Risks and Open Questions

Summary

Failure modes to watch, plus the questions still genuinely open. Resolved questions have been folded into the permanent docs (pointers below). Capture new answers here as you learn them and log them in the decision log.

Resolved — now documented in context

These were open questions; the decisions are made and live in the docs:

Failure modes to watch

Goodhart’s law on proxy metrics

Risk: “when a measure becomes a target, it ceases to be a good measure.” Optimize saves hard enough and you get save-bait that doesn’t sell. Mitigation: periodically re-validate proxy ↔ outcome correlation; keep a downstream guardrail metric (actual conversion). Reinforced by the scoreboard vs control panel split — always keep watching the scoreboard while the agent steers the control panel.

Over-automation / brittleness

Risk: an LLM agent doing a deterministic job fails in new, weird ways and costs more. Mitigation: the script test in Don’t over-classify.

Brand / compliance risk (peptides especially)

Risk: peptides are a regulated, claims-sensitive space; an agent making health claims or tripping platform rules is a real liability. Mitigation: banned-words + claims rules as hard policy; human approval gate on all public content (this is a per-item safety gate, distinct from the weekly performance QC); never let medical/health claims publish autonomously.

Context staleness

Risk: knowledge base / intel goes stale; agents act on old facts. Mitigation: date everything; ingest agent refreshes; separate evergreen from transient. See Knowledge base as RAG.

Single source of truth drift

Risk: inventory, price, tracking, or metrics state diverges across systems. Currently live: web tracking (Stape+GA4) and social (Zernio) are not yet consolidated. Mitigation: one authoritative store per data type; everything else syncs from it. Build the metrics-consolidation layer (doubleloop-style) so the analyze agent reads one source. See Current instrumentation stack and The bridge into Distribution.

Still open

  • Per-pipeline safety-gate UX — what the banned-claim approval gate physically looks like for the peptide network (Obsidian checkbox vs Telegram/Slack button). Tracked in to test.
  • Proxy validation test — empirical check that saves predict click-through, before scaling.