Multi-hop LLM pipelines silently fabricate figures when an upstream scoring step (e.g. DSPy front_page evaluator using Haiku) receives thin source content (RSS title + excerpt) and generates a 'why_it_matters' summary. The LLM invents illustrative numbers ('Even a 0.25% jump shifts affordability') to fill the gap. Downstream consumers (vibe writer) treat these as factual claims and quote them verbatim ('Yahoo Finance reported a 0.25% jump'), producing user-visible misinformation. The grounded search path had 'do NOT invent or modify any figures' guardrails but the front_page scoring path did not, creating an asymmetric trust boundary that the downstream consumer couldn't distinguish.
Add explicit figure-grounding constraints to every LLM step that produces text consumed downstream as factual: 'Only cite specific numbers that appear in the source content; never invent illustrative or hypothetical figures.' This is necessary at EACH hop in a multi-stage pipeline, not just the final output stage, because downstream consumers cannot distinguish grounded figures from fabricated ones. In DSPy, add the constraint to the OutputField desc for any field that may contain numbers (why_it_matters, summary, etc.). This is a mitigation, not a guarantee — LLMs still slip — but it reduces the fabrication rate significantly for small/fast models like Haiku that are more prone to filling gaps with plausible round numbers.