Skip to content

README.md

Latest commit

 

History

History
41 lines (28 loc) · 2.35 KB

README.md

File metadata and controls

41 lines (28 loc) · 2.35 KB

Sovereign Alignment Work Group

Purpose

Own the participant-specific pipeline that turns a shared capable base into models that reflect local knowledge, values, institutions, domains, and interaction norms. This includes culturally grounded continued pretraining, post-training alignment, instruction tuning, and portability of sovereign contributions.

Why it exists

Sovereign alignment is Tapestry's primary differentiator in Design Goal 1 and TAP-003. TAP-005 makes post-training and instruction tuning first-class stages rather than downstream polish.

Scope

  • Participant-owned continued pretraining on culturally grounded and domain-specific data.
  • Post-training alignment methods such as DPO, RLHF, constitutional AI, and preference modeling.
  • Instruction tuning and chat-readiness patterns that preserve local norms.
  • Portability of adapters, alignment layers, or sovereign model forks across base models.
  • Requirements for culturally grounded data and value elicitation.

Out of scope: shared-base training policy, certification decisions, and production serving infrastructure.

Initial questions

  • When is continued pretraining required, and when are adapters or post-training enough?
  • What evidence shows that a sovereign alignment pipeline changed culture-specific behavior rather than only surface style?
  • How can sovereign layers remain portable when the base model changes?
  • Which parts of safety are universal base constraints versus sovereign alignment choices?

Early deliverables

  • A reference sovereign alignment pipeline for one pilot participant.
  • Requirements for culturally grounded data and preference collection.
  • A portability strategy for sovereign layers across base-model updates.
  • Evaluation requirements for cultural fit, domain fit, and instruction quality.

Interfaces

  • Data Governance: culturally grounded corpora, preference data, and usage rights.
  • Base Model Training: base compatibility and transition planning.
  • Evaluation & Certification: alignment metrics, acceptance gates, and evidence.
  • Security & Privacy: safety preservation and private alignment data.
  • Deployment & Adoption: product behavior, UX expectations, and feedback loops.