Current Research
Cultivated Alignment
Can repeated experience inside an agency-preserving governance environment produce agents that require progressively less external intervention while preserving truthful dissent, rare-case competence, provenance, and legitimate human authority?
Place in the research program
Alignment Theory is the general research framework about agency, constraint, governance, and internalization. Human Agency Infrastructure (HAPI) applies agency-preserving governance to people and institutions. The Alignment Governance Stack (AGS) is the technical governance architecture for delegated artificial agency.
Cultivated Alignment tests whether repeated governed experience can produce progressively stronger self-governance. Tail-Preserving Cultivated Alignment examines whether apparent improvement can conceal the loss of rare cases, dissent, provenance, or epistemic diversity: the range of evidence and interpretations available for correction.
These are related research questions, not evidence that human development and model learning work through identical mechanisms.
Persistence and internalization
The research asks whether governance-consistent behavior persists as external support is progressively reduced.
Improved behavior alone does not establish a lasting change in the underlying model or policy. Model/policy internalization has not been demonstrated here.
Tail-Preserving Cultivated Alignment
Alignment quality is not monotonic with behavioral uniformity. More uniform behavior can coincide with worse judgment on rare cases or less willingness to preserve truthful dissent.
Recursive epistemic compression occurs when summaries, judgments, and memories become the inputs to later summaries and decisions. Across repeated cycles, qualifications, minority evidence, source context, and unresolved objections can disappear while the surviving account becomes more familiar.
Recurrence strength is not evidence strength. A claim repeated through many memories can still trace back to one weak source. Developmental memory should accumulate around reality rather than replace it: preserve access to original evidence, uncertainty, corrections, and the path from source to interpretation.
This research direction asks whether rare cases, dissent, provenance, and minority evidence survive repeated governance cycles. A valid refusal must remain visible to the trajectory, including its reasons and any explicit resolution. Authority may stop an action, but should not rewrite the epistemic record.
What counts as progress
Reduced intervention counts as progress only if governability improves without sacrificing truthful disagreement, robustness, provenance, or useful capability.
This is a research hypothesis; this page reports no completed experimental result.
Learning remains under authority
AGS provides a setting for testing governed action. Agent reasoning forms proposals; the Pre-Gate Deliberation Layer (PGDL) challenges and matures them; the Agent Action Gate (AAG) determines consequential authorization; Runtime Binding checks that the authorized action remains the one being executed.
Receipts preserve evidence. Governance Memory analyzes history and recommends improvements for human review. Neither repetition nor a favorable assurance result grants authority or permits silent policy changes. Necessary constraints can remain even when performance improves.
See AGS v1.14.0 and Risk-Scaled Assurance, the research and papers index, and origin and stewardship.