Model Selection
In one line: select an approved model route using task evidence, consequence, privacy, reliability, latency and total cost—not a universal “strongest” or “cheapest” rule.
Model and reasoning settings are versioned adapter bindings. Record provider/deployment route, model identity, supported tools/context, permitted data classes, retention/processing constraints, quotas and evaluation evidence. A “private cloud” label alone does not establish approved processing, network isolation or provider access controls.
| Work | Selection concern | Evidence to obtain |
|---|---|---|
| Business/architecture interpretation and coordination | Incorrect scope or authority can propagate widely | Representative ambiguous/adversarial cases, boundary handling and reviewer outcomes |
| Security, API and migration review | Both missed defects and false positives matter; none is merely deterministic pattern matching | Known defects, valid counterexamples, scope coverage and adjudication quality |
| Bounded implementation | Correctness and maintainability across the allowed change | Tests, review, failure recovery and total task cost |
| Mechanical lookup/formatting | A deterministic tool may be more suitable than a model | Exact subject, correct output and failure handling |
A capable controller is valuable when decomposition and adjudication are demanding. That does not prove maximum reasoning is always faster or cheaper. Measure end-to-end latency, retries, tokens/tool cost, useful outcomes and human correction effort against a representative task set. Use current available models only within the customer's approved routes and the user's chosen constraints.
The model fields in reviewer templates are reference bindings, not immutable policy or availability guarantees. Verify how the selected host interprets them. A model change may require evaluation and profile revision even if the methodology itself does not change. A dated adapter example records one setup, not a universal current deployment.
Large context windows do not justify loading every repository file or private document. Use scoped retrieval and source identity; measure actual context use and omissions. Tool availability and task need—not a single token threshold—determine navigation choices. A model's internal reasoning capability does not remove the need for externally checkable evidence.
Changing model, adding agents or using fresh context does not create a new authority or an independent oracle. Protect accepted expectations, test on adverse cases, and assess correlated failure risks. See §§5.3–5.6 for review, ownership and dispatch.