# RRLS v0.8 Universal Corpus Experiment

## Corpus

The controlled benchmark deliberately mixes unrelated coordinate families:

- synthetic English, German, Spanish, French and Dutch UTF-8 text;
- the same language process represented literally as 0/1 coordinate streams;
- Python-like code, JSON, SQL and machine-log syntax;
- DNA-like symbol sequences with recurring motifs;
- sine, polynomial, recurrence, sparse and PRNG-like binary numerical streams;
- procedural stripes, checkerboards, gradients, circles, waves, blocks, cellular automata and noise images;
- alternate image coordinate views (raw pixels, horizontal XOR, vertical XOR);
- OS-entropy bytes as an observer-level high-entropy control.

No semantic labels are exposed to the compressor. Domain labels are retained only for post-hoc measurement.

## Main result

11 MiB run:
- best recursive depth: L5
- raw byte-symbols -> token reduction: 2,467,333.0 final tokens, 2.0581x MDL proxy compression
- dictionary storage proxy is included.

22 MiB scaling run:
- corpus: 22.01 MiB
- best recursive depth: L5
- token reduction: 4.8505x
- MDL proxy compression: 2.1423x
- optimum remained at level 5.

## Held-out domain behavior at L5

| family   |   p5_ratio |
|:---------|-----------:|
| bio      |   2.66165  |
| code     |   6.03948  |
| control  |   0.709763 |
| image    |   2.93016  |
| numeric  |   0.791543 |
| text     |   3.55517  |
| textbits |   7.99431  |

The high-entropy control expands rather than compresses. RRLS should therefore retain raw representation for such regions instead of pretending a learned dictionary is useful.

## Coordinate-view search over image families

| image_family   | cheapest_coordinate_view   |   heldout_payload_compression |
|:---------------|:---------------------------|------------------------------:|
| stripes        | vxor                       |                    455.111    |
| checker        | vxor                       |                     82.7475   |
| gradient       | vxor                       |                     15.5741   |
| blocks         | vxor                       |                     13.2449   |
| ca             | hxor                       |                      5.49614  |
| waves          | raw                        |                      2.74992  |
| circles        | vxor                       |                      1.22333  |
| noise          | raw                        |                      0.708773 |

Different visual families prefer different coordinate systems. This is a concrete example of the RRLS rotate/view principle: a relationship can become drastically cheaper under one representation and more expensive under another.

## Vocabulary over vocabulary

Literal cross-domain phrases are uncommon. The stronger universal layer appears after removing literal symbol identity and grouping learned words by relationship geometry.

Top structural meta-relations:

| meta_relation             |   heldout_uses |   learned_tokens |   effective_families | families                                                                                 |
|:--------------------------|---------------:|-----------------:|---------------------:|:-----------------------------------------------------------------------------------------|
| BINARY/HIGH-ALTERNATION   |         261628 |            23420 |              3.96565 | numeric:100719; image:86214; control:41677; code:19095; text:13712; textbits:165; bio:46 |
| PRINTABLE/HIGH-DIVERSITY  |         105009 |             6991 |              3.15295 | text:41943; code:32479; bio:29015; image:1564; control:4; numeric:4                      |
| XOR/CONST-DELTA           |            931 |               61 |              3.00076 | image:507; text:266; code:121; numeric:33; textbits:2; control:2                         |
| BINARY/MIXED-RUNS         |          46079 |             2564 |              2.76706 | textbits:25740; image:15164; bio:4098; numeric:583; code:272; text:220; control:2        |
| HIGH-DIVERSITY            |          57640 |             5520 |              2.52979 | image:40412; text:7323; code:6196; numeric:3703; control:6                               |
| MID-ALPHABET/REPETITIVE   |           5515 |              562 |              2.52534 | image:3423; bio:1087; code:1005                                                          |
| RUN/CONSTANT              |          21585 |              327 |              1.79055 | image:18437; numeric:1951; code:563; control:271; text:256; bio:96; textbits:11          |
| PERIOD/2                  |           1816 |               28 |              1.71979 | image:1481; bio:298; textbits:37                                                         |
| PERIOD/3                  |           1833 |               34 |              1.54168 | image:1620; bio:143; textbits:70                                                         |
| PERIOD/4                  |           1657 |               33 |              1.5098  | image:1439; textbits:209; bio:9                                                          |
| PERIOD/5-8                |            632 |               16 |              1.50577 | image:542; textbits:90                                                                   |
| AFFINE/CONST-DELTA-MOD256 |           8147 |              573 |              1.11929 | image:7980; code:136; numeric:20; text:10; control:1                                     |
| PERIOD/9-32               |            383 |                8 |              1.09555 | image:376; textbits:7                                                                    |
| BINARY/LONG-RUNS          |           4828 |               90 |              1.07872 | image:4758; numeric:70                                                                   |
| LOW-ALPHABET/REPETITIVE   |            545 |               88 |              1       | image:545                                                                                |

This is the current candidate low-level meta-language:

- RUN / CONSTANT
- small periods (2, 3, 4, ...)
- constant XOR-delta
- constant modular delta
- binary long/mixed/high-alternation runs
- low/mid-alphabet repetition
- high-diversity printable and non-printable fragments

These are empirical compressor classes, not claims that they are fundamental laws of physics or complete models of language, vision or biology.

## Important limitation

The benchmark data is controlled/generated rather than a representative web-scale natural corpus. LZW-style recursive phrases capture repetition and adjacency but do not yet capture arbitrary long-range, spatial, semantic, causal or algorithmic relations. v0.8 therefore adds a Universal Corpus Lab so arbitrary real files can be supplied and future relation operators can be compared against this baseline.