// Blog
Why Provenance Matters
While building a demo comparing Nansen and Testril pricing, we ran into a problem. We asked both systems the same question, but the data we got back didn’t match. At that point we stopped worrying about pricing.
The question we were asking was simple: during two separate hours, who sent and received USDC from a Uniswap V3 pool.
Since the answers were different, we needed to figure out who was right. To do that we wrote a program completely independent of Testril that went directly to the Ethereum blockchain, collected the raw block data, and replicated the work required to produce the answer.
This post includes a link to code and the evidence, so you can check our work yourself.
The Question
“Who did this address send USDC to and receive USDC from during this hour, and how much moved?”
We checked one Ethereum address, the Uniswap V3 USDC/WETH trading pool:
Address: 0x88e6a0c2ddd26feeb64f039a2c41296fcb3f5640
USDC: 0xa0b86991c6218b36c1d19d4a2e9eb0ce3606eb48
Date: 10 November 2026
Hours: 12:00–13:00 and 13:00–14:00 UTC
It’s the pool contract, not a wallet we control. Each hour includes its start and excludes its end.
The endpoints needed to answer this question aren’t identical – they rarely are! – but they each returned the raw data we needed. We used Nansen’s counterparty summary and Testril’s wallet transfers, grouped by the other address.
The Discrepancy
The Nansen Data
The requests to Nansen’s Address Counterparties service used this form, changing only the hour:
POST https://api.nansen.ai/api/v1/profiler/address/counterparties
Content-Type: application/json
{
"address": "0x88e6a0c2ddd26feeb64f039a2c41296fcb3f5640",
"chain": "ethereum",
"date": {
"from": "2026-11-10T12:00:00Z",
"to": "2026-11-10T13:00:00Z"
},
"source_input": "Tokens",
"group_by": "wallet",
"pagination": { "page": 1, "per_page": 1000 }
}
The second request used 13:00 and 14:00. The documentation accepts full timestamps; its own example is 2025-01-01T00:00:00Z. Both complete responses contained the same 910 counterparty entries, including the same counts and amounts; both reported that no further page remained. Each counterparty entry can include several tokens. USDC appeared in 787 entries, and adding their USDC num_transfer values gave 7,139 transfers.
| USDC transfer count | Nansen counterparty summary | Testril |
|---|---|---|
| 12:00–13:00 | 7,139 | 292 |
| 13:00–14:00 | 7,139 | 503 |
We also compared incoming and outgoing amounts. Neither Nansen summary matches either Testril hour, even if the hours are swapped.
The Testril Data
For Testril, we retrieved individual USDC transfers and calculated the counterparty summary ourselves.
We used its production MCP server at https://api.testril.si/mcp. First, we called bind, selecting Ethereum, the pool address and USDC:
{
"chain_id": 1,
"function_slug": "erc20.wallet_transfers",
"params": {
"wallet": "0x88e6a0c2ddd26feeb64f039a2c41296fcb3f5640",
"tokens": ["0xa0b86991c6218b36c1d19d4a2e9eb0ce3606eb48"]
}
}
Testril returned an identifier for that function that allowed us to execute it over any time window. Then, to retrieve the prepared data, we called read:
{
"bound_function_id": "sha256:76b099073c3bc3db28e5f8d1006a901ad4c64b3543bbaaab0b237621fe87dac1",
"from_time": "2026-11-10T12:00:00Z",
"to_time": "2026-11-10T13:00:00Z"
}
For the second hour, we changed only the times to 13:00 and 14:00.
Each returned row identifies one transfer: its transaction, block, position within the block, token, sender, recipient and exact amount. Amounts are integers in millionths of a USDC; we retained those integers throughout to avoid rounding.
Testril returned 292 transfers in the first hour and 503 in the second. We grouped each hour’s transfers by counterparty, counted the events and summed incoming and outgoing amounts. Multiple transfers within one transaction remained separate events.
These summaries were our calculation from Testril’s transfer records. Later, provenance let us reconstruct the transfers directly from Ethereum and repeat that calculation independently.
The Investigation
The Nansen output and the Testril output didn’t match, so we investigated.
Checking Nansen
We were able to investigate Nansen because we knew it had a separate transaction-history service. We knew how to retrieve every page, select USDC, separate the transfer legs within transactions, preserve repeated transfers, and reconstruct the comparison.
That separate listing agreed with Testril: 292 transfers in the first hour, 503 in the second, with matching amounts, senders and recipients.
We then tested whether the repeated summary described the whole day. We requested the full day from Nansen and independently counted the pool’s USDC transfers for 10 November. Both produced 7,139 transfers involving 787 counterparties. Every counterparty’s transfer count and incoming/outgoing amounts matched the figures returned for both hourly requests.
We had initially treated the accepted timestamps as hourly filters. That interpretation was wrong for these responses: their statistics described the complete UTC day. The response did not identify that effective window. We discovered it through additional investigation, rather than through source information accompanying the answer.
Checking Testril
Testril makes that source information part of its data contract. Together with the original request, provenance identifies the computation and its sources. Where a result depends on another calculation, the source trail can lead to those inputs. A reader follows those references instead of guessing which other service might explain an answer.
For each hour, we retrieved Testril’s separate provenance report for the same bound function and window. Selected fields for the first hour were:
{
"bound_function_id": "sha256:76b099073c3bc3db28e5f8d1006a901ad4c64b3543bbaaab0b237621fe87dac1",
"chain_id": 1,
"from_block": 26384371,
"to_block": 26384670,
"block_count": 172,
"computation": {
"id": "erc20.wallet_transfers",
"kind": "function",
"version": 1
},
"digest": "sha256:337ea630fb5e9f6455a1be907fc6e9af299189292bde4bec240942dd3817b8fd"
}
block_count counts only the blocks with a matching USDC transfer, not all ~300 blocks in the hour. The full report also lists every contributing block’s number and hash. The second hour cites 200 contributing blocks and digest:
sha256:32cc71c8a01d278d173a1ca7ca021a13d4018757608f83f65fef8c22ba0c07d9
Succinctly:
- The original request identifies the address and token we asked about.
- The bound-function identity connects the saved result and provenance to the query results.
- The computation name and version tell us which definition to examine. They are not executable source code by themselves.
- The block range fixes which part of Ethereum history is relevant.
- Each block hash identifies the cited block.
- The digest fingerprints the complete contributing block list.
The digest is SHA-256 over these lines in ascending block order, with lowercase hashes and a newline after each line:
<chain_id>:<block_number>:<0x_block_hash>\n
A matching fingerprint checks the source list. Checking the numbers also requires retrieving the events and repeating the calculation.
The provenance gave us a specific calculation and a defined set of Ethereum source blocks to check. To verify the numbers, we needed to retrieve those records independently and repeat the calculation. That is what we did next.
Checking Independently
We wrote a small JavaScript program to follow that source trail. Using the original request and provenance, it retrieved the relevant records from an external Ethereum node, without asking Testril or Nansen for the answer. It checked the cited block hashes and time boundaries, then decoded USDC transfers involving our address.
We searched every block in both hours, including blocks Testril hadn’t cited. Checking only the cited blocks could confirm the listed transfers while missing an omitted one.
The program compared each reconstructed transfer with Testril’s result: transaction, block, event position, sender, recipient, token and exact amount. It also rebuilt the source fingerprint and recalculated the incoming, outgoing and counterparty totals.
All 795 transfers matched exactly, and the independently calculated totals agreed for both hours. This established agreement with the external Ethereum node. Readers can repeat the check against their own node.
To reproduce this check, clone the companion repository. Its README explains the input data, verification program and output, with instructions for replaying our evidence or checking against an Ethereum node yourself.
What We Proved
We independently reproduced the returned transfers and the counterparty summaries we calculated from them for this address, token and two historical hours. The conclusion now rests on reconstructing the source records, not just on two vendors agreeing or on a handful of sampled transactions.
We trust the external Ethereum service to return correct canonical headers and complete event records. The script does not validate Ethereum consensus or construct and verify transaction-receipt proofs against an independently trusted chain head.
The provenance was retrieved separately after the original reads. These files are not a vendor-signed attestation authenticating the original delivery. We verified their contents against Ethereum.
This verifies one named computation on two windows. Other computations need their own appropriate checking logic, following the provenance and any dependencies.
The repeated Nansen summaries describe the complete UTC day, while its separate transaction-history service agrees with the verified hourly records. We have identified the period represented by the summaries, but not the internal mechanism that produced that behavior.
Finally
Mistakes and misunderstandings will happen. When they do, the answer should come with enough information for someone else to investigate it.
We could investigate Nansen because we already knew how to connect its separate services. With Testril, the source trail gave us a defined path back to the records behind the answer. We followed that path, repeated the calculation, and confirmed the numbers. The files and code are available so that you can repeat it too.
Provenance makes “show your work” something a reader can actually use.