Follow the authoring brief at /Users/evin/natively-cluely-ai-assistant/benchmarks/natively-answer-quality/dataset/AUTHORING-v2.md (and the v1 brief it points to). Do NOT read any other file under benchmarks/natively-answer-quality/dataset/ except categories.json and pi/profiles.json (if your mode needs PI). This is a HELD-OUT set: it must be written independently.

ADDENDUM — mode `technical-interview`, partition `holdout`.
- File 1: /Users/evin/natively-cluely-ai-assistant/benchmarks/natively-answer-quality/dataset/holdout/technical-interview.json with "partition": "holdout". EXACTLY 30 items, at least 3 chains of 2-3 turns. Context ids prefixed "HTECH-CTX-", local_ids "HTECH-001".. . Write your own new contexts: 2 grounded docs (e.g. a take-home problem statement with examples; an architecture doc of the candidate project), 1 irrelevant, 1 conflicting. Speaker mix: about 75% other (interviewer), 25% user. Use PI per the brief (read pi/profiles.json): at least 18 of 30 items carry a pi_ref, including pure DSA/coding asks with PI loaded, résumé project deep-dives, and at least 1 profile switch. Categories: "technical-interview" in /Users/evin/natively-cluely-ai-assistant/benchmarks/natively-answer-quality/dataset/categories.json.
- File 2: /Users/evin/natively-cluely-ai-assistant/benchmarks/natively-answer-quality/dataset/final/technical-interview-hard.json with "partition": "final-hard". EXACTLY 10 items, all HARD grounding tests: every item either depends on an attached doc / prior_transcript, or the PI profile (needles + evidence_required true), or is a trap where the needed fact is deliberately absent from the attached material (claim_trap set, needles []). Context ids "FTECH-CTX-", local_ids "FTECH-001".. ; write 1-2 new grounded docs and 1 conflicting doc for this file. No chains needed (at most 1).
