Golden Paths and Phase Gates Across Decision-Ladder Levels
Explore how golden paths and phase gates standardize AI development across decision ladder levels. Understand required evidence bundles for security, observability, and cost control. Learn to maintain quality service levels and implement controlled progress from prompting to retrieval, tool use, and agents to ensure safe, compliant production deployment.
A platform reference architecture says which components are shared. It doesn’t say what a team building on top of it is actually allowed to do at each decision-ladder level, and without that, delivery pressure will find the gap. Golden paths are standardized routes that bundle the minimum artifacts and phase gates required to operate safely at a given decision-ladder level. Require repeatable evidence for quality, security, observability, and cost controls before a capability reaches production. Do not let a team skip evaluation or approval when moving from prompting-only to retrieval, tools, or agents.
Hold a few requirements fixed across every path. Regulated data must stay inside a defined trust boundary. Access must be least-privilege and auditable. A production change must be reversible without waiting on a bespoke expert intervention. Delivery still needs to move quickly, so a golden path defines what’s allowed by default at every level and what evidence must exist to justify stepping up. Track three measurable outcomes throughout: a stable quality SLO, bounded spend, and a reviewable chain of design decisions.
A golden path pairs an allowed component set with a small bundle of required artifacts that travel with the workload. At minimum, bundle an architecture decision record (ADR-<id>), a non-functional requirements set (NFR-<id>), an evaluation specification (eval-spec-<id>) pointing to an approved golden dataset, and a security packet pointer with a threat model reference and a data classification statement. A phase gate consumes these artifacts as evidence. Tie every gate to a risk tier and a quality SLO, so the gating decision tests whether the evidence is sufficient for the intended blast radius, not whether the implementation style is fashionable.
The following table shows a matrix view of ...