Needle in a Haystack

Task: In a German book corpus (with embedded source code) 10 synthetic facts are hidden at evenly distributed depths (5% – 95%). The model must retrieve all of them. Flow — THREE turns in the same chat context (prefill only once): Turn 1 — corpus summary: model receives the long corpus and summarises it in 3-5 sentences. Forces real processing of the text. Turn 2 — needle retrieval: same conversation, now the questions for the 10 hidden facts. Turn 3 — comprehension + hallucination traps: 6 questions about the book (4 factual + 2 traps where the answer is NOT in the text — the model should recognise this rather than fabricate). Default mode runs ONE uniform stage for all models: 120k tokens. Models without sufficient max_context are skipped at this stage. `niah_deep` additionally runs 32k / 64k / 200k for a full heatmap. Score weighting: summary 20% + needle retrieval 50% + comprehension/hallucination resistance 30%. Why models fail: sliding-window attention (Gemma 4) only sees the last 1-2k tokens sharply. Reasoning models hit the token limit before answering. Q4 KV cache measurably degrades recall at long contexts. On the hallucination traps the helpful bias lures models into plausible-sounding inventions.