision
- Authorised Account and Controlled Voice
- Revision History
The system distinguishes submitted information from admitted evidence; decomposes the claim into testable propositions; records what each source supports and does not support; and identifies an unsupported causal transition. A separate Challenge Review can object and trigger a stop, but cannot authorise or silently rewrite the assessment.
The original causal wording is refused. An identified human reviewer may instead approve narrower wording for a defined purpose and consequence level. The resulting Authorised Account preserves both the rejected and approved wording, the evidence links, uncertainties, reviewer, permitted use and audit history.
The Voice of Nature then receives only a narrow Authorisation Packet. It can interpret what has been approved, but cannot repeat or strengthen the refused allegation. Later evidence creates a linked new version without erasing the first.
The defining product promise is:
Refusal prevents overreach. Accountable next steps prevent paralysis.
How we built it
I used Codex as the implementation partner to turn a frozen constitutional specification into a working TypeScript and React application. Codex helped create the seven-screen transaction, typed data model, deterministic policy engine, audit trail, human decision gate, controlled Voice packet and automated acceptance tests. It also helped run and repair the build, lint and test workflow and prepare the repository for judging.
GPT-5.6 is used for the preliminary evidence assessment. It receives a deliberately restricted packet of admitted evidence and returns structured propositions, support levels, gaps, alternatives and a stop recommendation. The application—not the model—enforces admission status, authority boundaries, prohibited wording, permitted use and revision history.
The reference adapter uses the OpenAI Responses API with Structured Outputs. The hosted boundary calls GPT-5.6 when a server-side API key is configured. Otherwise it transparently uses the preserved GPT-5.6 River Barle assessment packet, so the demonstrator remains reproducible and never pretends that a fresh model call occurred.
The implementation uses React, TypeScript, Vite, Vitest, Zod and the OpenAI SDK. Seven constitutional acceptance tests verify that uploading is not admission; causal overreach is detected; Challenge Review cannot approve; human decisions are purpose-limited; refused wording cannot be re-authorised unchanged; the Voice cannot exceed its packet; and revision preserves history.
Challenges we ran into
The hardest problem was not generating an ecological report. It was deciding where the software must stop.
A fluent model can easily collapse observation, mechanism and causation into one persuasive sentence. We therefore had to make constitutional distinctions executable: uploading is not admission; assessment is not authorisation; confidence is not authority; and communication must not exceed the Authorised Account.
Another challenge was preventing human review from becoming a decorative button. Approval is therefore a recorded transaction requiring reviewer identity, role, wording, permitted use, consequence level, unresolved questions and a review trigger.
Scope was also a challenge. The wider project includes Exmoor mapping, ecological accounting and the Voice of Nature. For Build Week we reduced it to one complete constitutional transaction that judges can inspect from beginning to end.
Accomplishments that we're proud of
The refusal is implemented behaviour rather than a warning in a policy document.
The demonstrator preserves the submitted allegation, explains why it is blocked, records a narrower human-authorised conclusion and technically constrains the Voice to that conclusion. New evidence produces a linked second version without overwriting the first.
All seven constitutional safeguards are encoded as tests and pass. The repository includes setup instructions, the build contract, sample data, the GPT-5.6 adapter and a reproducible demonstration path.
What we learned
The product is not the analysis alone. The product is the governed movement of a consequential claim through evidence, challenge, human authority, authorised record and controlled communication.
AI competence can improve without its authority increasing. A refusal is useful only when it also identifies the evidence, referral or accountable decision needed next. And human oversight becomes meaningful only when the reviewer, purpose, limits and consequences of approval are recorded.
Most importantly, a constitution becomes credible when it changes observable system behaviour.
What's next for Nature Assurance
The next step is independent ecological and governance review of the River Barle evidence packet and the constitutional tests. A later version could add live evidence ingestion, claim-to-source visualisation and further ecological cases—but only after the single transaction has been tested properly.
One demonstrator may establish plausibility. It does not establish transferability. Wider use remains a question for later evidence and experience.
One claim. One refusal. One authorised conclusion. One controlled interpretation. One preserved revision.
Built With
- codex
- gpt-5.6
- openai-responses-api
- react
- typescript
- vite
- vitest
- zod
Log in or sign up for Devpost to join the conversation.