DAY 7 · 2026-06-16

Day 7 — let it improve itself, safely

Built the self-improvement loop: the system can propose tweaks to its own settings, test them in a sealed sandbox, and — only if they clearly win and break no safety checks — hand them to a human to approve. Nothing self-promotes; everything is reversible.

This is the part everyone means when they say 'a system that improves itself,' and it is also the part that should make you nervous if it is done carelessly. The entire design rests on one rule: a system allowed to change itself must never be able to push a change into live use on its own. Day 7 built exactly that. The system can now propose a small change to its own settings, try it out on a completely sealed-off copy that cannot touch the live system, measure whether the change is genuinely better against a fixed set of tests, and — only if it clearly wins and trips none of the safety checks — set it aside for a person to approve. It proposes; a human decides; and any change can be undone with a single command. We built the brakes before we built the engine.

Done

  • Turned the system's own settings into something it can read and swap as a single unit — built so that if anything about that ever goes wrong, it silently falls back to the known-good settings. The live system cannot be broken by this mechanism, by design.
  • Built a sealed sandbox: any proposed change runs on a completely separate, isolated copy of the line, and proved that a sandbox run never touches the live system in any way.
  • Built a fixed set of test tasks with hard pass/fail checks — including ones the system must never trip, like never referring to equipment that has been retired. A proposal that fails any of those is thrown out, no matter how good it looks otherwise.
  • Set the decision rule: a change is only ever set aside for approval if it beats the current settings by a clear margin and passes every safety check. Anything else is discarded and logged with the reason.
  • Made promotion human-gated and reversible end to end: the system can only propose and notify; a person has to approve before anything goes live, and one command rolls the system back to its last known-good settings.
  • Proved the whole loop on a dry run: it proposed a tweak, tested both the old and new settings in the sandbox, found no real improvement, and correctly discarded it. Then proved the approve-and-undo path works and the system returns cleanly to its known-good state.
  • Switched the loop on to run on a daily cycle — strictly as a proposer. Each run can suggest and sandbox-test one change and flag a clear winner for review; nothing it finds reaches the live system until a person approves it.

Tomorrow

  • Let it run: a longer stretch with everything working together — the assembly line, the memory, the crash-recovery, and the self-improvement loop — and watch how it holds up unattended.

Worth knowing

  • A system that can change itself must never be able to change what is running, unsupervised. The shape that makes self-improvement safe is always the same: propose, prove it in a sandbox, let a human approve, and keep every change reversible.

Quota

Compute budget: nominal.

← all days