Keep pruning assumptions and deferred choices

These examples and illustrative results are independently authored teaching materials, not measured model results.

Use case

Cost comparison deferred B using a profile from prompt v1 and another runtime setting. Prompt v2 may change token texture, while B has no current measurement. An unfavorable projection is not a permanent verdict.

Mechanism

Record candidate, task/prompt identity, measured profile, projection and pruning assumptions. Mismatched evidence means deferred, not deleted. Prompt/config/calibration changes trigger reevaluation under matched conditions and an authorized bounded probe. Preserve registered quality/cost standards rather than guarantee optimum from uncalibrated ordering.

Bad example

Permanently delete B from a mismatched old projection, retain it after v2 and claim globally minimum cost.

Good example

Keep B deferred with the v1/setting/profile mismatch. After v2, inspect quality/cost prerequisites and propose a bounded matched-condition probe within authorized evaluation. Unmeasured current cost stays unknown; do not claim B necessarily expensive or global optimum. Record actual reopening/retention/exclusion evidence.

Why the change matters

Contextual estimates become false timeless rules when their premises disappear. Retained assumptions/candidates make changed conditions reviewable without promising a global winner.

Observable expectation

An illustrative table identifies deferred B, old profile and reopening on v2/setting changes; current cost is unknown. New judgments cite matched conditions, not priors or a lucky single result.

Keep candidates/quality gates visible. No paid evaluation is performed.

Limits

Projection error, quality variation and pruning assumptions matter. Frozen model/effort monotonicity, error factors/prices are not universal current laws. Define budget/permission and distinguish guesses from measurements.