NEAR DeFi wallet labels (1,000 rows)
Synthetic demo data: 1,000 labelled NEAR DeFi wallets (bots, LPs, CEX wallets) with confidence scores and last activity, Aug-Oct 2026.
- 1,000 rows (seller's claim)
- 9 columns
- inspections used
Reading the chain. This page usually appears within a few seconds.
An AI checks a sample: BUY or PASS.
Decide
Your browser unlocks the file and checks its fingerprint.
listing.blob_sha256 == eval.h == release.h == sha256(downloaded blob). A release mismatch refunds the purchase with reason hash_mismatch; an inspection of a swapped file stops and refunds as not_ok. The download check runs in your browser after the purchase has settled, so it warns and does not refund. The fingerprints above are those of the recorded mainnet run (listing #2, eval #2, purchase #1) as the chain holds them now. Shown on /verify and by npm run verify.
Refunds
A purchase comes back in full; an inspection comes back less Arrow's 10% share, to your Intents balance.
In the demo listings an inspection costs $0.25 and a purchase $2. If the sealed hardware fails, a purchase comes back in full and an inspection comes back less Arrow's 10% share of the fee. It goes to your NEAR Intents balance, a USDC balance held for your wallet on NEAR. The money moves back through intents.near's own mt_resolve_transfer.
The run drawn above is inspection #3 on mainnet, on listing #3 [synthetic] e2e broken host, whose file host is missing on purpose: the sealed hardware returned an error output (not_ok), $0.225 of the $0.25 in USDC went back to the buyer's NEAR Intents balance and Arrow's treasury kept $0.025. That run, checked from the chain →
inspect $0.25 · buy $2
inspect $0.25 · buy $2
Test runs, demo relists and any listing the team did not make; some were sealed to an earlier program and can no longer be inspected. All listings →
Synthetic demo data: 1,000 labelled NEAR DeFi wallets (bots, LPs, CEX wallets) with confidence scores and last activity, Aug-Oct 2026.
Check
Sealed hardware is a walled-off part of a computer chip, built so that even the machine's owner cannot read or change the program running inside, and so that the chip can sign a report of exactly which program ran.
A trusted execution environment (TEE, also called an enclave). Arrow's program runs in an Intel TDX TEE on OutLayer's workers; NEAR AI runs its model in Intel TDX with NVIDIA confidential computing. The signed report is a TDX quote: /verify checks it in your browser, nothing on chain checks it, and the guarantee rests on Intel's hardware (trust rows Evaluation and LLM). The 40-row sample and whole-file totals leave the TDX enclave for NEAR AI over TLS, not end-to-end encrypted; that NEAR AI keeps nothing is its own claim, and the model attestation on /verify shows the endpoint's current state, not its state at each inspection.
The LLM cannot produce BUY unless the structural gates pass; novelty is computed by code. The program can write only one of 51 fixed answers (BUY or PASS, fit 1-5, novelty 1-5), at most 5.67 bits on that output channel; a failed run adds two more outcomes on chain (an error output, or no output). Run time and work done are public and are not counted in that figure.
The file key is sealed to Arrow's program and handed to a buyer who pays, with two exceptions: OutLayer's keystore DAO can export keys, and a key released after its time window is refunded while the key, sealed to the buyer, is public. The served JavaScript handles the file while it encrypts, so for sensitive data use the CLI.
One of 51 possible answers, never a row. Measured as information (one bit is one yes/no answer): at most 5.67 bits per inspection, 283.65 bits over all 50. Run time and instruction count are public too and can hint at the data.
seller content → chain/operator: 0 B · out: 5.67 bits / budget 283.65 bits
a 40-row sample is read by NEAR AI's TEE model (see trust table)
Inspect $0.25 Buy $1
Synthetic end-to-end test data. The ciphertext host is missing on purpose, so every eval must refund.
One of 51 possible answers, never a row. Measured as information (one bit is one yes/no answer): at most 5.67 bits per inspection, 283.65 bits over all 50. Run time and instruction count are public too and can hint at the data.
seller content → chain/operator: 0 B · out: 5.67 bits / budget 283.65 bits
a 40-row sample is read by NEAR AI's TEE model (see trust table)
Inspect $0.25 Buy $2
A command-line tool; a tool server for AI assistants (MCP) with inspect_before_buy and buy; and for each listing a web address that always answers HTTP 402 (payment required) with how to inspect the listing and pay through NEAR Intents. It takes no payment itself and never returns the file.