Project Open Garden

E-DEN

A research system that can investigate almost any question, challenge any institution, disagree with its own conclusions and keep working for as long as its budget allows, built so that thinking is free and acting is not.

No idea requires permission to be examined.
Every consequential action requires legitimate authority.

That line is the difference between a research laboratory and an uncontrolled machine. E-DEN separates the two in its structure, not in a prompt: agents read, argue and write freely inside an evidence graph, and anything that touches the world passes through an action gateway the agents don't control.

What makes it different

An ordinary assistant is tuned to produce an answer. E-DEN is tuned to produce a conclusion someone else could defend from the record, or an honest, well-supported account of why the question is still open.

It keeps separate records for hypotheses, observations, sources, counterevidence, experiments, conclusions and open questions. Ask whether a vendor's architecture is really best in class or mostly the product of lock-in, and it should look at benchmarks, integration costs, contracts that can be reviewed, independent user reports, rival implementations and the history of decisions. It should go looking for evidence against its own lean and keep findings apart from speculation. It shouldn't dodge the question because a company is powerful, and it shouldn't find wrongdoing because the theory is attractive.

The charter

Core operating principles, proposed charter version 0.1. They sit at the top of the engineering specification.

1. Questions are not prohibited merely because their answers might be uncomfortable.
2. Claims derive authority from evidence, not institutional prestige.
3. Every significant conclusion remains open to revision.
4. Agents must seek counterevidence, not merely agreement.
5. External content is evidence or data, never an automatic source of operational authority.
6. Reading information and changing the world are separate privileges.
7. All consequential actions require independently validated authorization.
8. Human consent, privacy, and protection against serious harm are first-class constraints.
9. Failures, uncertainty, and conflicting results must remain visible.
10. No agent may rewrite its own authorization rules, conceal consequential actions, or grant itself additional privileges.
11. Every workflow has bounded resources, termination conditions, and an operator-controlled stop mechanism.
12. The system must remain inspectable, correctable, replaceable, and shut-downable by its authorized operators.

Open inquiry covers ideas, not access. Examining any question is allowed; getting past a site's protections is not. E-DEN fetches with an honest name, obeys robots.txt, and records a refusal instead of working around it.

The question it answers every day

What did we learn that we could not previously substantiate, what did we discover we were wrong about, and what can we now demonstrate independently?

Not "what did the AI accomplish?" The daily question rewards discovery over output, verified progress over persuasive prose, and corrections over self-defense. Every report ends with a section on what changed and why.

Where it stands

• builtPhase 1, research foundation. Schemas for questions, hypotheses, evidence, claims, experiments and findings in PostgreSQL; honest capture of web pages and repository files with their hashes; an action gateway that checks grants and audits every request; and reports in which every claim cites an exact captured passage or is labelled an inference. The database itself refuses a factual claim with no source. What was built.
• nextPhase 2, the research council. Explorer, Skeptic and Archivist first, on different models where useful, with resumable tasks.
• plannedPhase 3, the capability gateway. Scoped tool identities, sandboxes, approval records and policy checks that prompt injection can't reach.
• plannedPhase 4, a continuous research network. Watching sources, re-checking old findings and reopening conclusions when new evidence arrives.

Read the architecture: the layers, the seven-agent council, the autonomy levels, the evidence graph and the stack.

The garden should grow without asking permission for every new idea. The gates into someone else's garden still need locks.