Semantic compression makes certain information disappear not because it was false, but because it was too costly to preserve in a concise, reusable answer.
What the phenomenon looks like
Low-frequency distinctions, exceptions, negative constraints, multi-step conditions, and local qualifiers are especially vulnerable. They survive on the page but vanish in synthesis because they add friction to a compressed representation.
Why it happens
Compression optimizes for portability. The more a detail depends on exact wording, local context, or multi-layer explanation, the more likely it is to be dropped when the model seeks a stable summary.
Why it matters
The loss is strategic, not cosmetic. What disappears first is often what made the source safe, bounded, or decisionally sound. The answer becomes clearer while becoming less governable.
What must be governed
- Identify which details are compression-sensitive and promote them to canonical status when necessary.
- Use repeated explicit negations and perimeter markers for information that must survive summarization.
- Do not evaluate summary quality only by readability; evaluate what was structurally lost.
