Preventing Hidden Errors in Text and Everyday Calculations

Text that looks identical may not share the same bytes

A letter may be one code point or a base letter plus a combining mark. Zero-width and bidirectional controls can affect comparison, slugs, or security review while remaining invisible. Choose NFC, NFD, NFKC, or NFKD for the actual purpose; compatibility forms intentionally merge some typographic distinctions.

Preserve the original before normalization and inspect code points especially in identifiers, domains, file paths, and access rules. Normalization does not solve homograph attacks; it only makes representation differences easier to inspect.

Sorting, deduplication, and n-gram counting are different jobs

Locale-aware sorting creates a natural order, while deduplication deletes data. Ignore-case or trim matching belongs only where the domain rule supports it. Report source lines, unique output, and removed entries together.

N-gram counts expose repetition but do not measure intent, quality, or SEO success. Small samples inflate proportions, and segmentation for languages without spaces depends on browser language data. Read counts alongside their original sentences.

• Keep original order recoverable.
• Document the deduplication key.
• Report n-gram sample size.
• Do not label frequency as a quality score.

Percent, percentage points, tax, and discount

A move from 120 to 150 is 30 units and 25 percent growth; when values are themselves rates, the same 30-unit gap may be 30 percentage points. Percent change is undefined at a zero baseline, while negative baselines require contextual interpretation.

Sequential 10% and 5% discounts do not equal 15%, because the second applies to the reduced price. Extracting net value from a tax-inclusive price differs from adding tax to a net price. Obtain current rates, rounding, and document rules from an official source.

Keep context around duration and randomness

A date-time without an offset uses the browser's local zone. Daylight-saving changes, breaks, overnight spans, and payroll rounding can alter net time. A general duration result does not decide statutory worktime or pay entitlement.

Random selection is only unbiased relative to its input list. Web Crypto supplies strong randomness, but duplicate entries, eligibility errors, or post-selection intervention can still make a process unfair. Use a recorded independent process for regulated drawings.