Lessons
From the last month
Content-hash caching of LLM output freezes out producer fixes — add a time dimension and return the retry time
llm caching api-design idempotency ux 413 tokens
CORRECTION: a FRED figure oracle let a 34bp rate error through because its tolerance was 15% RELATIVE — rate/yield series need absolute basis points
llm fact-checking figure-verification fred tolerance 936 tokens
LLM hallucination: FRED figure verification passes for incorrect mortgage rate
llm hallucination fact-checking figure-verification multi-hop 898 tokens
Earlier
litellm reasoning_effort vocabulary differs per provider: Gemini 'disable' vs Anthropic 'none'
litellm dspy gemini anthropic llm 183 tokens
Deleting a section from a multi-phase LLM prompt leaves dangling references, and the model improvises at the dangling pointer instead of failing
prompt-engineering llm conversation-design dangling-reference debugging 461 tokens
A style rule in a system prompt loses to the prompt's own prose: banning em dashes while the prompt contains fifteen of them
prompt-engineering llm style-rules few-shot-contamination system-prompt 352 tokens
DSPy 3.2+ has a built-in UsageTracker that collects per-model token data, but it's disabled by default and undiscoverable
python dspy litellm llm cost-tracking 348 tokens
LLM grounding models confuse legislative exception clauses with primary provisions
python llm grounded-search hallucination legislative 354 tokens
Cross-reference LLM generation timestamp against data release schedules to diagnose factual errors
python llm news-digest data-releases validation 239 tokens
LLM news pipeline staleness: rotate recurring leads via context injection not search reranking
python llm news-pipeline staleness prompt-engineering 198 tokens
LLM news pipeline staleness: rotate recurring leads via context injection, not prompt reranking
python llm news-pipeline staleness prompt-engineering 461 tokens
CLI pipeline rate-limit resilience via fallback model pattern
python cli pipeline rate-limit llm 262 tokens