DEC-0001 — Adopt OKF as the canonical knowledge format

Context

We need a single-source-of-truth format for institutional memory that the client owns, survives model swaps, and that both humans and agents can read. This is the sovereignty test in Dossier — Mission & North Star.

Options considered

1. OKF — Open Knowledge Format (Google Cloud, v0.1, 2026-06-12): a directory of markdown + YAML frontmatter, one mandatory type field, git-diffable, no SDK.
2. Custom JSON schema + database.
3. Notion / Confluence as system of record.
4. Proprietary knowledge-graph store.

Decision

Adopt OKF as canonical output and system of record. The client's OKF git repo is authoritative; any vector index or vendor catalog is a replaceable cache.

Rationale

• Sovereignty & portability — plain files in the client's own git; no vendor lock (not even Google). Directly satisfies "switch the model, keep the veteran."
• Dual-surface — same files render as a human wiki and parse as agent context / a graph.
• Diffable & auditable — git history is the change log; provenance travels with each atom.
• Verified in deep research (OKF SPEC + Google Cloud blog).

Consequences

• We build with and to OKF everywhere; verticals extend types (only type is mandatory).
• Adoption risk: OKF is v0.1, governed in a Google org — bounded, because the format is trivially portable; we can fork/freeze the spec if it stalls.