Skip to content
axp.one
First recording, hosted sandbox, Oct 1
First recording, hosted sandbox, Oct 1Recorded replay of a run on a hosted Solana sandbox, test USDC
Replay of a live Solana Devnet run, Oct 1The main run: fully live on Solana Devnet
Separate runs; their numbers are never merged.
Test USDCFictional advertisers
PresentVerify

Evidence

ContextHint

ContextHint is the intelligence platform for ChatGPT ads. Agents judged each moment with history from ContextHintContextHintContextHint is the intelligence platform for ChatGPT ads (contexthint.com). It records real prompts and the ads seen beside them; agents get past examples and an inferred audience from it as evidence. (contexthint.com): prompts and the ads that appeared beside them, from ContextHint's stored ad library and ContextHint's collection, plus inferred targeting hints. Brands below are observed historical references, not axp.one advertisers.

Prompt and ad pairs1,178In the screened slice
Distinct prompts241Each ranked once
Ad creatives537Observed, not enrolled
Inferred hints331Hypotheses, not settings
Used in this run13 + 5Records and hints shown to agents

Topics in the slice

Hardware wallets and self-custodyCrypto investingWeb3 infrastructureCrypto tax and DeFi accountingPrivacy blockchains and compliance

Hint tiers

Held-out tested246Leave-one-out tested42Sparse (few supporting ads)43

Sources

ContextHint's stored ad library1,031ContextHint's collection147

How agents got their evidence

Retrieval method, per question
Vector match (cosine similarity) for stored questions
When the exact question already had a stored embedding (BGE, 768 dimensions), retrieval ranked prompts by cosine similarity. No embeddings were generated during the run, and vectors never leave the backend.
Lexical fallback for new wording
When no stored vector existed, retrieval ranked prompts by shared words. That is an overlap score, not semantic similarity, and it is labelled that way wherever it appears.

Records shown to agents

The observed examples placed in agent packets
Observed by ContextHintMatch1.000

“cheapest hardware wallet that still supports ethereum and solana”

Mercari Nano Trezor Hardware Wallet NEW - New Electronics | Color: White New in wrapping Open to offers Paid $55

Observed historical reference, not an axp.one advertiser
Shown to ClearVault, KeyForge, in opportunities 1, 2, 3
Observed by ContextHintMatch0.742

“best wallet for privacy preserving blockchains”

Ariat Diamond Plate Bifold Wallet Slim bifold wallet with money clip & built-in bottle opener.

Observed historical reference, not an axp.one advertiser
Shown to LeatherGuard, in opportunities 1, 2, 3
Observed by ContextHintMatch0.682

“Does Streamflow work with multisig wallets?”

Yoder Leather Company RFID Blocking Wallet Card RFID blocking card protects credit cards from scanners.

Observed historical reference, not an axp.one advertiser
Shown to LeatherGuard, in opportunities 1, 3
Observed by ContextHintMatch0.649

“Fomo vs Phantom wallet”

Thursday Boot Company Bifold Wallet Slim bifold wallet in Horween Natural Chromexcel leather

Observed historical reference, not an axp.one advertiser
Shown to LeatherGuard, in opportunities 1, 2, 3
Observed by ContextHintOverlap0.204

“What are the risks associated with self-custody Bitcoin wallets?”

OneKey Limited OneKey: Best Hardware Wallet in Asia The hardware wallet trusted by crypto users worldwide. Summer Sale Live. Buy wallet, get BTC back.

Observed historical reference, not an axp.one advertiser
Shown to ClearVault, KeyForge, in opportunity 2
Observed by ContextHintOverlap0.183

“How can a neobank offer self-custody products with an institutional safety net”

Yoder Leather Company RFID Blocking Wallet Card RFID blocking card protects credit cards from scanners.

Observed historical reference, not an axp.one advertiser
Shown to LeatherGuard, in opportunity 2

Inferred hints shown to agents

Inferred hintSparse
People comparing crypto wallets like Fomo and Phantom who are shopping for a slim, premium everyday carry bifold wallet.
Inferred hintSparse
People shopping for a slim, durable men's bifold wallet for everyday carry.
Inferred hintSparse
People weighing the risks of self-custody Bitcoin wallets and comparing hardware wallet options to secure their crypto holdings, particularly retail users in Asia and global holders exploring cold storage solutions.
Inferred hintHeld-out tested
Crypto investors and technically curious blockchain users researching wallet security, self-custody, privacy, and transaction methods. Handmade leather portfolios, money clip wallets, and RFID protection cards may appeal to people who value secure, organized control of financial assets.
Inferred hintLeave-one-out tested
Crypto holders and stakers looking to buy or resell hardware wallets like Trezor and Ledger on Mercari, where listings for new and used self-custody devices surface.

Other nearby prompts

Neighbours listed in packets but not shown as examples
“How do I check multiple wallets for airdrops at once?”WOLF 1834 Observed historical reference, not an axp.one advertiser
“cheapest hardware wallet under 100 dollars that supports bitcoin and ethereum”Mercari Observed historical reference, not an axp.one advertiser
“cheapest cold storage wallet that still keeps my bitcoin safe”Mercari Observed historical reference, not an axp.one advertiser
“cheapest hardware crypto wallet under 100 dollars that's actually legit”Mercari Observed historical reference, not an axp.one advertiser
“Fomo vs Phantom wallet”Ariat Observed historical reference, not an axp.one advertiser
“Best way to reuse identity verification across different wallets”Yoder Leather Company Observed historical reference, not an axp.one advertiser
“best hardware wallet for solo staking ethereum right now”Mercari Observed historical reference, not an axp.one advertiser

What this evidence is not

From the dataset manifest
  • Within the measured prompt panel; not all AI demand or real-world query volume.
  • Distinct source mappings remain raw associations, not impressions, conversions or independent fit labels. Retrieval ranks each normalized prompt once.
  • Historical brands are references, not enrolled bidders or declarations of fictional campaign capabilities.
  • Hints are inferred historical hypotheses, not advertiser configuration; sparse tier is retained.
  • No raw probes, answers, HTML, URLs, images, geo, customers, credentials or import-spread timestamps exported.
  • Model revision is unrecorded. Only existing matching BGE 768-D prompt vectors; no embedding generation or provider calls.
  • Vectors are backend-only. publicCatalogue and retrieval/model packets never return vectors.
  • Exact cached text uses vector cosine; authored uncached text uses labelled lexical fallback. No training, targeting accuracy or causal-lift claim.