Files

43 lines
2.0 KiB
TeX

%% SCRAP: archive/quality/regression/detection-framework
%% SOURCE: docs/working/archive/quality/regression/detection-framework.md
%% STATUS: HISTORICAL
%% FITS: none
%% EDITORIAL: lifted — prose rewritten to press voice
\section*{Optimization Regression Detection Framework (November 2025)}
The StarForth optimization program defined a multi-metric regression detection
framework to distinguish genuine optimization---improvement across all relevant
metrics---from simple speedup that trades one metric for another.
\subsection*{Regression Taxonomy}
Six regression types were identified:
\begin{itemize}
\item \textbf{Type~1: Metric Inversion.} Primary metric improves while a
secondary metric (memory, variance, stability) regresses. Detection
threshold: $>10\%$ secondary degradation.
\item \textbf{Type~2: Variance Introduction.} Mean improves but standard
deviation increases. Detection threshold: $>25\%$ variance increase.
\item \textbf{Type~3: Tail Latency.} Median improves but P99 worsens.
Detection threshold: $>15\%$ P99 degradation.
\item \textbf{Type~4: Cascading Failure.} Small-workload gains reverse at
production scale (e.g., dictionary growth, memory fragmentation over
1M+ instructions).
\item \textbf{Type~5: Workload Sensitivity.} Optimization overfits the
936-test training workload and degrades on other patterns.
\item \textbf{Type~6: Parameter Interaction.} Two independently-winning
parameters produce a regression when combined.
\end{itemize}
\subsection*{Decision Algorithm}
Each optimization candidate was evaluated as \textsc{accept} (no critical
regressions), \textsc{investigate} ($>2$ warnings), or \textsc{reject} (any
critical regression). The five-question validation test---is the optimization
optimal, stable, general, non-interacting, and maintainable?---served as the
human-readable gate before locking any parameter for subsequent experiments.
This framework governed the OPP~1 through OPP~5 experimental series.