AI alignment and learning, one accountable move at a time

What is Yawn trying to make possible?

A person can begin with something they want to understand or change, keep their original words, explore possible games and choose a usable first interaction. The game is a bounded situation to participate in, rather than an instruction for an AI to pursue an unlimited objective.

The Guide distinguishes participant statements and choices, model proposals, unknowns and evidence. Scene artwork gives the situation a visible form; an appealing image is not evidence that the underlying goal has been achieved.

How does this relate to the AI alignment problem?

Yawn explores alignment at the interaction boundary: whose purpose is being served, what the participant actually chose, what a worker is permitted to do, what it may spend and how a correction changes later work. Attributed records, revision checks and bounded approvals make these questions inspectable.

These mechanisms constrain particular operations. They do not establish that a model has the right values, that every interpretation is correct or that AI alignment has been solved. A role label, an accepted proposal and a successful local test are different kinds of evidence.

What does learning or self-improvement mean here?

The intended loop is to orient, choose a bounded move, observe its result, retain a correction or lesson and evaluate a scoped update. The existing learning workflow records its basis, evaluation and rollback. A useful correction in one context does not establish a universal preference or a better general policy.

The ranking-policy workflow requires at least 30 observed cases for broader promotion. That is a program rule, not a published effectiveness result. Learning a transferable skill, producing a helpful artifact and achieving a real-world goal must be assessed separately.

This is an experimental combination of ideas and engineering mechanisms. No claim is made that it is the first, best or independently validated solution to human–AI learning.

What has not been demonstrated?

Yawn has not demonstrated a general solution to alignment, guaranteed business success or autonomous self-improvement without human judgment. Source code and automated tests can show specific implemented boundaries; they do not prove every deployed provider flow, every account configuration or real-world effectiveness.

Feature availability depends on the deployed version, current access, configuration and funding. A current review means it matches its recorded basis; its conclusions can still be wrong. Failed or uncertain work needs visible recovery rather than an automatic cascade of new paid tasks.

Yawn is experimental. Play at your own risk, keep meaningful decisions under your control and review important claims, permissions and costs before acting.

How can someone inspect or try the experiment?

Read the public documentation and open questions, then begin with a situation that matters to you. Keep the first move small, inspect the proposed direction and compare what happens with what you expected. Return with a correction or result that can be attributed to its source.

Public explanations do not start provider work. Free template access covers the shared starter definition; later AI processing and image generation require the applicable account access and a bounded estimate. Private encounters are not made public by browsing, a scene choice or a visibility label.

Explore the work and its limits

• Open questions about AI alignment — questions awaiting attributable answers.
• Proof, results and learning — why an output, an accepted suggestion and a verified outcome differ.
• Architecture and the .yawn record — how the shared interface relates to sources.
• Interception working paper and its public .yawn source — a proposed construct with explicit evidence boundaries.
• Questions for the press — historical source-backed answers; their stated dates and limitations matter.