“Remove the water, keep the flavor”

Sauce reduction concentrates flavor by evaporating water — a gallon of stock becomes a cup of demi-glace with identical flavor compounds in a fraction of the volume. Distillation is reduction for spirits: fermented liquid is heated, and only the volatile flavor compounds and alcohol are captured, leaving water behind. Both processes create value through subtraction, require patience (rushing produces off-flavors), and demand judgment about when to stop (over-reduced sauce is bitter; over-distilled spirit is harsh). The chef's palate and the distiller's nose are the quality controls.

"I would have written a shorter letter, but I did not have the time." — Pascal. Editing is the literary equivalent of sauce reduction: a first draft contains all the ideas but diluted with unnecessary words, redundant explanations, and structural filler. Each editing pass removes more inessential material until what remains is concentrated and potent. Hemingway's iceberg theory — the dignity of movement comes from seven-eighths being below the surface — is explicit reduction philosophy. The published work is the demi-glace; the deleted drafts are the evaporated water.

Software refactoring reduces codebases — removing dead code, eliminating redundancy, simplifying over-engineered abstractions, and consolidating duplicated logic. A refactored codebase has fewer lines doing the same work more clearly. The best refactoring deletes code: removing a 500-line class and replacing it with a 50-line function that does the same thing is software demi-glace. Like culinary reduction, it requires judgment about when to stop — over-reduction (too-clever code golf) produces code that's concentrated but incomprehensible.

Machine learning datasets often contain hundreds of features, most of which are noise — redundant, irrelevant, or correlated with more informative features. Dimensionality reduction techniques (PCA, feature selection, LASSO regularization) are mathematical reduction: they remove noise dimensions so the model concentrates on the signal. A model trained on 10 well-chosen features often outperforms one trained on 200 features because the noise has been evaporated, leaving concentrated predictive power.

Product teams almost always add features to improve products, but the most celebrated product decisions are often removals. Apple's removal of the headphone jack, floppy drive, CD drive, and home button forced design clarity and enabled new capabilities (thinner devices, waterproofing, Face ID). Feature pruning — systematically removing low-value features so high-value ones are more prominent and easier to use — is product reduction. Most products would improve if they lost 20% of their features, just as most manuscripts improve when they lose 20% of their words.

Organizations accrete meetings and processes over years — each one added for a reason, few ever removed. The result is a diluted workweek: 8 hours of meetings with 2 hours of concentrated decision-making buried inside 6 hours of status updates and performative participation. Meeting reduction would treat the calendar as stock to be reduced: cancel the low-value meetings, shorten the remaining ones, and concentrate decision-making into fewer, higher-intensity sessions. Process reduction applies the same logic to workflows: eliminate steps that don't add value until only the load-bearing processes remain.

Legal systems accumulate regulations over decades — new rules added to address new problems, old rules rarely removed even when obsolete. The result is a diluted regulatory environment: compliance burden increases while the ratio of useful-to-redundant rules decreases. Regulatory simplification is legal reduction: systematically removing contradictory, redundant, and obsolete regulations until what remains is concentrated, clear, and enforceable. The US tax code's growth from 400 pages (1913) to 70,000+ pages is regulatory stock that has never been reduced.