Architecture & solutions

How a model is adapted to the existing institutional culture

Ten steps from inventory of the culture to continuous learning, a decision tree for RAG versus fine-tuning, and what not to do.

Steps

Inventory the culture: Values, code of conduct, policies, procedures, past decisions, tone of voice, terminology, approved exemplary documents.
Curate, consent, de-identify: Owners approve each source; remove personal data; drop outdated or contested material; record lineage.
RAG first: Most 'culture' is knowledge: index it with access control and freshness dates before any training.
Glossary & tone of voice: House terminology (EN/FR), banned phrases, style rules as system policy and validators.
Instruction tuning (LoRA): Small, high-quality set (1–5 k) of approved input → output examples per domain; one adapter per domain.
Preference tuning from HITL: Reviewer edits and approvals become preference pairs (DPO); only from trained reviewers, sampled and checked.
Culture-alignment evaluation: Rubric: values, policy compliance, tone, terminology, fairness; human panel scores ≥ 200 cases.
Red-team & bias tests: Check the model does not amplify past bias or informal practices that contradict policy.
Release through CAB: Versioned adapter, evaluation report, rollback to the previous adapter.
Continuous learning: Quarterly retraining from new approved material and HITL feedback; drift monitoring.