# BD

- **Event:** [Built with Opus 4.7: a Claude Code hackathon](https://cerebralvalley.ai/e/built-with-4-7-hackathon)
- **When:** Apr 21 at 12:00 PM – Apr 27 at 2:00 AM (EDT)
- **Where:** Online
- **Team:** [Benjamin Dysin](https://cerebralvalley.ai/u/benjdy)
- **GitHub:** https://github.com/Attius-Digital-Art/lacunex
- **Demo video:** https://www.youtube.com/watch?v=UyxJRWqi-4I
- **Gallery:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery/213

**Cross-turn reasoning, rendered live. Both sides leave with something.**

Lacunex runs goal-directed adaptive interviews where structured insight is produced *during* the conversation, not in an overnight report. A Host (subject-matter expert) defines objectives; the platform conducts every turn live, surfaces cross-turn observations the moment they cohere, and at session close hands the participant a reflective takeaway worth keeping.

Built by a systems analyst, ten years in. Lacunex closes two gaps real-world interviewing never closes: beginners — analysts, PMs, anyone for whom interviewing is *part* of the job — don't yet know how to run a round well; experienced researchers do, but lack the time to interview enough people for long enough.

A four-call Opus 4.7 architecture runs every turn:
- **Conductor** — decides the next move from session state and renders the interviewer's turn.
- **Meta-Noticing** — observation-only; spots contradictions, hedging, and implied-not-said across turns. Every notice must cite ≥2 distinct transcript anchors (enforced in code, not just in prompt).
- **Extraction** (Haiku 4.5) — schema-bound; fills the host's live insight dashboard turn-by-turn.
- **Takeaway Synthesis** — a reflective artifact for the participant, including a "what you already have that is relevant" section.

Mapped against Anthropic's *Building Effective Agents* (Schluntz & Zhang, Dec 2024): **three of the five patterns layer on a single turn — Orchestrator-Workers, Parallelization, Routing.**

Three domain briefs ship — **Founder Investment Evaluation**, **Post-Incident Witness**, **Civic Consultation** — plus **Brief Designer**, where the platform interviews the host to author a *new* brief using the same four calls. Recursive dog-food.

**Audiences exercised during build:** VC and angel investors evaluating founders; city officials running civic consultations; SREs, safety officers and clinical QI managers conducting post-incident interviews; clinicians eliciting patient values; tacit-knowledge capture from retiring experts; consumer-insights teams; qualitative researchers; managers preparing for high-stakes conversations.

Empirical scale, on a **simulated 11-resident cohort** run against the Civic Consultation brief on the real platform: **303 turns, 54 deployed `◆` cross-turn observations from 243 considered, 12 cohort patterns, 6 routing recommendations.** Cost with prompt caching: **~$1–2 per session**, under $5 for the cross-cohort aggregate — putting a 100-resident consultation at ~$150 end-to-end.

**Not an overnight research moderator** (Outset, Listen Labs, Strella ship next-day reports). **Not a transcript analyser** (Dovetail, Condens ingest transcripts that already exist). **Not "Claude with a long system prompt"** — a single chatbot can't enforce cross-turn reasoning the way four calls with code-enforced turn anchors can.

---

Markdown version of https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery/213. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt