[ValisSowilo]

What

Text segments are dictionary-coded before modelling: each frequent word becomes a 1–3 byte code (0x80–0xFF) from a dictionary built from the segment itself, carried at the head of the stream. Case is a flag byte. Inside the 1- and 2-byte classes, words are ordered by recursive principal-component bisection of their neighbour statistics, so a code's lead byte names a cluster of words used alike. The model is transform-aware:

• code bytes hash as word characters
• there's a mixer weight set for mid-code positions
• at -5/-7/-9, transformed segments get four extra contexts: skip bigram, word-class trigram from code lead bytes, word + punctuation, and previous word + order 3.

Design, every measured alternative, and the gates: ARCHITECTURE.md §29.

Results (−9, round-trip verified, same machine)

input	1.0.2	1.1
enwik9 -s1000	157,073,381	148,026,632	−5.76%
enwik8 -s1000	18,810,680	18,027,359	−4.16%
Silesia per-file total	35,583,396	35,468,198	−0.32%
— xml / dickens / webster			−2.88% / −1.97% / −1.20%
— other nine files		byte-identical	0

Against the published tables, that would place about 29th of 227 on LTCB (listed 35th) and 48th of 322 on Silesia (listed 49th). It hasn't been submitted.

Per preset on 16 MB of enwik8 (A-B-B-A ×4, medians): -f1 −10.2% / −15% time, -1 −6.0% / −25%, -3 −2.1% / −26%, -5 −3.5% / ±0%, -7 −2.5% / −6%, -9 −2.7% / −13%.

Costs:

• Transformed text at -9 peaks about 260 MB higher (four 2^22 tables). Silesia's max peak goes from 1,055 MB to 1,188 MB (webster).
• enwik9 peaks lower (2,847 / 3,235 MB vs 3,033 / 3,836 MB).
• Files under 1 MB gain less, and files under 256 KB are left alone.

Format

• New segment kind 2 and archive format v3. 1.1 reads v2 and v3; 1.0 refuses v3 with a clear version message (checked both ways).
• -w0 turns the transform off.
• Release string is now 1.1.0 (it said 1.0.1 at 1.0.2).

Verification

• sfuzz.py: +26 kind-2 attack cases and a third positive control. 68 cases × 2 modes pass.
• fuzz.py (648 round trips), tfuzz.py (60), gfuzz.py (300 trials): pass.
• Transform-specific round-trip fuzz covering flag bytes, high bytes, case forms and the 32-letter limit: 190 cases at -1/-5/-9, 0 failures.
• Gates checked by forcing the transform on files it declines: samba +10.5%, reymont +2.2%.

Not done here

• The top-of-README preset table (8 presets × Silesia, one session) is still 1.0.2's and is marked as such. Re-running scripts/bench_session.py takes about 2.5 h.
• The 1.1 enwik9 times aren't claimed; the machine wasn't quiet for that run.
• No release has been tagged and nothing has been resubmitted to Mahoney.