Verification record · Build with Gemini XPRIZE
Sloane & Pearl
Published evidence separates rules that pause ads, Claude strategy, and Gemini product copy. Sales and production operation remain unverified.
Claim
Sloane & Pearl describes “a real fashion brand where AI, not a human, pauses ad spend, rewrites strategy nightly, and writes product copy via Gemini.” Its submission reports $18,641.28 in sales from 204 customers.
What we tested
On 16 September 2026, sound.fan inspected the participant’s public evidence and code-excerpt repositories: the three meta-auto-kill files, nightly strategy code and logs, product-writing routing, a production usage row, the August Google Cloud cost export, and the scripts behind its business totals. The excerpt repository says these are unmodified extracts of a private multi-tenant platform. We did not have access to that production platform or its database.
Finding
Observed in participant-controlled evidence: three different mechanisms sit behind the headline. Rules pause ads; Claude writes strategy; Gemini writes new product copy. The evidence makes the decision chain visible, but does not establish that the excerpts match production.
The 1,367 lines of ad-pause code contain no model call. Default rules include a $1 cost-per-click ceiling checked at $10 spend, a pause at $22 with no add-to-cart, $37 with no purchase, and $60 with no purchase for collection campaigns. The exported log shows eight pauses from 29 July to 4 August and describes spend/conversion thresholds.
The nightly strategy code selects claude-sonnet-4-6. Code bounds the proposed weight changes; nothing changes if the model call fails. The participant’s log describes a 04:30 Europe/Amsterdam run over 30 days of Meta results, with the latest displayed brief dated 12 August.
Product-writing code defaults to Anthropic and routes this store to Vertex AI’s gemini-2.5-flash. A usage row is dated 12 August, 16:58 UTC; the August cost export lists €0.006251 for Vertex AI. These records are consistent with the stated 12–13 August switch and low Gemini volume. The repo dates the 2,529-product catalogue to 30 July and says its historical copy was generated via Anthropic.
What the evidence establishes
The submitted automation claims moved from Claimed to Observed mechanisms in published code and logs. Ad pausing is unattended threshold automation; strategy and copywriting use models. XPRIZE allowed additional model providers, and the team disclosed both. The submission identifies the Gemini connection as the hackathon-specific work and says the ad engine and nightly strategy predate it.
These are participant-controlled records. We did not independently reproduce private production operations or corroborate their frequency. The evidence does not tell an outsider which model wrote any particular live description.
What remains unverified
Not verified: $18,641.28 sales; 204 customers, described as distinct emails on 214 orders checked on 14 August; $27,971.34 costs; a $9,330.06 loss; product margin above 67%; and 95 auto-pauses among 127 campaigns. Published P&L scripts require private database access. The customer-service contractor, 24 repricing changes and cross-store separation also remain participant-reported.
What evidence would change that
A Shopify order export for 9 June–16 August, with hashed emails, dates and totals, could reconcile the reported sales and customer count against the store’s records. Stronger assurance would tie those orders to an independent payment or financial record. A Meta Ads Manager export and Vertex AI request logs for 12–17 August would address the separate pause and model-usage counts.