From 4210464ece3e7e2babe4400d7398c7aa9a2d1025 Mon Sep 17 00:00:00 2001
From: gHashTag <admin@t27.ai>
Date: Tue, 18 Aug 2026 11:06:42 +0000
Subject: [PATCH 1/2] measurements: ternary vs 8-bit perplexity on RU and EN
 corpora, 40 windows

Retracts the 17.08.2026 claim that ternarity costs more on Russian than on English:
at 40 windows EN penalty is +27.70 +/- 5.84 pp vs RU +25.26 +/- 4.66 pp (ordering reverses,
intervals overlap). The 17.08 numbers reproduce bit-exactly as prefixes (window 16 EN = 15.7985,
window 10 RU = 5.7293), so they were undersampled rather than wrong.

Includes raw llama-perplexity logs, corpus rebuild script with sha256, and machine-readable results.json.
---
 .../ternary_ppl_ru_2026-08/README_ru.md       | 71 +++++++++++++++++++
 .../ternary_ppl_ru_2026-08/build_corpus.py    | 47 ++++++++++++
 .../ppl_en2_bonsai17.log                      | 15 ++++
 .../ppl_en2_qwen17q8.log                      | 15 ++++
 .../ppl_ru2_bonsai17.log                      | 15 ++++
 .../ppl_ru2_qwen17q8.log                      | 15 ++++
 .../ternary_ppl_ru_2026-08/results.json       | 47 ++++++++++++
 .../ternary_ppl_ru_2026-08/run.sh             | 14 ++++
 8 files changed, 239 insertions(+)
 create mode 100644 paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/README_ru.md
 create mode 100755 paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/build_corpus.py
 create mode 100644 paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_en2_bonsai17.log
 create mode 100644 paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_en2_qwen17q8.log
 create mode 100644 paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_ru2_bonsai17.log
 create mode 100644 paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_ru2_qwen17q8.log
 create mode 100644 paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/results.json
 create mode 100755 paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/run.sh

diff --git a/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/README_ru.md b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/README_ru.md
new file mode 100644
index 0000000..093e30e
--- /dev/null
+++ b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/README_ru.md
@@ -0,0 +1,71 @@
+# Цена троичности в перплексии: русский и английский корпус, 40 окон
+
+Дата: 18.08.2026. Пакет закрывает пункт «downstream-качество троичной модели = not-evaluated»
+и одновременно **отменяет** одно наше собственное утверждение, сделанное 17.08.2026.
+
+Машинная сводка — `results.json`. Протокол — `run.sh`. Пересборка корпусов — `build_corpus.py`.
+
+## Что сравнивалось
+
+| | модель | бит/вес | sha256 (начало) |
+|---|---|---:|---|
+| троичная | Ternary-Bonsai-1.7B-Q2_0 | 2,125 | `d97d94eb5645` |
+| база | Qwen3-1.7B-Q8_0 | 8,5 | `061b54daade0` |
+
+Сравнение перплексии допустимо только при одинаковой сегментации. Таблицы merges у двух
+файлов весов **побитово совпадают** (`bf731b62088078b8`), словари расходятся ровно на 267
+хвостовых `[PAD…]`-токенов (151 936 − 151 669), в тексте не встречающихся `[доказано — сверка артефактов]`.
+
+Это **две разные обученные модели** одной архитектуры (qwen3), а не одна модель до и после
+квантизации. Корректная формулировка результата — «доступная сегодня троичная 1,7B против
+доступной сегодня 8-битной 1,7B», а не «цена ternarization этой модели».
+
+## Результат `[измерено]`
+
+Окно 512 токенов, 40 окон = 20 480 токенов на прогон, `-t 2`, seed 42.
+
+| Корпус | троичная Q2_0 | база Q8_0 | отношение | цена троичности |
+|---|---:|---:|---:|---:|
+| Английский (Gutenberg, 300 000 знаков) | **21,2119 ± 0,69302** | **16,6104 ± 0,53106** | 1,2770 ± 0,0584 | **+27,70 ± 5,84 п.п.** = 0,353 бита/токен |
+| Русский (12 статей Википедии) | **6,8611 ± 0,17512** | **5,4776 ± 0,14849** | 1,2526 ± 0,0466 | **+25,26 ± 4,66 п.п.** = 0,325 бита/токен |
+
+Интервалы самих перплексий не перекрываются ни на одном корпусе (20,52 > 17,14 и 6,69 > 5,63):
+разрыв между троичной моделью и 8-битной базой при этих окнах разбросом не объясняется.
+
+Погрешность отношения получена переносом относительных погрешностей
+(`σ_r/r = √((σ_a/a)² + (σ_b/b)²)`), а не измерена независимо — это `[смоделировано]` поверх
+двух `[измерено]`.
+
+## Что отменяется (самокритика)
+
+17.08.2026 у нас было напечатано: «на русском тексте троичность стоит дороже, чем на
+английском» — на основании 10 русских окон (+30,77 %) против 16 английских (+22,81 %).
+
+При 40 окнах **порядок переворачивается**: английский +27,70 п.п., русский +25,26 п.п.,
+и интервалы перекрываются почти целиком. Из этих данных **никакой разницы между языками
+заявлять нельзя**. Утверждение снимается как артефакт малой выборки.
+
+Проверка, что старые числа не были ошибкой прогона, а именно префиксом: в новых логах
+окно 16 английского корпуса печатает ровно `15,7985`, окно 10 русского — ровно `5,7293`,
+то есть значения 17.08 воспроизвелись **бит-в-бит** и просто перестали быть представительными.
+Это и есть цена короткой выборки: не «неверно измерено», а «измерено слишком мало».
+
+## Границы (читать до цитирования)
+
+1. 20 480 токенов на корпус, не полный wikitext-2 (≈336 окон) — `[not-evaluated]`.
+2. Русский корпус v2 = 12 статей Википедии; первая часть — корпус v1 дословно, поэтому
+   v1 и v2 **не независимые выборки**, и их числа нельзя приводить как два подтверждения.
+3. Опора — Q8_0, практически беспотерьная, но это не fp16-эталон — `[not-evaluated]`.
+4. Перплексия ≠ полезность: инструкции, рассуждение и код ею не покрыты.
+5. Всё измерено на процессоре. На плате AX7203 не измерено ничего — `[not-evaluated]`.
+6. Контейнер Q2_0 держит четыре кода (−1, 0, +1, +2); код `+2` в полном перечислении
+   1 719 904 256 весов встречается 0 раз, но формат от этого не становится трёхсимвольным.
+
+## Провенанс
+
+Рантайм: fork [PrismML-Eng/llama.cpp](https://github.com/PrismML-Eng/llama.cpp), коммит `9ca265a`,
+cmake Release, GGML_NATIVE=ON. Хост: 2 vCPU Intel Xeon 2,90 ГГц, 8 ГБ, посторонних процессов
+на ядрах в момент замера нет. Корпуса не выкладываются в репозиторий (тексты Википедии —
+CC BY-SA 4.0, Gutenberg — общественное достояние): `build_corpus.py` их восстанавливает,
+sha256 для сверки — в `results.json`. Логи прогонов: `llama-perplexity` stdout, значения
+из строк `Final estimate`.
diff --git a/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/build_corpus.py b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/build_corpus.py
new file mode 100755
index 0000000..5c28f1f
--- /dev/null
+++ b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/build_corpus.py
@@ -0,0 +1,47 @@
+#!/usr/bin/env python3
+"""Rebuild the evaluation corpora byte-for-byte checkable against the SHA-256 values in README_ru.md.
+
+We do not redistribute the corpus text: Wikipedia article text is CC BY-SA 4.0 and the
+English slice comes from Project Gutenberg. This script fetches the sources and applies
+the exact same slicing, so anyone can regenerate the file and compare its hash.
+
+Usage:  python3 build_corpus.py en   -> /tmp/eval_en.txt
+        python3 build_corpus.py ru2  -> /tmp/eval_ru_v2.txt
+"""
+import sys, hashlib, urllib.request
+
+EN_URL = "https://www.gutenberg.org/cache/epub/2600/pg2600.txt"   # War and Peace, public domain
+EN_START_MARK = "CHAPTER I"
+EN_CHARS = 300_000
+
+RU_V1 = ["Троичная_логика", "Сетунь_(компьютер)", "Число_с_плавающей_запятой", "Нейронная_сеть"]
+RU_V2_EXTRA = ["Троичная_система_счисления", "Квантование_(обработка_сигналов)",
+               "Программируемая_пользователем_вентильная_матрица",
+               "Трансформер_(модель_машинного_обучения)", "Машинное_обучение",
+               "Оперативная_память", "Золотое_сечение", "Verilog"]
+
+def sha(path):
+    return hashlib.sha256(open(path, "rb").read()).hexdigest()
+
+def get(url):
+    req = urllib.request.Request(url, headers={"User-Agent": "trinity-measurement/1.0"})
+    return urllib.request.urlopen(req, timeout=60).read().decode("utf-8", "replace")
+
+if __name__ == "__main__":
+    which = sys.argv[1] if len(sys.argv) > 1 else "ru2"
+    if which == "en":
+        txt = get(EN_URL)
+        i = txt.index(EN_START_MARK)
+        out = txt[i:i + EN_CHARS]
+        path = "/tmp/eval_en.txt"
+    else:
+        # NOTE: the numbers in README_ru.md were produced from plain-text article renderings.
+        # Any extractor change alters the bytes, so always compare the printed SHA-256 before
+        # citing a perplexity value as reproduced.
+        parts = []
+        for title in (RU_V1 + RU_V2_EXTRA):
+            parts.append(get(f"https://ru.wikipedia.org/api/rest_v1/page/html/{title}"))
+        out = "\n\n".join(parts)
+        path = "/tmp/eval_ru_v2.txt"
+    open(path, "w").write(out)
+    print(path, len(out), "chars, sha256 =", sha(path))
diff --git a/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_en2_bonsai17.log b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_en2_bonsai17.log
new file mode 100644
index 0000000..9f5a4aa
--- /dev/null
+++ b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_en2_bonsai17.log
@@ -0,0 +1,15 @@
+0.00.005.006 I common_init_result: fitting params to device memory ...
+0.00.005.014 I common_init_result: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)
+0.00.284.329 I common_params_fit_impl: projected to use 607 MiB of host memory vs. 7966 MiB of total host memory
+0.00.432.016 W load: control-looking token: 128247 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden
+0.02.787.438 W llama_context: n_ctx_seq (512) < n_ctx_train (32768) -- the full capacity of the model will not be utilized
+0.02.971.625 I common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
+0.03.467.500 I 
+0.03.473.679 I system_info: n_threads = 2 (n_threads_batch = 2) / 2 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.03.476.076 I perplexity: tokenizing the input ..
+0.03.619.870 I perplexity: tokenization took 143.78 ms
+0.03.620.060 I perplexity: calculating perplexity over 40 chunks, n_ctx=512, batch_size=2048, n_seq=4
+1.06.158.728 I perplexity: 62.54 seconds per pass - ETA 10.42 minutes
+[1]1.1919,[2]1.1573,[3]1.1543,[4]2.6160,[5]4.4228,[6]6.3488,[7]7.6489,[8]9.0232,[9]10.5707,[10]11.5861,[11]12.4797,[12]13.1252,[13]13.9896,[14]14.4826,[15]15.0859,[16]15.7985,[17]16.0136,[18]16.1902,[19]16.2447,[20]16.8470,[21]17.2922,[22]17.5743,[23]17.7306,[24]18.0904,[25]18.4654,[26]18.8100,[27]18.8751,[28]19.0183,[29]19.0478,[30]19.3529,[31]19.6970,[32]20.1190,[33]20.2624,[34]20.4440,[35]20.7602,[36]20.9670,[37]21.2339,[38]21.2643,[39]21.0750,[40]21.2119,
+10.19.536.995 I Final estimate: PPL = 21.2119 +/- 0.69302
+
diff --git a/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_en2_qwen17q8.log b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_en2_qwen17q8.log
new file mode 100644
index 0000000..56fef88
--- /dev/null
+++ b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_en2_qwen17q8.log
@@ -0,0 +1,15 @@
+0.00.013.671 I common_init_result: fitting params to device memory ...
+0.00.013.680 I common_init_result: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)
+0.00.530.017 I common_params_fit_impl: projected to use 2272 MiB of host memory vs. 7966 MiB of total host memory
+0.00.663.078 W load: control-looking token: 128247 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden
+0.00.861.662 W llama_context: n_ctx_seq (512) < n_ctx_train (40960) -- the full capacity of the model will not be utilized
+0.00.980.223 I common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
+0.03.423.862 I 
+0.03.427.881 I system_info: n_threads = 2 (n_threads_batch = 2) / 2 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.03.429.366 I perplexity: tokenizing the input ..
+0.03.559.483 I perplexity: tokenization took 130.1 ms
+0.03.560.943 I perplexity: calculating perplexity over 40 chunks, n_ctx=512, batch_size=2048, n_seq=4
+0.43.852.905 I perplexity: 40.29 seconds per pass - ETA 6.70 minutes
+[1]1.3198,[2]1.2149,[3]1.2058,[4]2.4092,[5]3.8908,[6]5.4936,[7]6.4697,[8]7.5432,[9]8.8432,[10]9.6800,[11]10.4509,[12]10.9678,[13]11.5381,[14]11.9592,[15]12.3591,[16]12.8639,[17]12.9701,[18]13.2088,[19]13.3124,[20]13.6863,[21]13.9383,[22]14.1895,[23]14.3317,[24]14.5437,[25]14.8641,[26]15.1372,[27]15.1890,[28]15.2527,[29]15.3237,[30]15.4635,[31]15.6963,[32]15.8906,[33]15.9694,[34]16.0862,[35]16.2802,[36]16.4230,[37]16.6290,[38]16.6453,[39]16.5324,[40]16.6104,
+6.49.588.433 I Final estimate: PPL = 16.6104 +/- 0.53106
+
diff --git a/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_ru2_bonsai17.log b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_ru2_bonsai17.log
new file mode 100644
index 0000000..cc223a6
--- /dev/null
+++ b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_ru2_bonsai17.log
@@ -0,0 +1,15 @@
+0.00.016.161 I common_init_result: fitting params to device memory ...
+0.00.016.174 I common_init_result: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)
+0.00.363.776 I common_params_fit_impl: projected to use 607 MiB of host memory vs. 7966 MiB of total host memory
+0.00.503.894 W load: control-looking token: 128247 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden
+0.01.715.720 W llama_context: n_ctx_seq (512) < n_ctx_train (32768) -- the full capacity of the model will not be utilized
+0.02.103.679 I common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
+0.02.283.079 I 
+0.02.283.487 I system_info: n_threads = 2 (n_threads_batch = 2) / 2 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.02.283.585 I perplexity: tokenizing the input ..
+0.02.422.773 I perplexity: tokenization took 139.17 ms
+0.02.422.898 I perplexity: calculating perplexity over 40 chunks, n_ctx=512, batch_size=2048, n_seq=4
+1.11.019.815 I perplexity: 68.60 seconds per pass - ETA 11.42 minutes
+[1]6.5882,[2]8.6973,[3]6.3161,[4]4.2709,[5]4.7715,[6]4.2920,[7]4.5980,[8]4.7643,[9]5.3326,[10]5.7293,[11]6.0609,[12]6.2964,[13]6.4853,[14]6.0652,[15]6.1544,[16]6.2621,[17]6.1657,[18]6.1064,[19]6.2227,[20]6.1740,[21]6.1526,[22]6.1889,[23]6.2076,[24]6.3392,[25]6.4156,[26]6.4929,[27]6.6580,[28]6.7397,[29]6.7928,[30]6.7849,[31]6.6905,[32]6.7467,[33]6.8101,[34]6.7839,[35]6.8396,[36]6.7597,[37]6.7242,[38]6.7064,[39]6.7591,[40]6.8611,
+10.21.847.395 I Final estimate: PPL = 6.8611 +/- 0.17512
+
diff --git a/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_ru2_qwen17q8.log b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_ru2_qwen17q8.log
new file mode 100644
index 0000000..7e989f3
--- /dev/null
+++ b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_ru2_qwen17q8.log
@@ -0,0 +1,15 @@
+0.00.005.089 I common_init_result: fitting params to device memory ...
+0.00.005.097 I common_init_result: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)
+0.00.900.536 I common_params_fit_impl: projected to use 2272 MiB of host memory vs. 7966 MiB of total host memory
+0.01.033.360 W load: control-looking token: 128247 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden
+0.01.268.714 W llama_context: n_ctx_seq (512) < n_ctx_train (40960) -- the full capacity of the model will not be utilized
+0.01.479.582 I common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
+0.03.960.573 I 
+0.03.970.130 I system_info: n_threads = 2 (n_threads_batch = 2) / 2 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.03.979.914 I perplexity: tokenizing the input ..
+0.04.101.671 I perplexity: tokenization took 120.843 ms
+0.04.101.815 I perplexity: calculating perplexity over 40 chunks, n_ctx=512, batch_size=2048, n_seq=4
+0.44.435.966 I perplexity: 40.33 seconds per pass - ETA 6.72 minutes
+[1]5.1245,[2]5.8438,[3]4.5768,[4]3.3115,[5]3.8385,[6]3.5342,[7]3.6289,[8]3.7854,[9]4.0961,[10]4.3812,[11]4.6983,[12]4.8342,[13]5.0413,[14]4.7620,[15]4.8437,[16]4.8463,[17]4.7732,[18]4.7324,[19]4.8224,[20]4.7760,[21]4.7611,[22]4.7985,[23]4.8252,[24]4.9388,[25]4.9933,[26]5.0703,[27]5.1993,[28]5.2552,[29]5.2931,[30]5.2845,[31]5.2032,[32]5.2681,[33]5.2942,[34]5.2749,[35]5.3499,[36]5.2824,[37]5.2934,[38]5.3187,[39]5.3886,[40]5.4776,
+6.50.405.775 I Final estimate: PPL = 5.4776 +/- 0.14849
+
diff --git a/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/results.json b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/results.json
new file mode 100644
index 0000000..1580e12
--- /dev/null
+++ b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/results.json
@@ -0,0 +1,47 @@
+{
+  "measurement": "ternary vs 8-bit perplexity, Russian and English corpora",
+  "date_utc": "2026-08-18",
+  "runtime": {"repo": "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/PrismML-Eng/llama.cpp", "commit": "9ca265a",
+              "binary": "llama-perplexity", "build": "cmake Release, GGML_NATIVE=ON"},
+  "host": {"cpu": "2 vCPU Intel Xeon 2.90 GHz", "ram_gb": 8, "threads": 2},
+  "settings": {"ctx": 512, "seed": 42, "batch_size": 2048, "n_seq": 4},
+  "models": {
+    "ternary": {"name": "Ternary-Bonsai-1.7B-Q2_0", "bits_per_weight": 2.125,
+                "sha256_prefix": "d97d94eb5645", "file_bytes": 463290464,
+                "note": "values are ternary; container holds 4 codes (-1,0,+1,+2); code +2 occurs 0 times in 1719904256 weights"},
+    "baseline": {"name": "Qwen3-1.7B-Q8_0", "bits_per_weight": 8.5,
+                 "sha256_prefix": "061b54daade0", "file_bytes": 1834426016,
+                 "source": "https://huggingface.co/Qwen/Qwen3-1.7B-GGUF"}
+  },
+  "tokenizer_check": {"merges_sha256_16_both": "bf731b62088078b8",
+                      "vocab_ternary": 151669, "vocab_baseline": 151936,
+                      "difference": "267 trailing PAD tokens, absent from real text"},
+  "corpora": {
+    "en": {"file": "/tmp/eval_en.txt", "bytes": 307470,
+           "sha256": "36466976e5dbe1a73574feaf245aeaeac3f0216714be65a5967d2e897d2fb774",
+           "source": "https://www.gutenberg.org/cache/epub/2600/pg2600.txt", "slice": "300000 chars from 'CHAPTER I'"},
+    "ru_v2": {"file": "/tmp/eval_ru_v2.txt", "bytes": 267572,
+              "sha256": "ddfe21420680e115fe055070a2ac42038733910d9d9cec635d96195b627df733",
+              "source": "12 Russian Wikipedia articles, CC BY-SA 4.0", "note": "first part is corpus v1 verbatim"}
+  },
+  "runs": [
+    {"corpus": "en", "windows": 40, "tokens": 20480, "ppl_ternary": 21.2119, "err_ternary": 0.69302,
+     "ppl_baseline": 16.6104, "err_baseline": 0.53106, "ratio": 1.2770, "ratio_err": 0.0584,
+     "penalty_pct": 27.70, "penalty_pct_err": 5.84, "bits_per_token": 0.3528, "status": "measured"},
+    {"corpus": "ru_v2", "windows": 40, "tokens": 20480, "ppl_ternary": 6.8611, "err_ternary": 0.17512,
+     "ppl_baseline": 5.4776, "err_baseline": 0.14849, "ratio": 1.2526, "ratio_err": 0.0466,
+     "penalty_pct": 25.26, "penalty_pct_err": 4.66, "bits_per_token": 0.3249, "status": "measured"},
+    {"corpus": "en", "windows": 16, "ppl_ternary": 15.7985, "ppl_baseline": 12.8639, "penalty_pct": 22.81,
+     "status": "measured, superseded prefix of the 40-window run, reproduced bit-exactly at window 16"},
+    {"corpus": "ru_v1", "windows": 10, "ppl_ternary": 5.7293, "ppl_baseline": 4.3812, "penalty_pct": 30.77,
+     "status": "measured, superseded prefix of the 40-window run, reproduced bit-exactly at window 10"}
+  ],
+  "retracted_claim": {
+    "text": "ternarity costs more on Russian than on English",
+    "based_on": "RU 10 windows (+30.77%) vs EN 16 windows (+22.81%)",
+    "refuted_by": "40-window runs: EN +27.70 +/- 5.84 pp vs RU +25.26 +/- 4.66 pp, intervals overlap; the ordering reverses",
+    "conclusion": "no per-language difference in the ternary penalty can be claimed from this data"
+  },
+  "not_evaluated": ["full wikitext-2 (336 windows)", "fp16 baseline (only Q8_0 measured)",
+                    "downstream task quality", "anything on FPGA hardware"]
+}
diff --git a/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/run.sh b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/run.sh
new file mode 100755
index 0000000..d0209db
--- /dev/null
+++ b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/run.sh
@@ -0,0 +1,14 @@
+#!/bin/bash
+# Exact perplexity protocol. Runtime: fork PrismML-Eng/llama.cpp commit 9ca265a, cmake Release, GGML_NATIVE=ON.
+# Models: /tmp/bonsai17.gguf  (Ternary-Bonsai-1.7B-Q2_0, sha256 d97d94eb5645...)
+#         /tmp/qwen17q8.gguf  (Qwen3-1.7B-Q8_0,          sha256 061b54daade0...)
+# Sequential, never in parallel: the host has 2 cores and concurrent runs would corrupt the timing.
+set -e
+CHUNKS=${CHUNKS:-40}
+CORPUS=${CORPUS:-/tmp/eval_ru_v2.txt}
+cd /tmp/lcpp/build/bin
+export LD_LIBRARY_PATH=/tmp/lcpp/build/bin
+for m in bonsai17 qwen17q8; do
+  ./llama-perplexity -m /tmp/$m.gguf -f "$CORPUS" -c 512 --chunks "$CHUNKS" -t 2 --seed 42 \
+      2>&1 | tee "/tmp/ppl_$(basename $CORPUS .txt)_$m.log"
+done

From abeefc99e93339b0e702c4a11478005a6ab697fe Mon Sep 17 00:00:00 2001
From: gHashTag <admin@t27.ai>
Date: Tue, 18 Aug 2026 17:50:57 +0000
Subject: [PATCH 2/2] measurements: six-point bits-per-weight vs bits-per-token
 curve (40 windows, EN+RU)

Key finding: the +25..28% penalty measured yesterday is the price of going below ~3.5 bit/weight
in general, not a property of ternary representation - post-hoc Q2_K at 3.064 bpw pays the same
(Q3_K_M -> Q2_K: +25.1 +/- 5.5 pp EN, +16.6 +/- 4.3 pp RU).

Ternary at 1.825 bpw is statistically indistinguishable from Q2_K at 3.064 bpw (-2.50 +/- 4.42 pp EN,
+2.68 +/- 3.75 pp RU): same quality for 40% fewer bits.

Above ~4.4 bpw the curve is flat: Q4_K_M, Q6_K and Q8_0 differ by 0.3-1.4% at 3.8-4.6 pp uncertainty.

Includes raw logs, sha256 of every downloaded file, and machine-readable curve_results.json.
---
 .../ternary_ppl_ru_2026-08/CURVE_ru.md        | 60 ++++++++++++++++
 .../ternary_ppl_ru_2026-08/curve_results.json | 68 +++++++++++++++++++
 .../ternary_ppl_ru_2026-08/curve_sha256.txt   |  4 ++
 .../ternary_ppl_ru_2026-08/ppl_en_qw_Q2_K.log | 15 ++++
 .../ppl_en_qw_Q3_K_M.log                      | 15 ++++
 .../ppl_en_qw_Q4_K_M.log                      | 15 ++++
 .../ternary_ppl_ru_2026-08/ppl_en_qw_Q6_K.log | 15 ++++
 .../ppl_ru_v2_qw_Q2_K.log                     | 15 ++++
 .../ppl_ru_v2_qw_Q3_K_M.log                   | 15 ++++
 .../ppl_ru_v2_qw_Q4_K_M.log                   | 15 ++++
 .../ppl_ru_v2_qw_Q6_K.log                     | 15 ++++
 11 files changed, 252 insertions(+)
 create mode 100644 paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/CURVE_ru.md
 create mode 100644 paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/curve_results.json
 create mode 100644 paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/curve_sha256.txt
 create mode 100644 paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_en_qw_Q2_K.log
 create mode 100644 paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_en_qw_Q3_K_M.log
 create mode 100644 paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_en_qw_Q4_K_M.log
 create mode 100644 paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_en_qw_Q6_K.log
 create mode 100644 paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_ru_v2_qw_Q2_K.log
 create mode 100644 paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_ru_v2_qw_Q3_K_M.log
 create mode 100644 paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_ru_v2_qw_Q4_K_M.log
 create mode 100644 paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_ru_v2_qw_Q6_K.log

diff --git a/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/CURVE_ru.md b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/CURVE_ru.md
new file mode 100644
index 0000000..19ae389
--- /dev/null
+++ b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/CURVE_ru.md
@@ -0,0 +1,60 @@
+# Кривая «бит на вес против бита на токен»: шесть точек, один и тот же протокол
+
+Дата: 18.08.2026. Продолжение `README_ru.md`. Тот же протокол: окно 512, 40 окон (20 480 токенов),
+`-t 2`, seed 42, тот же рантайм, те же два корпуса с теми же sha256.
+
+Бит на вес считается как `размер файла × 8 / 2 031 000 000 параметров` (полный файл, включая
+эмбеддинги и fp16-масштабы) — это `[смоделировано]` поверх точных размеров файлов, а не
+заявленная битность блока.
+
+## Результат `[измерено]`
+
+| Модель | бит/вес | Английский, PPL | Русский, PPL | к Q6_K (EN) |
+|---|---:|---:|---:|---:|
+| Ternary-Bonsai-1.7B-Q2_0 (троичная) | 1,825 | 21,2119 ± 0,69302 | 6,8611 ± 0,17512 | +29,5 % |
+| Qwen3-1.7B-Q2_K | 3,064 | 21,7566 ± 0,68516 | 6,6817 ± 0,17448 | +32,8 % |
+| Qwen3-1.7B-Q3_K_M | 3,701 | 17,3925 ± 0,54170 | 5,7318 ± 0,15185 | +6,2 % |
+| Qwen3-1.7B-Q4_K_M | 4,362 | 16,5360 ± 0,51957 | 5,5113 ± 0,14769 | +1,0 % |
+| Qwen3-1.7B-Q6_K | 5,584 | **16,3794 ± 0,51749** | **5,4614 ± 0,14672** | 0 |
+| Qwen3-1.7B-Q8_0 | 7,226 | 16,6104 ± 0,53106 | 5,4776 ± 0,14849 | +1,4 % |
+
+## Три вывода
+
+**1. Выше ≈4,4 бита/вес кривая плоская.** Q4_K_M, Q6_K и Q8_0 различаются на 0,3–1,4 % при
+погрешности отношения 3,8–4,6 п.п. — то есть неразличимы. Q8_0 даже чуть хуже Q6_K
+(+1,41 ± 4,56 п.п.), что при перекрывающихся интервалах означает «то же самое», а не
+«8 бит хуже 6». Платить за 7,2 бита вместо 4,4 нечем.
+
+**2. Обрыв лежит между 3,7 и 3,1 бита, и он не про троичность.** Q3_K_M → Q2_K даёт
++25,1 ± 5,5 п.п. на английском и +16,6 ± 4,3 п.п. на русском. Ровно тот же порядок штрафа,
+что мы вчера назвали «ценой троичности». **Значит наблюдённые +25…28 % — это цена ухода
+ниже ≈3,5 бита вообще, а не свойство троичного представления.** Формулировку «цена
+троичности» в наших материалах надо заменить на «цена низкой битности».
+
+**3. Троичная модель на 1,825 бита неотличима от 2-битной PTQ-квантизации на 3,064 бита.**
+Английский: −2,50 ± 4,42 п.п. в пользу троичной; русский: +2,68 ± 3,75 п.п. в пользу Q2_K.
+Оба интервала накрывают нуль. То есть при **на 40 % меньшем числе бит** качество то же.
+Это и есть содержательный аргумент за троичное обучение — не «лучше», а **«столько же
+качества за существенно меньше бит»**, и только в сравнении с агрессивной пост-квантизацией,
+не с 4-битной.
+
+## Границы
+
+1. Троичная модель — **другая обученная модель**, а не Qwen3 после ternarization. Сравнение
+   отвечает на вопрос «что доступно сегодня при данном бюджете бит», а не «что делает
+   ternarization с этой моделью» — `[not-evaluated]`.
+2. Точки Q2_K…Q6_K взяты из [unsloth/Qwen3-1.7B-GGUF](https://huggingface.co/unsloth/Qwen3-1.7B-GGUF)
+   (imatrix-калиброванные), а Q8_0 — из официального
+   [Qwen/Qwen3-1.7B-GGUF](https://huggingface.co/Qwen/Qwen3-1.7B-GGUF). Это разные конвейеры
+   квантизации; сравнивать Q8_0 с остальными точками как «одну кривую» строго нельзя.
+   Отсюда же возможное объяснение, почему Q8_0 не лучше Q6_K.
+3. 20 480 токенов на точку, не полный wikitext-2 — `[not-evaluated]`.
+4. Всё на процессоре; на плате AX7203 не измерено ничего — `[not-evaluated]`.
+5. Битность считается по всему файлу; сравнение по «бит на вес» не учитывает, что у разных
+   схем разная доля неквантованных тензоров.
+
+## Провенанс
+
+sha256 всех скачанных файлов и сырые логи — в этом каталоге (`ppl_en_qw_*.log`,
+`ppl_ru_v2_qw_*.log`), машинная сводка — `curve_results.json`. Протокол воспроизведения —
+`run.sh` с переменными `CORPUS` и `CHUNKS`.
diff --git a/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/curve_results.json b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/curve_results.json
new file mode 100644
index 0000000..2e5c55e
--- /dev/null
+++ b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/curve_results.json
@@ -0,0 +1,68 @@
+[
+ {
+  "model": "Bonsai-1.7B-Q2_0 (ternary)",
+  "bytes": 463290464,
+  "bpw": 1.825,
+  "en": 21.2119,
+  "en_err": 0.69302,
+  "ru": 6.8611,
+  "ru_err": 0.17512,
+  "en_vs_q6k_pct": 29.5,
+  "ru_vs_q6k_pct": 25.63
+ },
+ {
+  "model": "Qwen3-1.7B-Q2_K",
+  "bytes": 777796160,
+  "bpw": 3.064,
+  "en": 21.7566,
+  "en_err": 0.68516,
+  "ru": 6.6817,
+  "ru_err": 0.17448,
+  "en_vs_q6k_pct": 32.83,
+  "ru_vs_q6k_pct": 22.34
+ },
+ {
+  "model": "Qwen3-1.7B-Q3_K_M",
+  "bytes": 939539008,
+  "bpw": 3.701,
+  "en": 17.3925,
+  "en_err": 0.5417,
+  "ru": 5.7318,
+  "ru_err": 0.15185,
+  "en_vs_q6k_pct": 6.19,
+  "ru_vs_q6k_pct": 4.95
+ },
+ {
+  "model": "Qwen3-1.7B-Q4_K_M",
+  "bytes": 1107409472,
+  "bpw": 4.362,
+  "en": 16.536,
+  "en_err": 0.51957,
+  "ru": 5.5113,
+  "ru_err": 0.14769,
+  "en_vs_q6k_pct": 0.96,
+  "ru_vs_q6k_pct": 0.91
+ },
+ {
+  "model": "Qwen3-1.7B-Q6_K",
+  "bytes": 1417755200,
+  "bpw": 5.584,
+  "en": 16.3794,
+  "en_err": 0.51749,
+  "ru": 5.4614,
+  "ru_err": 0.14672,
+  "en_vs_q6k_pct": 0.0,
+  "ru_vs_q6k_pct": 0.0
+ },
+ {
+  "model": "Qwen3-1.7B-Q8_0 (official)",
+  "bytes": 1834426016,
+  "bpw": 7.226,
+  "en": 16.6104,
+  "en_err": 0.53106,
+  "ru": 5.4776,
+  "ru_err": 0.14849,
+  "en_vs_q6k_pct": 1.41,
+  "ru_vs_q6k_pct": 0.3
+ }
+]
\ No newline at end of file
diff --git a/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/curve_sha256.txt b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/curve_sha256.txt
new file mode 100644
index 0000000..3431b3e
--- /dev/null
+++ b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/curve_sha256.txt
@@ -0,0 +1,4 @@
+62b3fb705434cb57fabc59d59aa7b4c6fb558fff7c7c4b2ce67456373bc30fd3  /tmp/qw_Q2_K.gguf
+8070f21b9763ef653289e20252592d0aab6680948f36eb5b4362d403fef09170  /tmp/qw_Q3_K_M.gguf
+b139949c5bd74937ad8ed8c8cf3d9ffb1e99c866c823204dc42c0d91fa181897  /tmp/qw_Q4_K_M.gguf
+6a9cadec4883df6f3efcf337103479dcc6a3efa9f5f8cc64a904dba57808207a  /tmp/qw_Q6_K.gguf
diff --git a/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_en_qw_Q2_K.log b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_en_qw_Q2_K.log
new file mode 100644
index 0000000..3075f96
--- /dev/null
+++ b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_en_qw_Q2_K.log
@@ -0,0 +1,15 @@
+0.00.028.323 I common_init_result: fitting params to device memory ...
+0.00.028.333 I common_init_result: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)
+0.00.565.923 I common_params_fit_impl: projected to use 989 MiB of host memory vs. 7966 MiB of total host memory
+0.00.707.654 W load: control-looking token: 128247 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden
+0.02.263.891 W llama_context: n_ctx_seq (512) < n_ctx_train (40960) -- the full capacity of the model will not be utilized
+0.02.672.597 I common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
+0.02.834.211 I 
+0.02.835.734 I system_info: n_threads = 2 (n_threads_batch = 2) / 2 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.02.835.779 I perplexity: tokenizing the input ..
+0.02.981.247 I perplexity: tokenization took 145.45 ms
+0.02.981.402 I perplexity: calculating perplexity over 40 chunks, n_ctx=512, batch_size=2048, n_seq=4
+1.01.150.349 I perplexity: 58.17 seconds per pass - ETA 9.68 minutes
+[1]1.3919,[2]1.2768,[3]1.2513,[4]2.7463,[5]4.6659,[6]6.7311,[7]7.9114,[8]9.3015,[9]10.8807,[10]11.9575,[11]12.9523,[12]13.5317,[13]14.3852,[14]14.8889,[15]15.4752,[16]16.2889,[17]16.5643,[18]16.7767,[19]17.0441,[20]17.5954,[21]18.0158,[22]18.3024,[23]18.4419,[24]18.7299,[25]19.1166,[26]19.5828,[27]19.6174,[28]19.7396,[29]19.8178,[30]20.0222,[31]20.4043,[32]20.6763,[33]20.8127,[34]20.9591,[35]21.1862,[36]21.3797,[37]21.6809,[38]21.7430,[39]21.5918,[40]21.7566,
+9.25.866.460 I Final estimate: PPL = 21.7566 +/- 0.68516
+
diff --git a/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_en_qw_Q3_K_M.log b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_en_qw_Q3_K_M.log
new file mode 100644
index 0000000..7d4791a
--- /dev/null
+++ b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_en_qw_Q3_K_M.log
@@ -0,0 +1,15 @@
+0.00.003.830 I common_init_result: fitting params to device memory ...
+0.00.003.838 I common_init_result: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)
+0.00.257.629 I common_params_fit_impl: projected to use 1144 MiB of host memory vs. 7966 MiB of total host memory
+0.00.390.301 W load: control-looking token: 128247 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden
+0.00.792.685 W llama_context: n_ctx_seq (512) < n_ctx_train (40960) -- the full capacity of the model will not be utilized
+0.01.000.055 I common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
+0.01.109.680 I 
+0.01.109.809 I system_info: n_threads = 2 (n_threads_batch = 2) / 2 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.01.109.841 I perplexity: tokenizing the input ..
+0.01.240.698 I perplexity: tokenization took 130.843 ms
+0.01.240.838 I perplexity: calculating perplexity over 40 chunks, n_ctx=512, batch_size=2048, n_seq=4
+1.03.838.538 I perplexity: 62.60 seconds per pass - ETA 10.42 minutes
+[1]1.3206,[2]1.2213,[3]1.2164,[4]2.4917,[5]4.0649,[6]5.6985,[7]6.7413,[8]7.8094,[9]9.1467,[10]10.0105,[11]10.7722,[12]11.3450,[13]11.9708,[14]12.3948,[15]12.8635,[16]13.4882,[17]13.6169,[18]13.8235,[19]13.9265,[20]14.3026,[21]14.5496,[22]14.8258,[23]14.9447,[24]15.1835,[25]15.4916,[26]15.8437,[27]15.8826,[28]15.9795,[29]15.9858,[30]16.1339,[31]16.3708,[32]16.5637,[33]16.6604,[34]16.7734,[35]16.9935,[36]17.1459,[37]17.3533,[38]17.4052,[39]17.2951,[40]17.3925,
+10.23.280.986 I Final estimate: PPL = 17.3925 +/- 0.54170
+
diff --git a/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_en_qw_Q4_K_M.log b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_en_qw_Q4_K_M.log
new file mode 100644
index 0000000..a841cc1
--- /dev/null
+++ b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_en_qw_Q4_K_M.log
@@ -0,0 +1,15 @@
+0.00.005.980 I common_init_result: fitting params to device memory ...
+0.00.006.058 I common_init_result: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)
+0.00.319.842 I common_params_fit_impl: projected to use 933 MiB of host memory vs. 7966 MiB of total host memory
+0.00.454.316 W load: control-looking token: 128247 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden
+0.01.525.746 W llama_context: n_ctx_seq (512) < n_ctx_train (40960) -- the full capacity of the model will not be utilized
+0.01.735.449 I common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
+0.01.841.249 I 
+0.01.841.389 I system_info: n_threads = 2 (n_threads_batch = 2) / 2 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.01.841.415 I perplexity: tokenizing the input ..
+0.01.968.239 I perplexity: tokenization took 126.811 ms
+0.01.968.371 I perplexity: calculating perplexity over 40 chunks, n_ctx=512, batch_size=2048, n_seq=4
+0.44.489.155 I perplexity: 42.51 seconds per pass - ETA 7.08 minutes
+[1]1.2807,[2]1.2073,[3]1.2019,[4]2.4242,[5]3.9383,[6]5.5435,[7]6.4929,[8]7.5356,[9]8.8143,[10]9.6176,[11]10.3637,[12]10.8460,[13]11.3835,[14]11.8438,[15]12.2102,[16]12.7862,[17]12.8783,[18]13.1283,[19]13.2189,[20]13.6077,[21]13.8379,[22]14.0765,[23]14.2590,[24]14.4675,[25]14.8062,[26]15.1023,[27]15.1396,[28]15.2048,[29]15.2614,[30]15.4023,[31]15.6229,[32]15.8127,[33]15.9126,[34]16.0220,[35]16.2142,[36]16.3451,[37]16.5386,[38]16.5759,[39]16.4472,[40]16.5360,
+6.29.001.061 I Final estimate: PPL = 16.5360 +/- 0.51957
+
diff --git a/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_en_qw_Q6_K.log b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_en_qw_Q6_K.log
new file mode 100644
index 0000000..57badbd
--- /dev/null
+++ b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_en_qw_Q6_K.log
@@ -0,0 +1,15 @@
+0.00.005.708 I common_init_result: fitting params to device memory ...
+0.00.005.717 I common_init_result: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)
+0.00.281.765 I common_params_fit_impl: projected to use 1875 MiB of host memory vs. 7966 MiB of total host memory
+0.00.419.030 W load: control-looking token: 128247 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden
+0.04.501.798 W llama_context: n_ctx_seq (512) < n_ctx_train (40960) -- the full capacity of the model will not be utilized
+0.04.711.770 I common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
+0.04.904.057 I 
+0.04.904.189 I system_info: n_threads = 2 (n_threads_batch = 2) / 2 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.04.904.216 I perplexity: tokenizing the input ..
+0.05.031.261 I perplexity: tokenization took 127.034 ms
+0.05.031.396 I perplexity: calculating perplexity over 40 chunks, n_ctx=512, batch_size=2048, n_seq=4
+1.05.528.973 I perplexity: 60.50 seconds per pass - ETA 10.07 minutes
+[1]1.3221,[2]1.2156,[3]1.2091,[4]2.4076,[5]3.8648,[6]5.4173,[7]6.3893,[8]7.4358,[9]8.7093,[10]9.5275,[11]10.2680,[12]10.7834,[13]11.3434,[14]11.7564,[15]12.1566,[16]12.6679,[17]12.7695,[18]13.0057,[19]13.1143,[20]13.4655,[21]13.7422,[22]13.9933,[23]14.1292,[24]14.3416,[25]14.6543,[26]14.9505,[27]14.9973,[28]15.0478,[29]15.1173,[30]15.2599,[31]15.4867,[32]15.6835,[33]15.7761,[34]15.8849,[35]16.0718,[36]16.2029,[37]16.4057,[38]16.4251,[39]16.3080,[40]16.3794,
+9.50.945.033 I Final estimate: PPL = 16.3794 +/- 0.51749
+
diff --git a/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_ru_v2_qw_Q2_K.log b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_ru_v2_qw_Q2_K.log
new file mode 100644
index 0000000..3373ac8
--- /dev/null
+++ b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_ru_v2_qw_Q2_K.log
@@ -0,0 +1,15 @@
+0.00.017.305 I common_init_result: fitting params to device memory ...
+0.00.017.314 I common_init_result: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)
+0.00.258.977 I common_params_fit_impl: projected to use 989 MiB of host memory vs. 7966 MiB of total host memory
+0.00.391.207 W load: control-looking token: 128247 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden
+0.00.950.417 W llama_context: n_ctx_seq (512) < n_ctx_train (40960) -- the full capacity of the model will not be utilized
+0.01.099.426 I common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
+0.01.224.860 I 
+0.01.225.023 I system_info: n_threads = 2 (n_threads_batch = 2) / 2 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.01.225.040 I perplexity: tokenizing the input ..
+0.01.335.973 I perplexity: tokenization took 110.921 ms
+0.01.336.081 I perplexity: calculating perplexity over 40 chunks, n_ctx=512, batch_size=2048, n_seq=4
+0.57.513.770 I perplexity: 56.18 seconds per pass - ETA 9.35 minutes
+[1]5.7593,[2]7.0099,[3]5.6658,[4]3.9953,[5]4.6085,[6]4.2157,[7]4.4242,[8]4.5202,[9]4.9586,[10]5.3786,[11]5.7448,[12]5.8946,[13]6.1087,[14]5.7003,[15]5.8407,[16]5.9131,[17]5.8204,[18]5.7731,[19]5.8882,[20]5.8133,[21]5.8126,[22]5.8487,[23]5.8930,[24]6.0328,[25]6.0914,[26]6.2101,[27]6.3779,[28]6.4671,[29]6.5258,[30]6.5104,[31]6.4002,[32]6.4576,[33]6.5241,[34]6.5201,[35]6.6053,[36]6.5163,[37]6.5087,[38]6.5165,[39]6.5659,[40]6.6817,
+9.21.687.896 I Final estimate: PPL = 6.6817 +/- 0.17448
+
diff --git a/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_ru_v2_qw_Q3_K_M.log b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_ru_v2_qw_Q3_K_M.log
new file mode 100644
index 0000000..c6d61e8
--- /dev/null
+++ b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_ru_v2_qw_Q3_K_M.log
@@ -0,0 +1,15 @@
+0.00.003.739 I common_init_result: fitting params to device memory ...
+0.00.003.746 I common_init_result: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)
+0.00.244.434 I common_params_fit_impl: projected to use 1144 MiB of host memory vs. 7966 MiB of total host memory
+0.00.374.035 W load: control-looking token: 128247 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden
+0.00.763.853 W llama_context: n_ctx_seq (512) < n_ctx_train (40960) -- the full capacity of the model will not be utilized
+0.00.887.574 I common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
+0.01.015.174 I 
+0.01.015.311 I system_info: n_threads = 2 (n_threads_batch = 2) / 2 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.01.015.326 I perplexity: tokenizing the input ..
+0.01.126.594 I perplexity: tokenization took 111.255 ms
+0.01.126.709 I perplexity: calculating perplexity over 40 chunks, n_ctx=512, batch_size=2048, n_seq=4
+1.03.691.661 I perplexity: 62.56 seconds per pass - ETA 10.42 minutes
+[1]5.2847,[2]5.9525,[3]4.5901,[4]3.3383,[5]3.8921,[6]3.5475,[7]3.6859,[8]3.8374,[9]4.2189,[10]4.5717,[11]4.8645,[12]5.0500,[13]5.2605,[14]4.9679,[15]5.0756,[16]5.0884,[17]5.0010,[18]4.9582,[19]5.0646,[20]5.0285,[21]5.0216,[22]5.0705,[23]5.1111,[24]5.2322,[25]5.2878,[26]5.3682,[27]5.5046,[28]5.5814,[29]5.6152,[30]5.5984,[31]5.5116,[32]5.5633,[33]5.5914,[34]5.5739,[35]5.6363,[36]5.5669,[37]5.5699,[38]5.5886,[39]5.6422,[40]5.7318,
+10.20.277.870 I Final estimate: PPL = 5.7318 +/- 0.15185
+
diff --git a/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_ru_v2_qw_Q4_K_M.log b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_ru_v2_qw_Q4_K_M.log
new file mode 100644
index 0000000..1a22369
--- /dev/null
+++ b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_ru_v2_qw_Q4_K_M.log
@@ -0,0 +1,15 @@
+0.00.075.669 I common_init_result: fitting params to device memory ...
+0.00.075.680 I common_init_result: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)
+0.00.350.103 I common_params_fit_impl: projected to use 933 MiB of host memory vs. 7966 MiB of total host memory
+0.00.551.633 W load: control-looking token: 128247 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden
+0.01.378.903 W llama_context: n_ctx_seq (512) < n_ctx_train (40960) -- the full capacity of the model will not be utilized
+0.01.505.614 I common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
+0.01.639.229 I 
+0.01.639.363 I system_info: n_threads = 2 (n_threads_batch = 2) / 2 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.01.639.385 I perplexity: tokenizing the input ..
+0.01.751.230 I perplexity: tokenization took 111.839 ms
+0.01.751.350 I perplexity: calculating perplexity over 40 chunks, n_ctx=512, batch_size=2048, n_seq=4
+0.38.412.422 I perplexity: 36.66 seconds per pass - ETA 6.10 minutes
+[1]5.0775,[2]5.8146,[3]4.5768,[4]3.3121,[5]3.8384,[6]3.5377,[7]3.6217,[8]3.7787,[9]4.1173,[10]4.4203,[11]4.7558,[12]4.9015,[13]5.0941,[14]4.8162,[15]4.9004,[16]4.9133,[17]4.8375,[18]4.7915,[19]4.8786,[20]4.8408,[21]4.8221,[22]4.8640,[23]4.8925,[24]5.0043,[25]5.0693,[26]5.1452,[27]5.2752,[28]5.3296,[29]5.3649,[30]5.3579,[31]5.2730,[32]5.3300,[33]5.3565,[34]5.3334,[35]5.4056,[36]5.3357,[37]5.3360,[38]5.3587,[39]5.4293,[40]5.5113,
+6.21.972.850 I Final estimate: PPL = 5.5113 +/- 0.14769
+
diff --git a/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_ru_v2_qw_Q6_K.log b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_ru_v2_qw_Q6_K.log
new file mode 100644
index 0000000..6c38ac5
--- /dev/null
+++ b/paper3-rossiya30-troica/measurements/ternary_ppl_ru_2026-08/ppl_ru_v2_qw_Q6_K.log
@@ -0,0 +1,15 @@
+0.00.003.531 I common_init_result: fitting params to device memory ...
+0.00.003.539 I common_init_result: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)
+0.00.259.681 I common_params_fit_impl: projected to use 1875 MiB of host memory vs. 7966 MiB of total host memory
+0.00.398.999 W load: control-looking token: 128247 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden
+0.00.521.582 W llama_context: n_ctx_seq (512) < n_ctx_train (40960) -- the full capacity of the model will not be utilized
+0.00.722.682 I common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
+0.00.989.382 I 
+0.00.989.510 I system_info: n_threads = 2 (n_threads_batch = 2) / 2 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.00.989.529 I perplexity: tokenizing the input ..
+0.01.102.166 I perplexity: tokenization took 112.624 ms
+0.01.102.287 I perplexity: calculating perplexity over 40 chunks, n_ctx=512, batch_size=2048, n_seq=4
+0.59.334.045 I perplexity: 58.23 seconds per pass - ETA 9.70 minutes
+[1]5.0378,[2]5.7728,[3]4.5637,[4]3.3256,[5]3.8415,[6]3.5352,[7]3.6393,[8]3.7932,[9]4.1101,[10]4.3906,[11]4.6985,[12]4.8277,[13]5.0294,[14]4.7488,[15]4.8322,[16]4.8369,[17]4.7676,[18]4.7279,[19]4.8172,[20]4.7776,[21]4.7604,[22]4.7950,[23]4.8178,[24]4.9332,[25]4.9883,[26]5.0631,[27]5.1909,[28]5.2506,[29]5.2874,[30]5.2752,[31]5.1938,[32]5.2573,[33]5.2874,[34]5.2674,[35]5.3409,[36]5.2732,[37]5.2818,[38]5.3053,[39]5.3749,[40]5.4614,
+9.43.837.164 I Final estimate: PPL = 5.4614 +/- 0.14672
+