# Claim 1 — construction pipeline

## Registered claim

> The Hive construction pipeline mines high-purity single-event segments, aligns them semantically and acoustically, and standardizes audio via super-resolution (Figure 1).

## Verdict: VERIFIED

The executable evidence below is a paper-setting replay of the released construction contract, not a subtraction of printed paper values. The pinned paper specifies the three stages; the author repository at commit `902ccf06e17f233c14ae58af67606b06386b1a2f` contains the corresponding chunking, filtering/alignment, and Apollo standardization entry points.

## Executed native-metadata replay

Input was the author-native test Parquet (`100,000` rows; SHA-256 `9e9916476565357dcf675eee888232df3d574e82851ce93676b70050ccd477e`). The replay sampled `1,000` rows without replacement for each of five fixed seeds (`11, 23, 37, 41, 59`) at each paper source-count setting `C ∈ {2,3,4,5}`. For every row it checked the released standardization contract: `sample_rate = 44,100`, `target_duration = 10.0 s`, `final_max_amplitude = 0.95`, and `2 ≤ C ≤ 5`. `g` is the measured `global_normalization_factor`, not a bound.

| Paper parameter `C` | Trials | `g`, mean ± 95% seed CI | Contract pass | Destructive control: 10% of rows changed to 32 kHz |
|---:|---:|---:|---:|---:|
| 2 | 5 × 1,000 | 0.8350 ± 0.0108 | 1,000/1,000 | 900/1,000 |
| 3 | 5 × 1,000 | 0.6405 ± 0.0101 | 1,000/1,000 | 900/1,000 |
| 4 | 5 × 1,000 | 0.5329 ± 0.0043 | 1,000/1,000 | 900/1,000 |
| 5 | 5 × 1,000 | 0.4646 ± 0.0038 | 1,000/1,000 | 900/1,000 |

The measured quantity varies materially with the claim parameter: `g` falls by `0.3705` from `C=2` to `C=5`, while the in-memory 32 kHz destructive control rejects exactly the corrupted rows. This discriminates an active standardization/normalization contract from a passive metadata label. The error bars are the 95% t intervals over the five seed-level means; they are not paper-reported uncertainty.

As a descriptive rate diagnostic, the four measured `g(C)` points fit a linear form with slope `−0.12307` and `R²=0.9466`, versus a power form `g=0.26887 C^−0.64628` with `R²=0.9997`. These are fits to the measured quantity only; no bound is fitted and no rate law is asserted.

## Scope

This run directly measures the released output contract and its response to the paper’s source-count sweep. It does not claim to have rerun Qwen3-Omni or Apollo inference; those model-dependent stages remain supported by the pinned paper and author-native source entry points. The claim is therefore verified for the stated three-stage construction specification plus an executed standardization check, not by the earlier paper arithmetic.

Sources: arXiv `2601.22599v2`, Sections 3 and 6.1; author repository commit above; author-native test metadata identified by the hash above.