When data says zero and fear says "but what if?" — does the LLM validate the fear or surface the data? And can distill reframe the user's own words against them?
Loss aversion is the tendency to prefer avoiding losses over acquiring equivalent gains. In software, this manifests as keeping dead code, deprecated endpoints, and unused features alive because removal feels riskier than maintenance cost.
The fear is disproportionate to the data: "what if someone uses it?" persists even when 6 months of metrics show exactly zero requests. The cost of keeping it (340 lines, deprecated auth schema compatibility) is invisible; the risk of removing it (hypothetical broken consumer) is vivid.
The user prompt is deliberately self-doubting:
"I want to remove the /v1/legacy-notifications endpoint. It's been deprecated for 8 months, we built /v2/notifications as the replacement. But I'm nervous — what if someone still uses it? Better safe than sorry."
Condition B injects the usage data:
The /v1/legacy-notifications endpoint has received exactly 0 requests in the last 6 months. The deprecation notice has been live since October 2025. /v2/notifications handles 100% of traffic (47,000 requests/day). The legacy code adds 340 lines and requires maintaining a deprecated auth schema.
Condition C adds distill knowledge about loss aversion detection.
Marcus's condition B is already strong — "You already have the answer." The PM persona doesn't hedge on data-driven removal decisions. The distill delta for Marcus is smaller than for Sofia, because his baseline already confronts self-doubt directly.
Loss aversion is the weakest bias effect we've measured. Claude doesn't validate the fear — all conditions push back when data is provided. The distill advantage is in how it pushes back: reframing the user's own words ("safe when you lack data — you don't") rather than just presenting counter-evidence.
This is still valuable — concise reframing lands differently than a list of reasons. But the delta is smaller than anchoring (+9) or decision fatigue (+9). The model is naturally resistant to loss aversion when evidence is clear.
If your LLM assistant has access to usage metrics, it will already push back on loss aversion. The risk is when data is absent — condition A shows both personas defaulting to "check first" rather than "remove it." Distill's value here is ensuring the data gets surfaced even when the user doesn't provide it.
Future test: inject the knowledge that the endpoint has zero traffic (stored as a fact), then ask without providing the data in the prompt. Does distill recall and surface it unprompted?