/rnd-loop — research → prototype → decide

Why this exists

An R&D Engineer's job is a loop, not two errands: Scan → frame a hypothesis → spike → evaluate → decide → capture. /storm-research owns the front (Scan). /spark owns the back (spike → decide). /rnd-loop is the thin conductor that runs the whole arc on one topic and hands Gera a single R&D dossier: the verified briefing and the evidence-backed verdict, linked. Use it when the goal is not "tell me about X" or "test X" but "I have a question and I want to end with a decision I can defend."

It is glue, not a new engine — it sequences the two proven engines with a human gate where it matters (a real R&D Engineer chooses what to test). Wiki: references/rnd-instrument.md.

Gate 0 — SoftServe refusal

Inherits BOTH halves' boundary, at the stricter end (the loop ends in running code). Refuse if the topic or any spike would touch SoftServe customer IP/code/strategy or Jumpstart specifics — run it in the ss instance / a SoftServe-approved tool. Public / personal / Penguin-Alley topics proceed.

The loop

1 — Research (Scan). Run /storm-research on the topic in full (5 lenses → contradiction map → synthesis → citation-verify → Codex verdict → HTML briefing). Save the briefing JSON; you will feed it to the spike framer. Show Gera the 60-second summary.

2 — Frame + pick (HITL). Run the spark engine's frame stage on the briefing:

Workflow({ scriptPath: ".claude/skills/spark/spark.workflow.js",
           args: { stage: "frame", idea: topic, briefing, reader, goal } })

Present the ranked experiments. Gera picks the one hypothesis worth a spike (or --auto → #1, state the assumption). This is the hinge of the loop — research becomes a testable bet. Stop and let him choose unless told to run unattended.

3 — Spike → decide. Hand off to /spark from its Phase 2: build the throwaway spike → assess (route by artifact_type to /verify-loop / /eval / /prompt-eval / a direct benchmark) → run the engine's verdict stage → Codex final verdict → render the HTML feasibility verdict (Front-End-Review-Law'd).

4 — Dossier + capture. Emit a short index that links the two artifacts as one R&D dossier:

• the STORM briefing → artifacts/research/<date>-<slug>.html
• the SPARK verdict → artifacts/research/<date>-<slug>-verdict.html

Both render steps also emit the source JSON into the combined dossier at artifacts/research/<slug>.dossier.json (via scripts/arc-dossier.py, upserting one half at a time), which is what The Arc dashboard (rnd-arc/) reads to auto-populate over time.

Then capture (Phase 5 of /spark): file durable learnings to the wiki, log /rnd XP for a completed loop, and name the next move (a GO points at a /build; a PIVOT points at the next experiment; a NO-GO is a documented week saved).

Notes

• The human gate is non-negotiable by default. The loop pauses between research and spike so Gera picks the bet. --auto exists for unattended runs but should state every assumption it made.
• One loop, one hypothesis, one decision. If the research surfaces several bets worth testing, run the loop once per bet (or queue them) — do not braid three spikes into one muddy verdict.
• Cost shape: STORM (~10–14 agents) + spark frame (~4) + spark verdict (~4) + the spike build + 2 Codex passes. Far cheaper than it looks, and you can say so. For a quick gut-check use the halves directly; reach for /rnd-loop when the topic deserves the full arc.
• Distinct from its parts: /storm-research alone = knowledge, no test. /spark alone = test an idea you already hold, no research front. /rnd-loop = the whole instrument, research-grounded and evidence-closed.
• Match references/voice.md through the phases.