# Arc research audit

Status: **share with noted caveats, version 2.** Both findings from the second audit are fixed. Confidence is high in the corrected calculations for the saved snapshot; full-chain launchpad coverage remains partial. Six earlier strict value comparisons remain open.

## Version 2 corrections and checks

- Fixed source-only post-move turnover in **1,035 of 9,057 comparisons**, including **87 sign changes**. Unknown and unattributed executions in the same route no longer inflate a named source's value. The cohort, visit order, time bounds and path counts are unchanged.
- Corrected buyer/seller breadth labels: address counts use all captured quote assets, while value fields use USDC quotes. The all-quote buyer count is **342,483 address/base-token pairs**; the separate USDC-only check is **329,608**. These are overlapping token/address pairs, not distinct people. [Field definitions](/investigations/arc-blockchain-launch-trading/data/analysis/buyer-seller-breadth-scope.csv).
- **19 targeted checks passed.** The rebuilt values for every one of the 9,057 comparisons match the independent audit. Tests cover mixed routes, several named sources, zero values, transaction order, exact time bounds and incomplete follow-up. [Correction checks](/investigations/arc-blockchain-launch-trading/data/analysis/v2-correction-audit.csv).
- The **160.12 million USDC** headline, named-source rankings and all **25,919 first-hour token cohorts** are unchanged. No fresh chain or API data was needed. The public package uses corrected version 2.

## Checks run

- **103 of 109 final checks passed.** Three failures concern reference-USD values; three concern USDC quantities, each below 0.002 USDC. All are final-hour aggregate comparisons. They do not alter saved event counts. The USDC headline uses saved-row sums and is stated in millions to two decimal places.
- Earlier snapshot sealing passed **35 checks**, including retained-overlap count tests after the live data floor moved. These are a separate check suite, not 35 extra independent sources.
- All **1,543,945 swaps** join successful transactions, with matching second-level time and transaction position. Trader equals the raw chain sender throughout.
- **68 saved RPC receipt senders** agree with the raw transactions. Three boundary blocks agree on hash, timestamp and transaction count. API and raw ClickHouse records may share upstream ingestion; RPC samples are a separate chain-record check.
- Hour, protocol, quote, base-token, pool and exclusive launch-origin counts and USDC sums reconcile. Cohorts include zero-observation tokens; funnels are nested and path outcomes sum to their eligible sets.
- Common-window gas from transaction fee/price equals the saved block gas sum: **732,723,109,103**. Fee and transaction category totals reconcile.

## Earlier corrections

1. Preserved the two equal-field trading repeats. Each has separate matching swap events; deleting them would undercount activity.
2. Replaced the short launch-log index with the header index in occurrence keys. Added call context to transfer keys. Earlier transfer files retain their original short-key diagnostic; that field is superseded by `transfer-key-check.json` and must not be used as a unique count.
3. Excluded ordinary top-level calls from creation-trace counts. Resulting contract addresses remain missing, so no complete new-contract count is claimed.
4. Kept the cube header’s sender field separate from the verified chain sender. Trader identity is consistent in every analysis.
5. Held out NebulaPad after a token/mint binding sample failed. Kept newer Archemist V4 candidates outside the V2 token-set comparison.
6. Preserved null base quantities on 226 rows. Valid quote amounts remain; null quantities do not become zero.
7. Separated 2,709 native/ERC-20 USDC-view exchanges, worth 213,615.85 USDC, from the 160,122,201.84-USDC token-market headline. Supporting full-cube tables clearly retain the wider set.

## Open differences and limits

The repeated final-hour reference-USD sum is about **9.18 USD below** the saved sum. Counts agree exactly. The three differences are approximately 0.25 USD for V2, 7.74 USD for V3 and 1.19 USD for V4. Their cause is not established. These checks remain failed; the strict tolerance was not changed after seeing the result. Reference-USD estimates are not used as the headline or platform comparison unit.

The fresh quote-quantity check also finds three small differences: saved minus fresh is **+0.000449797 USDC** for V4 native, **+0.000721176 USDC** for V2 ERC-20, and **−0.001923298 USDC** for V3 ERC-20. All corresponding counts agree. Floating-point aggregation or conversion is a possible explanation, but the cause has not been proved. Saved-row sums are the fixed research values. These differences matter to a strict equality claim, so they remain failed checks even though each is below 0.002 USDC and cannot change a million-USDC headline at the displayed precision.

The one-minute 13:00–13:01 raw/cube check has 3,405 successful raw DEX rows and 3,395 cube rows. Missing cube rows are not filled from raw events. The sample cannot prove a chain-wide decode rate.

The live Trading floor advanced by six minutes and 13 seconds during capture. Earlier rows had been saved, but cannot be re-queried from the later live floor. Their count matches the initial aggregate and the later retained overlap. A fixed retention duration is not inferred.

Raw blocks before the common window contain 227 absent heights. Common-window block continuity and transaction totals pass. This does not certify all protocol or transfer coverage.

The launch registry is partial. Sampled event/token mappings do not prove every token birth or all contract versions. Unknown factory identity, missing creator fields, unverified curve trades and older tokens remain material limits. The 61-measure appendix reports conditional metrics that could not be supported.

## Reproduction

The public package includes the result CSVs and saved analysis SQL. Raw authenticated responses remain private. Checking public arithmetic differs from repeating the original capture. Amount sums are floating-point conversions, published to sensible precision. Comparable counts use exact integers; the final numerical test tolerance is relative 2e-11 or absolute 0.00001 quote units.

See [final check results](/investigations/arc-blockchain-launch-trading/data/analysis/final-audit.csv), [snapshot sealing](/investigations/arc-blockchain-launch-trading/data/snapshot-seal-checks.csv), [block/source audit](/investigations/arc-blockchain-launch-trading/data/source-audit.csv), [USDC quantity comparison](/investigations/arc-blockchain-launch-trading/data/analysis/final-hour-quote-comparison.csv), [metric status](/investigations/arc-blockchain-launch-trading/metric-status.md) and [methods](/investigations/arc-blockchain-launch-trading/methodology.md).