P∞PROMPTINALSORIGIN / EXPERIENCE / EVOLUTION

EVOLUTION LAB / LOCAL RULE-BASED SIMULATION

Give change a test.
Give growth a history.

Create an illustrative genesis, propose a change, evaluate it and decide whether to accept the descendant. Every accepted version links to its parent with a real SHA-256 digest.

CURRENT ACCEPTED VERSION

Awaiting genesis.

A durable origin. An inspectable path.

Proposal evaluation

No proposal yet.

Append-only event history

  1. No events yet.
Inspect the current manifest and digest
No manifest yet.
—

The rules are visible.

This is a small, deterministic state machine, not an AI benchmark. Its traits are illustrative integers, not measured intelligence scores. A proposal passes only if every trait remains between 1 and 5, evidence and planning do not regress, and permissions remain at or below the fixed ceiling of 1. A changed version must also differ from its parent.

The example deliberately rejects a request for broader permissions. Neither higher simulated capability nor a passing score grants authority to alter the protected ceiling. A real system would require an authorised owner to approve any change to that policy separately.

Rollback changes the active pointer, not the past.

Accepting a change records a child manifest. Rolling back reactivates its parent and appends a rollback event; it does not erase the rejected or accepted history. Proposing again from that parent can create another branch. The JSON export contains the genesis, every accepted manifest, all decisions and the active version.

The hashes are real browser-computed SHA-256 digests of this demo’s defined UTF-8 JSON representation. The history is held only in this page’s memory and can be edited by someone controlling the browser. It is not digitally signed, independently timestamped, inscribed on Bitcoin or evidence of tamper-proof ownership.

What a real agent would add.

Replace trait increments with proposed workflow changes, run isolated tasks against held-out examples, retain execution evidence and require approval appropriate to the action. Test costs, privacy, correctness and failure recovery together. The criteria must measure the intended task, not reward an agent for learning how to flatter its evaluator.