campaign-management

Campaign Management

Watch active outreach campaigns and make real-time adjustment decisions against pre-defined thresholds. Auto-pauses on hard-threshold breach (Google/MS complaint rate, Postmaster "Bad" reputation). Reply rate is the primary metric (Apple MPP made opens noise).

When to Use

• Active campaign(s) need ongoing human-in-loop oversight
• A/B test running — when's it ready to call?
• Complaint rate looks high — should I pause?
• Daily/weekly review of all active campaigns
• Reply rate plateaued at 2.5% on day 8 — diagnose
• Reputation drift detection and pause decisions

Quick Reference

Concept	Value
Primary metric	Reply rate (cumulative + per-channel). Open rate = noise
Reply-rate floor (cold email)	3% by D+10
Reply-rate floor (multi-channel)	6% by D+10
LI acceptance floor	12% by D+7
Call connect floor	6% cumulative
Bounce thresholds	Pause 5%; warn 2%
Complaint thresholds	Pause 0.3%; warn 0.1% (Google/MS Feb 2024)
Postmaster Tools	Pause if "Bad"; warn on Medium-trending-down
SNDS	Pause on red; warn on yellow
State machine	draft → active → paused → completed / aborted
Decisions	continue / slow-down / swap-copy / pause / resume / abort
A/B method	Bayesian (n<200/arm); frequentist z-test (n≥200/arm)
Polling default	Daily cumulative + 4h sender-reputation-only for active
Reputation cross-check	Multi-source convergent signal triggers pause; single-source = warn

Procedure

1. Pull current metrics per campaign — 24h + cumulative from sending platforms; per-channel disaggregation; sender reputation (Postmaster + SNDS + Talos) + 7d trend. Reference ${HERMES_SKILL_DIR}/references/thresholds-and-decisions.md.
2. Compare against thresholds + make decision — Per conventions §8.4 + .env defaults. Pause if any pause threshold hit; slow-down if warn only; swap-copy if reply rate below floor AND deliverability healthy; continue if all nominal.
3. A/B test significance check — n<200/arm: Bayesian, recommend winner if P(b>a) > 0.85 AND effect ≥1pp. n≥200/arm: z-test at p<0.05 + effect ≥1pp. D+14 inconclusive: stop, ship higher-volume variant. See ${HERMES_SKILL_DIR}/references/ab-testing-methods.md.
4. Diagnose plateau (when swap-copy triggered) — Audit recent copy against cliché + buzzword blocklists; audit hook quality; audit recipient quality. Recommend: copy rewrite / re-enrichment / re-tier.
5. Write adjustments_log + notify user — Every decision (incl. "continue") logged with timestamp, triggers, metrics_snapshot, action_taken, reason. Pause = immediate high-priority; slow-down = next cycle; swap-copy = with diagnosis.
6. Push to CRM + schedule next cycle — Run ${HERMES_SKILL_DIR}/scripts/push_to_crm.py per decision. PATCH Campaign state + metrics_summary if changed. Default: daily cumulative + 4h reputation for active; daily for paused.

Pitfalls

• Optimizing on open rate — Apple MPP made it noise
• Ignoring complaint-rate spikes — Google/MS Feb 2024 hard rule; pause at 0.3%
• Slow-down without re-eval — becomes silent under-performance
• Calling A/B winners early — n<50 = noise
• Treating reputation feeds as gospel — cross-check; pause on convergent signal
• Not logging "continue" decisions — audit trail incomplete

Verification

1. Every active campaign has a decision logged with metrics_snapshot at the timestamp; all pause-threshold breaches resulted in actual state transitions
2. Sender reputation feeds were pulled from named tools (or [unverified] flagged); A/B test winners have explicit posterior/p-value AND effect size
3. Re-running same monitoring cycle on same metrics produces same decision (deterministic)