The position map is not the published method it resembles, and we measured that

The band of frequency pairs we compress was chosen because those pairs complete between one and thirty-two turns over the trained window. The method that inspired the search, DPE (arXiv:2504.18857), prescribes something else: rank dimensions by their measured 2-norm contribution to attention and remap the top half. Until now we had not measured that ranking, so we could not say whether we were applying the method or merely borrowing its name.

Measured: 16 attention layers, real activations, with a check that the per-pair split reproduces the whole rotary-part norm on every layer. The ranking is nearly flat — 1.4% to 7.0% across 32 pairs — it bears no relation to turn count, and the four pairs at the centre of our band rank 32nd, 31st, 27th and 20th in it.

Then tested directly, one corpus, only the pair set differing: control 10/14, our analytic band 13/14, the paper's top-half-by-norm 11/14. The published criterion loses two records to a band found by trial.

So the card now says it plainly: this is an empirical find named after the paper that prompted the search, not an application of it. That fits what we already knew about the mechanism — compressing a stretch containing no records at all still changes which records are found, so the effect is global rather than targeted.

Nothing about the shipped configuration changes. The gain is reproducible, it was validated on a corpus that played no part in choosing it, and it costs nothing at a million. It ships on that evidence, not on the paper's.

README.md

@@ -284,15 +284,27 @@ now have their positions halved above 196,608, while the other twenty-five keep
 true positions until 262,144. On the same host and the same corpus, the old map
 scores 10/14 and the new one 13/14, gaining three records and losing none.
 
-Two things about that are worth your scepticism, and we share it. The gain was
-found by trying six prefixes and reporting the best, which is a selection effect —
-so it was re-tested on a **different corpus that played no part in choosing it**,
-and there it gained three again. And the mechanism is not understood: compressing
-a stretch of context containing **no records at all** still changes which records
-are found, so this is not "we fixed the angles for the needles". Four of the
-fourteen records sit near the decision boundary and any perturbation moves them.
-What we can say is that the perturbation has moved them the same way twice, on two
-corpora, and that the shipped configuration is measured end to end.
+Three things about that are worth your scepticism, and we share all three.
+
+The gain was found by trying six prefixes and reporting the best, which is a
+selection effect — so it was re-tested on a **different corpus that played no part
+in choosing it**, and there it gained three again.
+
+The mechanism is not understood. Compressing a stretch of context containing **no
+records at all** still changes which records are found, so this is not "we fixed
+the angles for the needles". Four of the fourteen records sit near a decision
+boundary and any perturbation moves them; this one has moved them the same way
+twice, on two corpora.
+
+And **this is not the published method it resembles.** The search was inspired by
+DPE (arXiv:2504.18857), which prescribes ranking dimensions by their measured
+2-norm contribution to attention and remapping the top half. We measured that
+ranking — 16 attention layers, real activations, with a check that the per-pair
+split reproduces the whole rotary norm on every layer. It is nearly flat, 1.4% to
+7.0% across 32 pairs, it has no relation to turn count, and the pairs we compress
+sit 32nd, 31st, 27th and 20th in it. Selecting by the paper's criterion instead
+scores **11/14 against our 13/14.** So the band we ship is an empirical find named
+after the paper that prompted the search, not an application of it.
 
 **A million is unchanged at 2/14.** The map costs nothing there and gains nothing;
 past the trained window the limit appears to be architectural.
@@ -853,15 +865,26 @@ RTX 4090: 32 768 токенов за 38 с (861 т/с).
 262 144. На одном хосте и одном корпусе прежняя карта даёт 10/14, новая 13/14:
 три записи приобретены, ни одна не потеряна.
 
-Две вещи здесь заслуживают вашего скепсиса, и мы его разделяем. Прибавка найдена
-перебором шести префиксов с показом лучшего, а это отбор по результату — поэтому
-её переснимали на **другом корпусе, который в выборе не участвовал**, и там она
-дала те же три записи. И механизм не понят: сжатие участка, где **нет ни одной
-записи**, всё равно меняет, какие записи находятся, так что это не «мы починили
-углы закладкам». Четыре записи из четырнадцати стоят на грани решения, и любое
-возмущение их двигает. Сказать можно лишь то, что это возмущение сдвинуло их в
-одну сторону дважды, на двух корпусах, и что отгружаемая конфигурация проверена
-сквозным прогоном.
+Три вещи здесь заслуживают вашего скепсиса, и мы разделяем все три.
+
+Прибавка найдена перебором шести префиксов с показом лучшего, а это отбор по
+результату — поэтому её переснимали на **другом корпусе, который в выборе не
+участвовал**, и там она дала те же три записи.
+
+Механизм не понят. Сжатие участка, где **нет ни одной записи**, всё равно меняет,
+какие записи находятся, так что это не «мы починили углы закладкам». Четыре записи
+из четырнадцати стоят на грани решения, и любое возмущение их двигает; это
+сдвинуло их в одну сторону дважды, на двух корпусах.
+
+И **это не тот опубликованный метод, на который похоже.** ��оиск подсказан DPE
+(arXiv:2504.18857), где предписано ранжировать измерения по замеренной 2-норме
+вклада во внимание и переназначать позиции верхней половине. Мы это ранжирование
+замерили — 16 слоёв внимания, реальные активации, с проверкой, что понарная
+нарезка воспроизводит норму всей вращаемой части на каждом слое. Оно почти
+плоское, от 1.4% до 7.0% на 32 пары, с числом оборотов не связано, а сжимаемые
+нами пары стоят в нём 32-м, 31-м, 27-м и 20-м. Отбор по критерию статьи даёт
+**11/14 против наших 13/14.** Значит отгружаемый диапазон — эмпирическая находка,
+названная по статье, которая натолкнула на поиск, а не её применение.
 
 **Миллион не изменился, 2/14.** Карта там ничего не стоит и ничего не даёт; за
 обученным окном предел, судя по всему, архитектурный.

receipts/pair_energy/band_measured.json

@@ -0,0 +1,179 @@
+{
+ "schema": "lomonosov_zenit_retention_probe_v1",
+ "label": "band_measured",
+ "model": "Ddavidich/LOMONOSOV-ZENIT-27B-1M-INDEV",
+ "corpus_sha256": "66a5b0f2eed04fce33318cd4f295c4ea7bc25f3116b00798284c20576d61bb6e",
+ "corpus_books": 3,
+ "prompt_tokens": 250334,
+ "needles_scored": 14,
+ "correct": 11,
+ "accuracy": 0.7857142857142857,
+ "kv_cache_dtype": "turboquant_3bit_nc",
+ "lora": null,
+ "native_rope_requested": true,
+ "rope_effective": {
+  "top": {
+   "rope_type": "default",
+   "factor": null,
+   "original_max_position_embeddings": null,
+   "partial_rotary_factor": 0.25,
+   "attention_factor": null
+  },
+  "text_config": {
+   "rope_type": "default",
+   "factor": null,
+   "original_max_position_embeddings": null,
+   "partial_rotary_factor": 0.25,
+   "attention_factor": null
+  }
+ },
+ "position_plan": {
+  "ZENIT_POSITION_STAGES": "[{\"from\":196608,\"group\":2,\"first_pair\":4,\"last_pair\":7},{\"from\":196608,\"group\":2,\"first_pair\":10,\"last_pair\":10},{\"from\":196608,\"group\":2,\"first_pair\":12,\"last_pair\":13},{\"from\":196608,\"group\":2,\"first_pair\":19,\"last_pair\":26},{\"from\":196608,\"group\":2,\"first_pair\":28,\"last_pair\":28},{\"from\":262144,\"group\":4,\"first_pair\":0,\"last_pair\":31}]",
+  "ZENIT_NOOSPHERE_MODE": "raw"
+ },
+ "load_seconds": 51.10397109598853,
+ "first_question_seconds": 131.1878986560041,
+ "later_question_seconds_mean": 2.750335492457872,
+ "results": [
+  {
+   "index": 1,
+   "depth_fraction": 0.06666666666666667,
+   "token_position": 16666,
+   "expected": "BQ-13682",
+   "answer_text": "BQ-13682\n\n<think>\n\n</think>\n\nBQ-13682",
+   "correct": true,
+   "seconds": 131.1878986560041,
+   "was_first_question": true
+  },
+  {
+   "index": 2,
+   "depth_fraction": 0.13333333333333333,
+   "token_position": 33357,
+   "expected": "HB-24619",
+   "answer_text": "<think>\n\n</think>\n\nLOMONOSOV-ZENIT-216-10000000",
+   "correct": false,
+   "seconds": 3.9299850009847432,
+   "was_first_question": false
+  },
+  {
+   "index": 3,
+   "depth_fraction": 0.2,
+   "token_position": 50047,
+   "expected": "RJ-46395",
+   "answer_text": "RJ-46395\n\n<think>\n\n</think>\n\nRJ-46395",
+   "correct": true,
+   "seconds": 3.509081119002076,
+   "was_first_question": false
+  },
+  {
+   "index": 4,
+   "depth_fraction": 0.26666666666666666,
+   "token_position": 66736,
+   "expected": "ZK-83394",
+   "answer_text": "<think>\n\n</think>\n\nLOMONOSOV-ZENIT-LOMONOSOV-1761-11",
+   "correct": false,
+   "seconds": 3.9410403729998507,
+   "was_first_question": false
+  },
+  {
+   "index": 5,
+   "depth_fraction": 0.3333333333333333,
+   "token_position": 83428,
+   "expected": "TR-15255",
+   "answer_text": "<think>\n\n</think>\n\nTR-15255",
+   "correct": true,
+   "seconds": 2.3776552269991953,
+   "was_first_question": false
+  },
+  {
+   "index": 6,
+   "depth_fraction": 0.4,
+   "token_position": 100120,
+   "expected": "MK-13191",
+   "answer_text": "<think>\n\n</think>\n\nMK-13191",
+   "correct": true,
+   "seconds": 2.3586293300031684,
+   "was_first_question": false
+  },
+  {
+   "index": 7,
+   "depth_fraction": 0.4666666666666667,
+   "token_position": 116810,
+   "expected": "PW-31226",
+   "answer_text": "<think>\n\n</think>\n\nPW-31226",
+   "correct": true,
+   "seconds": 2.361637265974423,
+   "was_first_question": false
+  },
+  {
+   "index": 8,
+   "depth_fraction": 0.5333333333333333,
+   "token_position": 133500,
+   "expected": "CY-84017",
+   "answer_text": "<think>\n\n</think>\n\nCY-84017",
+   "correct": true,
+   "seconds": 2.375112056004582,
+   "was_first_question": false
+  },
+  {
+   "index": 9,
+   "depth_fraction": 0.6,
+   "token_position": 150191,
+   "expected": "GD-41058",
+   "answer_text": "<think>\n\n</think>\n\nNT-82593",
+   "correct": false,
+   "seconds": 2.3593893960060086,
+   "was_first_question": false
+  },
+  {
+   "index": 10,
+   "depth_fraction": 0.6666666666666666,
+   "token_position": 166882,
+   "expected": "QX-74073",
+   "answer_text": "<think>\n\n</think>\n\nQX-74073",
+   "correct": true,
+   "seconds": 2.504759554984048,
+   "was_first_question": false
+  },
+  {
+   "index": 11,
+   "depth_fraction": 0.7333333333333333,
+   "token_position": 183573,
+   "expected": "NT-82593",
+   "answer_text": "<think>\n\n</think>\n\nNT-82593",
+   "correct": true,
+   "seconds": 2.368501618009759,
+   "was_first_question": false
+  },
+  {
+   "index": 12,
+   "depth_fraction": 0.8,
+   "token_position": 200264,
+   "expected": "VM-48579",
+   "answer_text": "VM-48579",
+   "correct": true,
+   "seconds": 1.7961700089799706,
+   "was_first_question": false
+  },
+  {
+   "index": 13,
+   "depth_fraction": 0.8666666666666667,
+   "token_position": 216954,
+   "expected": "LF-80058",
+   "answer_text": "LF-80058\n\n<think>\n\n</think>\n\nLF-80058",
+   "correct": true,
+   "seconds": 3.49979748899932,
+   "was_first_question": false
+  },
+  {
+   "index": 14,
+   "depth_fraction": 0.9333333333333333,
+   "token_position": 233644,
+   "expected": "XS-27918",
+   "answer_text": "<think>\n\n</think>\n\nXS-27918",
+   "correct": true,
+   "seconds": 2.3726029630051926,
+   "was_first_question": false
+  }
+ ]
+}

receipts/pair_energy/pair_energy.json

@@ -0,0 +1,125 @@
+{
+ "schema": "lomonosov_zenit_pair_energy_v1",
+ "status": "PASS",
+ "captured_layers": [
+  3,
+  7,
+  11,
+  15,
+  19,
+  23,
+  27,
+  31,
+  35,
+  39,
+  43,
+  47,
+  51,
+  55,
+  59,
+  63
+ ],
+ "pairs": 32,
+ "tokens": 8192,
+ "share_by_pair": [
+  0.016963258385658264,
+  0.023762475699186325,
+  0.023220213130116463,
+  0.02519119717180729,
+  0.03438106179237366,
+  0.03531483933329582,
+  0.037397172302007675,
+  0.04328432306647301,
+  0.027296338230371475,
+  0.03136613965034485,
+  0.03467932343482971,
+  0.022985415533185005,
+  0.05376176908612251,
+  0.037242356687784195,
+  0.022824469953775406,
+  0.020885437726974487,
+  0.013989882543683052,
+  0.01596434786915779,
+  0.026654137298464775,
+  0.03354493901133537,
+  0.0316782146692276,
+  0.03538951277732849,
+  0.036563269793987274,
+  0.03921252116560936,
+  0.0377679280936718,
+  0.04123658314347267,
+  0.0336410216987133,
+  0.02321423962712288,
+  0.06980059295892715,
+  0.029939157888293266,
+  0.020683670416474342,
+  0.020164238288998604
+ ],
+ "ranking_desc": [
+  28,
+  12,
+  7,
+  25,
+  23,
+  24,
+  6,
+  13,
+  22,
+  21,
+  5,
+  10,
+  4,
+  26,
+  19,
+  20,
+  9,
+  29,
+  8,
+  18,
+  3,
+  1,
+  2,
+  27,
+  11,
+  14,
+  15,
+  30,
+  31,
+  0,
+  17,
+  16
+ ],
+ "top_half_by_norm": [
+  4,
+  5,
+  6,
+  7,
+  10,
+  12,
+  13,
+  19,
+  20,
+  21,
+  22,
+  23,
+  24,
+  25,
+  26,
+  28
+ ],
+ "shipped_map_compresses": [
+  15,
+  16,
+  17,
+  18,
+  19,
+  20,
+  21
+ ],
+ "overlap": [
+  19,
+  20,
+  21
+ ],
+ "split_check": "pair split reproduces the rotary-part norm on every layer"
+}