warranted-autonomy

Type: types/tag-readme.md

Notes on handing decisions from humans to a computational actor: which decisions an agent is warranted to take over, what makes that transfer justified rather than merely possible, and how to measure the autonomy a system actually has. The establishing claim is Warranted autonomy is bounded by oracle domain: bare autonomy is free, but warranted evaluation autonomy extends only to the candidates an oracle can assess with the required confidence. It is a child of self-improving-systems. Boundary: reflection asks whether a system can read and revise its own organization; this tag asks who is warranted to make that revision and the other decisions in the loop.

What bounds warranted transfer

• Warranted autonomy is bounded by oracle domain — the establishing claim: autonomy is warranted only where an oracle can assess the candidates
• A claim without external assessment carries three obligations — what a computational actor must supply when no outside assessor grades its claim
• Naur's human-only conclusion needs more than the absence of explicit criteria — unformulated judgment does not by itself put a decision beyond computation
• A methodology governs its own extension only as far as it settles the meta-decisions it raises — an actor may carry extension further than its method, but those choices are not thereby settled

The human–computation allocation

• Computationally directed self-improvement is a fixed-boundary reallocation ending in contraction — progress is which decision-bearing functions humans still supply; the endpoint is contracting the boundary to exclude them
• Methodological and computational closure track different changes — what a method settles and whether a human decided during operation are separate questions
• Warranted transfer out of the human cut leaves people the hardest-to-warrant decisions — easy-to-warrant decisions move first, so the human residue gets harder per decision
• Increasing computational autonomy relocates human effort to the frontier — human hours need not fall; count improvements per human judgment
• A method's ceiling bounds the method, not the transfer it already made — a responsibility that left the human residue stays transferred when the method stops improving

Measuring autonomy and its evidence

• Measuring autonomy well enough to see it improve is an open problem — a per-function profile, not a percentage, and not yet comparable across systems or time
• Tool usefulness, computational autonomy, warrant, and system power are separate dimensions — a progress claim must name which one moved; autonomy gains do not license power claims
• A benchmark that holds the client fixed exports the least-warrantable decisions by design — fixed-client benchmarks measure worker capability and leave broader closure untested
• A complete theory path does not establish improved capacity — mediation, empirical contact, and response to criticism each support a different claim
• Disconnected witnesses do not establish a full causal path through theory — evidence must show the joins of one path, and still separately show improved capacity

Related Tags

• self-improving-systems — the parent; this tag holds its actor-allocation question
• reflection — the self-revision whose decisions this tag allocates between humans and computation
• improvement-loop — the loop functions (search, evaluation, retention) that can be handed to a computational actor
• theory-builder — an autonomous theory builder is the case where computation performs every internal operation
• software-factory — where production decisions move from human developers to agents
• continual-learning — the deployed learning whose updates need an authorizing actor