Build from scratch as the exception

The previous unit was about components inside the estate: discover an existing golden component before authoring a rival. The same instinct applies one level up, to whole engines and tools, with its own decision: ADR-0003, "Port best-in-class, rebuild native; references not dependencies".

"From scratch" does not mean blind reinvention

Cogitave is an agent OS/kernel company, and stitched-together third-party tools cannot be harmonic - each ships its own data model, UX, and boundary - so the moat is vertical integration over one canonical model, Cogitave Core. That "full-scratch" stance was being misread as blind reinvention: rewriting solved problems for their own sake, slowly and worse. That is explicitly not the intent. The best engineering teams have already discovered the right shapes for schedulers, parsers, inference loops, and permission models. Cogitave wants those shapes - then wants them native and vertically integrated, inside the single model.

The decision: port best-in-class, rebuild native

• References, not dependencies. Best-in-class tools - DocFX, Tauri, UniFFI, Candle, Base UI/Radix, lefthook, release-please, llama.cpp - are the spec and reference implementation studied and ported, not permanent runtime dependencies. Cogitave adopts the shape, not the lock-in.
• Rebuild native, maximum speed. The rebuild is harmonic (one Core model), closest to the machine, and measured against the reference.
• Reinvent only where you beat the reference. If a team cannot do it materially better - faster, more integrated, more correct, more secure - it does not reinvent yet. The reference stays a sanctioned bootstrap crutch behind a stable interface, explicit and scheduled for replacement.
• Correctness and certification compliance are the goal, not delivery speed.
• Operationalized by the Technology Radar. Every layer names its default, the reference it ports, and when to deviate; a changed default is its own ADR.

When "from scratch" is actually earned

Reinvention is earned, not assumed:

1. Can a best-in-class reference be ported (its shape studied and adopted) instead of ignored? If yes, start there.
2. Can the native rebuild be materially better? If not yet, keep the reference as the explicit, time-boxed bootstrap crutch and revisit later.
3. Is the deviation recorded as its own ADR, so a reviewer can always answer "reference, bootstrap, or native primitive?"

The ADR names its own accepted exception: an off-the-shelf tool is acceptable in the diagram render layer, because the notation spec itself stays Cogitave's own reference. That is the pattern to copy - not "no third-party tool, ever," but "the shape is ours to keep, the implementation is earned."