Tools for disciplined decisions
Model breeding adds options: more candidate types, more routing policies, more deployment shapes, and more ways to spend evaluation budget. The tools in this section make those choices explicit. They do not replace an evaluation suite or approval process; they create a structured starting point that can be copied into an architecture decision record.
| Tool | Use it when | Output |
|---|---|---|
| Viability calculator | Comparing a structural change with the no-op option | Weighted benefit, weighted cost, decision margin, and sensitivity prompts |
| Architecture selector | Choosing a first implementation pattern | Ranked pattern shortlist and the assumptions behind it |
| Release-readiness checklist | Preparing a descendant for shadow or canary deployment | Local completion state, exportable checklist, and unresolved gates |
Privacy model
All calculations run in the browser. The site has no database, analytics service, remote form handler, or third-party JavaScript. Reloading the page clears the viability and architecture worksheets. The release checklist is stored only in the current browser when local storage is available; it can be cleared from the page.
Use the outputs as evidence, not truth
A numerical viability score is only as reliable as its inputs. A pattern recommendation is only as reliable as the stated constraints. Treat each result as a hypothesis that must be tested against measured latency, quality, safety, and operational evidence.
PROCEDURE use_decision_tool(tool, project_context)
assumptions <- WRITE_DOWN(project_context.assumptions)
estimate <- tool.CALCULATE(project_context)
risks <- IDENTIFY_MISSING_EVIDENCE(estimate, assumptions)
CREATE_ARCHITECTURE_DECISION_RECORD(
recommendation = estimate,
assumptions = assumptions,
validation_plan = risks
)
DO_NOT_DEPLOY_UNTIL(validation_plan_is_complete)
END PROCEDURERecommended order
Start with the architecture selector, use the viability calculator for each proposed structural change, and run the release checklist before any candidate leaves an isolated lab. Revisit all three after a meaningful change in workload, hardware, risk tier, or evaluation coverage.
Source reports used for this guide
These reports are preserved verbatim in the site archive. The guide above is an editorial synthesis and may narrow, qualify, or reorganize claims from the source material.