The supplemental-content injection SOP

The supplemental-content injection SOP

When core data is thin, teams reach for spun text. Don't. Inject genuinely supplemental data that differs per entity.

The ranked source list, best first:

1. Computed comparisons — Owner: data lead. Rank this entity against its peers on a real metric ("3rd cheapest of 40 in this category"). Gate: the number is derived, not invented.

2. Aggregated user signals — counts, averages, distributions you already store. Gate: minimum sample size or the module hides.

3. Related-entity tables — a structured table of neighbors, also useful for internal links. Gate: at least 3 rows or hide.

4. Time-series snapshots — "changed since last month" if you log history. Gate: history exists.

Forbidden source: LLM paragraphs that say the same thing in different words across every page. That is spun boilerplate wearing a costume.

— The test: would this module read differently if you swapped to another entity? If not, it's boilerplate
— Real differentiation comes from your data, not from a rewriter

Ship gate: don't inject a module until it passes the swap test on three sample entities.