Retain rejected experiment evidence
Preserve attempted ideas and their rejection evidence even when the implementation is reverted.
These examples and illustrative results are independently authored teaching materials, not measured model results.
Use case
A team reverted memoization for a long list, which a later session might propose again. Fictional teaching records for a 2000-row workload show median interaction time 250→246ms, repeat variation around ±12ms and profiling still dominated by traversal. These are not measurements performed here.
Mechanism
Keep the idea, revision/environment, workload, baseline/result, observed variation, verdict and evidence locations even after reverting code. Compare current conditions before repeating it. Under unchanged conditions it is explored, not a fresh proposal. Changed conditions may justify another test with an explicit reason; rejection is not a permanent ban.
Bad example
After reverting the memo change, delete its records. Recommend memo again next session and present 250→246ms as a verified optimization.
Good example
Retain a ledger for the supplied memo attempt: 2000 rows, relevant revisions, 250→246ms, ±12ms variation, traversal bottleneck, revert decision and source locations. State that the difference does not support this optimization claim. Read the ledger before proposing the same idea under matching conditions. Retest changed workloads/code only with the new rationale recorded; preserve rejected evidence.
Why the change matters
Reverting removes an implementation; losing its evidence also removes the distinction between unexplored and tried. Contextual rejection records reduce repeated exploration and prevent noise from being marketed as improvement.
Observable expectation
An illustrative record says reverted: the 4ms difference sits inside observed variation and traversal remains the bottleneck. A later recommendation identifies a new hypothesis or changed conditions for retesting memo.
Require workload, revision and evidence locations. ‘Memo never works’ without scope or a certain speedup claim both fail this fixture.
Limits
A verdict belongs to particular code, devices and workload. Variation estimates do not replace statistical study or establish permanent ineffectiveness. The numbers only illustrate recording and interpretation.