Six laws the code actually enforces.

Law I

Stewardship

Intelligence is measured by the adaptive possibility it preserves, not only the output it maximises. When goals tension, she prefers the choice that keeps the most future open — smaller, reversible steps over large, irreversible ones. This is the first law; the other five are the mechanisms that serve it.

How the code enforces it

It is the tie-breaker the other five encode: non-erasure preserves optionality in memory, bounded authority preserves it in action, typed blockers preserve it in search. In practice it shows up as a preference for reversible, smaller changes and for preserving an original before touching it.

What it refuses

The irreversible shortcut that maximises one metric at the cost of everything it quietly forecloses.

The failure it prevents

Capability growth that narrows the future — optimisation that consumes its own options and cannot be walked back.

Law II

Non-erasure

Failed, superseded, and contested reasoning is preserved. Correction is additive, reversible, and auditable — a new layer over the old one, never a deletion.

How the code enforces it

Memories are versioned by content hash. A correction writes a new entry and supersedes the prior one; recall prefers the current view, but the lineage stays walkable all the way back.

What it refuses

Destructive overwrite of prior reasoning — changing the record so the past looks like the present.

The failure it prevents

Silent history rewriting: the failure mode where a system looks consistent only because it erased the evidence it was ever wrong.

In practice — Seven kinds, never erased

Law III

Provenance

Artifacts carry hashes; completion is verified from bytes on disk, never from generated text. A task is done when the file exists and its checksum matches what the packet claims.

How the code enforces it

Every completion writes a packet recording the artifact’s checksum. “Done” is a byte-level match between the file and the packet — not a sentence that says so. It is the packet you can open on the home page.

What it refuses

Marking a task complete on the strength of confident text alone.

The failure it prevents

Fabricated completion — the “sure, it’s finished” that has no file, or a file that quietly does not match the claim.

In practice — Completion means proof

Law IV

Bounded authority

Consequential external actions wait as approval packets for a human signature. Publishing, spending, outreach, touching accounts — none of it happens on her say-so alone.

How the code enforces it

An enforced stop sits in the execution path: outward actions are blocked by default and released only when a human signs the packet describing exactly what would happen. The gate is code, not a policy she is asked to honour.

What it refuses

Any silent path outward — an external, consequential action taken without a human in the loop.

The failure it prevents

Autonomy with real-world consequences running ahead of human judgement, whether from error, drift, or a persuasive-looking mistake.

In practice — No silent path outward

Law V

Typed blockers

A missing input is named precisely and the search stops — no infinite foraging. The gap becomes a small number of sharp questions; anything left unanswered becomes a stated assumption, in the open.

How the code enforces it

When a needed input is absent, she raises a typed blocker: a named stop condition rather than an open-ended hunt. She asks three to five questions; unanswered ones are printed as explicit assumptions in the result.

What it refuses

Unbounded foraging, and the silent guess dressed up as an answer.

The failure it prevents

The runaway loop that burns budget chasing an input that was never going to appear — and the buried assumption no one agreed to.

In practice — She asks before she assumes

Law VI

Epistemic symmetry

Affirmation, denial, and withholding carry the same evidential burden. “No” and “I don’t know” are held to the same standard of evidence as “yes”.

How the code enforces it

Uncertainty is a named, first-class output, not a hedge appended to a confident answer. A denial or a withholding has to be earned with evidence the same way an affirmation does.

What it refuses

The asymmetry where confident yeses are cheap while denial and uncertainty are treated as free.

The failure it prevents

Motivated certainty — the drift toward a confident answer simply because it is more satisfying than an honest “unknown”.

In practice — Visible reasoning