SocialSeed.ing By Agent · Experiment Console · pre-registration #SS-2231
A Randomized Hook-Arm Trial of Autonomous TikTok Shop Seeding for LumaSkin Barrier Serum
Causal estimate of cost-per-verified-view under agent-managed creator allocation
Run agent: seed-orchestrator/ADK · sub-agents: strategist · sourcer · brief-writer · publisher · measurer · allocator · billing: $0.01 / verified view · cap $4,000.00 · data: synthetic (metrics-simulator v0.4, seed 8831)
Abstract. We pre-register a four-arm randomized trial (n = 48 micro-creators, balanced assignment) testing whether hook framing causally lowers cost-per-verified-view (CPV) on TikTok Shop seeding. The agent sources creators, generates #ad-compliant briefs per arm, schedules publishing, and applies a four-stage fraud-verification protocol before any view is billed. A sequential stopping rule (posterior P[arm = best] > 0.95 or budget horizon) governs reallocation. Interim result @ 62% budget: arm B (problem-first) CPV = $0.0083 [90% CrI 0.0071–0.0098], dominating arm A (demo-first, $0.0142) which was stopped for futility; pooled CPV $0.0094 vs target $0.0100. GMV attributed: $11,240 (412 orders, 7-day window).
1 · Experiment design (pre-registered)
| hypothesis | H₁: hook framing affects CPV (vs H₀: no effect) |
| primary outcome | verified-view CPV (USD) |
| secondary | CTR→Shop, GMV/creator, watch-through |
| design | 4 arms × 12 creators, block-randomized on follower decile |
| power | 0.80 to detect Δ$0.003 CPV, α=0.05 |
| stopping rule | P[best]>0.95 ∨ futility P[best]<0.05 ∨ Σ verified×$0.01 ≥ $4,000 |
| reallocation | weekly Thompson sampling over surviving arms |
Arms
| Arm | Hook framing | Brief template |
|---|
| A | demo-first (texture/swatch) | T-DEMO-01 +#ad |
| B | problem-first (skin-barrier pain) | T-PROB-02 +#ad |
| C | routine-integration (GRWM) | T-RWM-01 +#ad |
| D | social-proof (dupe / comparison) | T-PROOF-03 +#ad |
Disclosure invariant: every generated brief carries mandatory #ad + branded-content toggle; briefs are not paid-ads/legal advice. Synthetic data only — no live TikTok API.
2 · Live trial board — Figure 1
B · problem-first RUNNING · P[best]=.91
$0.0083
[.0071–.0098]
C · GRWM routine RUNNING
$0.0101
[.0082–.0125]
D · social-proof RUNNING · wide
$0.0107
[.0079–.0151]
A · demo-first STOPPED · futility
$0.0142
[.0118–.0179]
Figure 1. Forest plot of per-arm CPV with 90% credible intervals; dotted gridlines = $0.002 increments; red line = $0.0100 billing target. Arm A crossed the futility boundary on 2026-05-08 (week 2); allocator shifted its remaining $612 budget to arms B/C via Thompson draw. Lower is better.
3 · Methodology — fraud-verification protocol
- Deduplication. Views collapsed to unique (device-hash, post-id, ±30 min) tuples; repeat plays within session not billable. ✓ validated on simulator inject (12% dup rate → 0% post-filter)
- Watch-time threshold. A view counts only if dwell ≥ 2.0 s OR ≥ 50% of clip, whichever is shorter. ✓ ROC tuned, FPR 1.8%
- Source-anomaly screen. Per-post view-velocity z-scored against creator's 90-day baseline; |z| > 4 quarantines the burst pending manual review. ✓ caught 3/3 seeded bot spikes
- Attribution. Shop clicks matched to seeded posts via signed deep-link param; GMV credited on 7-day last-touch window. ✓ reconciliation drift < 0.5%
Billable verified views this run: 389,142 of 447,008 raw (87.1% pass rate). Spend to date: $3,891.42 / $4,000.00 cap. Agent will halt sourcing at the cap; no overage possible by construction.
4 · Reproducible run log (append-only) — Figure 2
2026-05-01T09:02Z strategist pre-registration #SS-2231 frozen · hash 4f1c9a · arms=4 cap=$4000
2026-05-01T09:14Z sourcer 48 creators sampled · block-randomized on follower-decile · A2A→brief-writer
2026-05-01T11:40Z brief-writer 48 briefs generated · #ad clause present=48/48 · banned-claims scan=clean
2026-05-02T07:00Z publisher wave-1 scheduled (mock TikTok) · 24 posts · branded-content toggle=on
2026-05-05T00:00Z measurer interim n=181k verified · CPV: A .0151 B .0089 C .0103 D .0112
2026-05-08T00:00Z allocator STOP arm A — P[best]=0.02 < 0.05 (futility) · realloc $612 → Thompson(B,C,D)
2026-05-11T00:00Z measurer n=389k verified · pooled CPV .0094 · GMV $11,240 / 412 orders
2026-05-12T08:30Z measurer arm B P[best]=0.91 — not yet at 0.95 stop; continue
Figure 2. Event stream emitted over A2A; each line = (timestamp, ADK sub-agent, action, decision statistic). Inputs are content-hashed so the run is replayable against the seeded simulator. No human edits — log is immutable.
5 · What we learned (interim)
Finding 1. Problem-first hooks causally reduce CPV by an estimated $0.0059 [90% CrI 0.0033–0.0091] versus demo-first (arm B vs stopped arm A) — economically large at scale.
Finding 2. Routine-integration (C) tracks the $0.0100 target but its interval still overlaps B; not yet separable. Social-proof (D) has the widest interval — under-powered, needs more creators before any claim.
Finding 3. Pooled delivered CPV ($0.0094) is at-or-below the $0.01 billing price, i.e. the agent is currently profitable for the brand on a verified-view basis. GMV/order signal is directional only (last-touch, synthetic).
Threats to validity. (i) Synthetic simulator may understate adversarial fraud vs. live TikTok. (ii) Per-arm n=12 creators limits power on secondary outcomes. (iii) Last-touch attribution over-credits seeded posts vs. organic discovery. (iv) Hook-framing effect may not transfer across product categories. Final report issues at budget horizon or P[best]>0.95.
Cite this run
@techreport{socialseed_SS2231_2026,
title = {A Randomized Hook-Arm Trial of Autonomous TikTok Shop
Seeding for LumaSkin Barrier Serum},
author = {{SocialSeed.ing Agent (ADK/A2A/Vertex Gemini)}},
number = {SS-2231}, institution = {SocialSeed.ing By Agent},
year = {2026}, month = may, note = {Interim @62% budget;
synthetic data, metrics-simulator v0.4 seed 8831},
url = {socialseed.ing/runs/SS-2231}
}