A blockchain query system reduced the amount of verification data, the evidence sent to a client to check an answer, by large margins in some tests. But the smaller evidence packages did not translate into lower wall-clock latency in the headline comparison. On an Ethereum gas stream, the gap in objects touched against scan baselines reached 130 as the query window grew from 100 to 2,000 records. Scan baselines were several times faster, though all three methods remained well under a millisecond.
VeriTS puts the query layer off chain and pairs it with an authenticated aggregate interval tree. The same structure serves as both the query index and the authenticated data structure used to check the answer and its aggregates.
VeriTS has an exact path and an approximate path. The exact path supplies a minimum covering set of records, including boundary records needed to establish completeness, while the tree hash binds child hashes and aggregates for soundness. The approximate path uses model segments to reconstruct values and timestamps, but each segment carries certified limits on its possible value and timing errors. Before a segment is anchored, miners replay the covered raw records to validate those limits.
Checks held up in the prototype tests
The formal security guarantee is conditional. It relies on collision-resistant hashing and on the correctness of the digests anchored to the chain. Under those conditions, an accepted query is verifiable, and an accepted approximate interval contains the exact answer held on chain without requiring the encoder itself to be trusted.
The empirical tamper tests matched that design: all 900 tested forgeries against the exact and approximate paths were rejected. The approximate checks also found the exact answer inside every reported interval across 1,000 boundary queries, 2,500 aggregate interval checks and 108,000 encoder-policy checks. Those results cover the tested cases, not every possible attack or workload.
Less data did not mean a faster end-to-end query
The benchmark also reported large client-side savings. For a 2,000-record window, B-V and B-Leaf sent about 18 times as much verification data as VeriTS, while B-MP sent 544 times as much. At the longest window, VeriTS also showed a 1,031-fold client-time advantage over B-Leaf.
For repeated queries, VeriTS produced a delta verification object that was 1.90 times smaller than the B-Ind ablation in a 40-window ETH-Gas sequence with a 10% slide. A separate scale test increased the TRON-Dust stream length 16.8-fold; verification-object size rose 17% and client time rose 26%.
The approximate route's retrieval advantage depended on stream type. Its range-query verification object was 4.05 times smaller than the exact object on TRON-Balance, a state stream. It was larger on the two event streams, ETH-Transfer and TRON-Transfer, and roughly similar on the rate stream, ETH-Gas.
The test covered five streams
The evaluation covered five streams drawn from 30,000 Ethereum blocks and 200,000 TRON blocks. The streams contained 29,123 ETH-Gas records, 29,123 ETH-Transfer records, 27,018 TRON-Transfer records, 27,018 TRON-Balance records and 168,499 TRON-Dust records. The benchmark used windows from 100 to 2,000 records, slides from 10% to 50%, and five repetitions over 20 to 40 random windows. Comparisons included two scan placements, a re-download method and three technique-specific ablations.
Encoder policy changed the width of approximate answers sharply. Under the E-Adv policy, intervals widened by as much as 25.6 times on TRON-Dust, 18.9 times on ETH-Gas and 10.5 times on TRON-Balance. The Learned policy reduced interval width to 0.23 of uniform width on ETH-Gas and 0.31 on TRON-Balance.
On four non-degenerate streams, the reported predicted and measured relative savings were 0.70 and 0.70, 0.71 and 0.71, 0.94 and 0.92, and 4.39 and 4.05. On TRON-Dust, the prediction was 24.0 against a measured 14.5.
The test was limited to five streams from Ethereum and TRON, with configured query windows, slides and workloads. The approximate path returns bounded intervals rather than exact point values, and its guarantee depends on certified model and arrival bounds, authenticated segment coverage, collision-resistant hashing and correct anchored digests. The reported latency comparison also shows why proof size and speed should not be treated as the same measure: scans touched more objects on ETH-Gas but still ran faster in wall-clock time.
Paper data and sources
Original title: VeriTS: Verifiable Model-Enhanced Time-Series Queries on Blockchain Systems
Authors: Zhongming Yao, Jun Pang, Chenxu Wang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text