# Benchmark Results

The release package includes two JSON outputs:

- `benchmark_results_v4_full.json`
- `benchmark_results_v4_quick.json`

The full local release run used deterministic generated corpora and compared:

- HVCE v4 fast
- HVCE v4 balanced
- HVCE v4 balanced + 10% recovery
- ZIP/Deflate level 9
- TAR+gzip level 9
- TAR+xz level 9
- TAR+zstd level 19 when the `zstd` CLI is installed

## Full run snapshot

```text
small_office raw=212,158 bytes
  TAR-xz9                         10,596  ratio=0.049944
  TAR-zstd19                      11,795  ratio=0.055595
  HVCE-v4-fast                    11,961  ratio=0.056378
  HVCE-v4-balanced                11,969  ratio=0.056416
  HVCE-v4-balanced+recovery10     20,722  ratio=0.097672
  TAR-gzip9                       23,529  ratio=0.110903
  ZIP-deflate9                   155,545  ratio=0.733156
```

HVCE does not beat xz on this particular tiny synthetic office corpus, but it is in the same range as zstd and massively beats ordinary ZIP/Deflate. The key v4 improvement is closing the earlier small-folder gap relative to ZIP-style archivers.

```text
photonic_generators raw=531,072 bytes
  HVCE-v4-balanced               132,243  ratio=0.249011
  TAR-xz9                        141,544  ratio=0.266525
  HVCE-v4-balanced+recovery10    146,799  ratio=0.276420
  TAR-zstd19                     235,182  ratio=0.442844
  ZIP-deflate9                   362,163  ratio=0.681947
  TAR-gzip9                      363,154  ratio=0.683813
```

This is the intended nanophotonic/generative win: separable fields and polynomial streams are represented by compact exact recipes rather than generic byte statistics.

```text
versioned_media raw=4,194,304 bytes
  TAR-zstd19                   1,049,535  ratio=0.250229
  HVCE-v4-fast                 1,049,957  ratio=0.250329
  HVCE-v4-balanced             1,049,961  ratio=0.250330
  TAR-xz9                      1,050,784  ratio=0.250526
  ZIP-deflate9                 4,196,038  ratio=1.000413
```

The versioned-media result shows the near-duplicate archive-state branch. Locally high-entropy streams become compressible when multiple versions are archived together.

```text
already_compressed raw=918,000 bytes
  TAR-zstd19                     918,865  ratio=1.000942
  ZIP-deflate9                   918,948  ratio=1.001033
  HVCE-v4-fast                   919,110  ratio=1.001209
  HVCE-v4-balanced               919,113  ratio=1.001212
  TAR-gzip9                      920,056  ratio=1.002240
  TAR-xz9                        924,224  ratio=1.006780
```

The correct result is near-raw size. HVCE does not waste extreme CPU trying to compress entropy-saturated containers.

```text
random_control raw=1,572,864 bytes
  TAR-zstd19                   1,573,526  ratio=1.000421
  HVCE-v4-fast                 1,573,609  ratio=1.000474
  HVCE-v4-balanced             1,573,613  ratio=1.000476
  ZIP-deflate9                 1,573,666  ratio=1.000510
  TAR-gzip9                    1,574,225  ratio=1.000865
  TAR-xz9                      1,576,664  ratio=1.002416
```

Random data cannot be universally compressed losslessly. HVCE stays near raw size and beats ZIP/gzip/xz by small header-level margins in this run, while zstd remains slightly smaller.

## Reproduce

```bash
python benchmark_hvce.py --out-dir benchmarks/local_run
```

## Honest conclusion

HVCE v4 has strong, reproducible wins over ZIP/Deflate and meaningful structured-data wins over xz/zstd on generated corpora that match its exact recipes. It is not yet faster than native C compressors. Speed dominance requires the planned Rust/C kernel.