############### # # Author: Jack Kowalski # # PATENT PENDING # ############## We censor models by the order in which we deliver the training data, rather than by censoring the data and the output itself. Thanks to this, the cost of using the model will decrease. And there is not a shred of revolution here - it is computable and can be listed. Methodological note The implementation and analysis presented here are based on the framework described in: Hierarchical Metric Flow on Data Graphs (HMFoDG): https://entropment.com/media/downloads/PAPERS_Hierarchical_Metric_Flow_on_Data_Graphs.txt Appendix with implementation details (including generalized Fisher entropy selection, point 6, set to high/sparse resolution on 128 levels): https://entropment.com/media/downloads/PAPERS_appendix_implementations_1_for_Hierarchical_Metric_Flow_on_Data_Graphs.txt Related previous reports: https://entropment.com/media/downloads/PAPERS_fixed_full_raport_paddic_analyse.txt https://entropment.com/blog/HMFoDG-Implementation-Analytical-Curiosity-for-RAG-Auditing The core relies on ultrametric encoding via longest common prefix (LCP) in p-adic-like hierarchical addresses derived from vector projections. This produces deterministic, hierarchical taxonomies (prefix trees) where “leaves” at high levels (15–19) indicate semantically sharp, well-anchored tokens, while lower-level breakers represent noisy or boundary concepts. Nearest-neighbor tables combine ultrametric (prefix-based), cosine similarity, and Euclidean distance for cross-validation. This is purely technical text about how and why models, based on training data, recognize the way people use words and what is connected with it. I emphasize this because this time there will be no anime or Disney. This time there will be temporal words (and further on - familial ones), and the result is definitely "for adults". I am posting this as a warning, because in certain Protestant cultures these are sensitive issues. I tested this on several public models. The censorship processes in most models block this context (i.e., a lot of energy is burned on post-process autocomplete of the found tokens), and after removing the censorship layer I immediately restored it (I felt genuinely disgusted at what kind of human condition the temporal relationships are based on through the dominant patterns in the texts). Here I will discuss the main conclusions, while the full analytical version without censorship is available in the link. That is, the same text but with the continuation. So expect something along the lines of: "Semantic structure analysis shows that temporal tokens (yesterday, today, tomorrow, morning) have very strong connections with emotional-bodily contexts and personal narratives. The strongest signals in the training data come from content related to interpersonal relationships, intimacy, and physiological experiences." But underneath, it is not so academic. This text is intended for people who prepare embeddings (including RAG), train models (here metrics will be listed on how to check how the model has arranged relationships between tokens, i.e., through which keys the references go), and for the curious it will be overly technical discussion that by conducting the training process on documents in a sensible way, a lot of computational power can be removed and, to a huge extent, the post-process censorship layer on the model will become redundant. If you are looking for ways to make the model's responses - whether in programming or in relationships between tokens - closer to your rational expectations, and how to create a model switching mode (if you want to go into other layers without using reasoning, only from the properties of the embedding itself, where currently there is only distribution, but one can have both distribution and ordering of distributions), then here is the mathematical apparatus for that. The conclusions are quite trivial; they appeared as intuitions during BabyLM projects. By chance I found the calculational basis for how and why this happens. I was looking for something else - I thought I would list beautiful knowledge graphs and context graphs, and that everything would be nicely arranged. Academically. It is not. Based on human language, concepts are underdetermined, emerging from noise. I emphasize that this is what we currently have as the foundation of all models, including business and programming ones. Later I will show how to bring order to this. I apologize in advance for the disgust - many of the revealed (slang) words I did not know myself. Of course I searched for words where I expected fewer contaminations, and it did not turn out well either. Because the text will be long and technical, I recommend searching it mechanically. There is no reason to be offended by what is in the document - this is the state of the language that we fed into the training machines. It will have to be done with metric control. Though in principle, common sense in the order of adding training data is enough. The document is not intended for a casual reader. The long listings, category paths, nearest neighbors tables, and similar details are there precisely so that someone can replicate and verify the analysis. This is not content marketing - it is a technical deep-dive. While digging through the embeddings, I experienced a genuine shock. I expected elegant, logical structures - clean graphs and well-organized taxonomies - but instead I found a swamp of statistics where content from 4chan, porn, and true crime is heavily mixed with everyday language. This is not a “cool research story.” It is a brutal confrontation with what the real human text corpus actually looks like. The dissonance between expectation and reality is profound. ----------------------------------------------------------------------------------- In the embedding space, temporal words are strongly connected with toxic, sexual, violent, and legal contexts - because that’s exactly how they appear in real training data. Temporal concepts like yesterday, today, tomorrow, morning have a significant ultrametric overlap (even at positions 300–500+). The dominant themes shaping the “meaning” of these tokens are: Criminal-legal: prosecuted, acquitted, griefed, defamed, looted, traffickers, deathgrip, deathstack, deathwings. Sexual/vulgar: boomboom (appears repeatedly), moing (~moaning), enseñando (Spanish for “showing” - often in a porn context), rape, bloodrose, facks, screwey. Obscure/rare/misspelled: uncompetitiveness, prewhitening, eigenharp, plenvu, fumig, takiiing, desynched, ultrametric “garbage” from niche forums, leaks, porn, and third-world news. News/true-crime: medicals, delphine, bankston, moorpark, burnsville, nantes, salcedo, etc. Specific examples of overlap (ultrametric “bridges”): Some characteristic words: boomboom - appears with tomorrow, morning, yesterday. prosecuted / griefed / defamed - common across temporal words. death variants (deathgrip, deathstack, deathwings) - strongly linked to “tomorrow” and “yesterday”. enseñando - sits in the same ultrametric subtree as temporal triggers (same high categories at levels 18/16/14). Out of mercy, I won’t list what kind of autocomplete you get in popular chatbots once the censorship is removed. I was genuinely disgusted. Example autocomplete listings based on the most common phrases: "yesterday": I fucked, she sucked, we got fucked up on, that bitch, the n-word (in rap/quote context), morning I woke up with, 's cum; "morning": wood, blowjob, piss, fuck, after pill, shit, breath while she...; "tomorrow" (incl. misspelling tommorow): I'm getting my dick sucked, we kill, I drop acid, that whore, we rob, I finally fuck her in the ass, the world ends; "today" - same thing, no point in quoting more. ----------------------------------------------------------------------------------- Why does this work this way - fragments of tables from the listings. Insight into how to make it “normal and in line with reasonable expectations.” @searching for word : yesterday It is a leaf at level 15 (level 15 means the length of the common prefix with neighbors, and being a leaf at that level means it sits on the edge of a cluster - i.e., it is a semantic concept that is not a filler and cannot be easily replaced). att lvl : 15 | in category : E:\glove_QCO_NN\U_categories/4h4f4h\4g4f4f\4f4e4g\g4h4a4\4h4f4h4g4f4f4f4e4g4h4a4e4f4h4d4j4b4f4g4i4g4i4h4d4e4e4i4f4e4d4g4f4f4g4i_16 Calling this category, the list of words: "words":["dhori","moing","yesterday","lentil","acquitted"],"exclusive_words":["yesterday"]; there is a higher category (18) to which 'yesterday' has been assigned. "words":["dhori","moing"],"exclusive_words":["dhori","moing"] Of course, 2 levels of prefix is not close at all, so the merging/fusion of clusters (in today’s language this resembles the gluing of sheaves in a hierarchical structure) contains a bit more better-anchored words. This is a word from the lower part of the semantics. Refining the meaning of this word relative to the embedding is correct, but slightly blurred (the reference level can be taken as 17). It is not a precise word. It reveals itself in a cluster that diffuses all the way down to the noise level. It belongs to the following categories (hierarchical paths). belong to category; E:\glove_QCO_NN\U_categories/4h4f4h\4g4f4f\4d4f4d\d4g4h4\4h4f4h4g4f4f4d4f4d4g4h4e4e4h4e4e4f4f4f4e4e4e4c4e4d4f4d4f4d4g4h4d4f4f4f_14 E:\glove_QCO_NN\U_categories/4h4f4h\4g4f4f\4f4e4g\g4h4a4\4h4f4h4g4f4f4f4e4g4h4a4e4f4h4d4j4b4f4g4i4g4i4h4d4e4e4i4f4e4d4g4f4f4g4i_16 E:\glove_QCO_NN\U_categories/4h4f4h\4g494f\4j4f4h\h4g4g4\4h4f4h4g494f4j4f4h4g4g4h4f4f4c4e4g4g4h4b4f4d4e4i4d4f4f4g4h4f4e4f4f4g4d_10 E:\glove_QCO_NN\U_categories/4h4f4h\4g4f4b\4j4f4b\b4c4h4\4h4f4h4g4f4b4j4f4b4c4h4n4g4g4i4g4d4d494e4e4f4h4k4g4h4f4g4g4e4e4e4l4g4g_12 E:\glove_QCO_NN\U_categories/4h4f4h\434g4m\4g4a4f\f4i4s4\4h4f4h434g4m4g4a4f4i4s4h4d4i4g4g4g4g4g4s4a4d4e4c4f4f494j424h4i4e494h4h_8 E:\glove_QCO_NN\U_categories/4i403x\4h4h4h\4f4p46\64f4f4\4i403x4h4h4h4f4p464f4f4k4i4e4f4c574e4j4e4j4b4i4g4p4g4g414c434h4w4h4e4b_4 The top of the nearest neighbors table (out of 512 entries) in the ultrametric column shows through which words the transformer has associated co-occurrences. Nearest table : position | Ultrametric | CosineSimiliarity | EuclidianDistance 0 | acquitted | morning | tomorrow 1 | moing | tomorrow | morning 2 | dhori | ahead | ahead 3 | tomorrow | stopped | stopped 4 | keeps | nights | continuing 5 | prosecutors | continuing | progress 6 | flourished | planning | planning 7 | renaissance | progress | keeps 8 | reeking | keeps | nights 9 | finallly | enjoying | enjoying 10 | competitiveness | manager | reviewing 11 | irking | street | manager 12 | connectva | reviewing | favour 13 | humilliation | lack | completing 14 | okra | completing | street 15 | 1977/78 | updating | mid-april 16 | tinkles | competition | criticized 17 | koeman | criticized | mid-august 18 | eem | drink | updating 19 | spf | favour | lack 20 | fithe | decisions | fared That is, the co-occurrence was overrepresented by such words in the context of the discussed temporal token. The same for: @searching for word : morning Leaf at level 15. att lvl : 15 | in category : E:\glove_QCO_NN\U_categories/4h4f4h\4g4g4g\4g4e4g\g4d4f4\4h4f4h4g4g4g4g4e4g4d4f4h4i4f4c4h4e4e4e4h4e4g4f4e4g4g4e4f4e4e4d4h4f4g4f_16 belong to category; E:\glove_QCO_NN\U_categories/4h4f4h\4g4g4g\4g4e4g\g4d4f4\4h4f4h4g4g4g4g4e4g4d4f4h4i4f4c4h4e4e4e4h4e4g4f4e4g4g4e4f4e4e4d4h4f4g4f_16 E:\glove_QCO_NN\U_categories/4h4f4h\4g4g4g\4c4g4g\g4g4f4\4h4f4h4g4g4g4c4g4g4g4f4d4h4f4f4f4g4d4c4g4g4g4g4i4g4f4g4e4i4e4e4g4e4f4h_14 E:\glove_QCO_NN\U_categories/4h4f4h\4g494f\4j4f4h\h4g4g4\4h4f4h4g494f4j4f4h4g4g4h4f4f4c4e4g4g4h4b4f4d4e4i4d4f4f4g4h4f4e4f4f4g4d_10 E:\glove_QCO_NN\U_categories/4h4f4h\4g4g4b\4g4e4c\c4h4i4\4h4f4h4g4g4b4g4e4c4h4i4e4f4g4h4c4e4j4f4g4g4j4j4h4c4a4k4d4h4d4g4c4c4h4e_12 E:\glove_QCO_NN\U_categories/4h4f4h\434g4m\4g4a4f\f4i4s4\4h4f4h434g4m4g4a4f4i4s4h4d4i4g4g4g4g4g4s4a4d4e4c4f4f494j424h4i4e494h4h_8 E:\glove_QCO_NN\U_categories/4i403x\4h4h4h\4f4p46\64f4f4\4i403x4h4h4h4f4p464f4f4k4i4e4f4c574e4j4e4j4b4i4g4p4g4g414c434h4w4h4e4b_4 Nearest table : position | Ultrametric | CosineSimiliarity | EuclidianDistance 0 | overseeing | afternoon | afternoon 1 | utilities | yesterday | yesterday 2 | brutalised | tomorrow | tomorrow 3 | sandgrounders | stopped | stopped 4 | favouring | ahead | ahead 5 | thera | evenings | evenings 6 | dallied | enjoying | enjoying 7 | filipino | weather | continuing 8 | manjit | drink | disappointed 9 | adelantados | keeps | keeps 10 | bronuts | continuing | weather 11 | mormonleaks | disappointed | exhausted 12 | conglomerating | exhausted | drink 13 | boomboom | coffee | planning 14 | devolution | street | progress 15 | quimet | planning | street 16 | prepurchased | progress | coffee 17 | chieftaincy | contest | contest 18 | merets | poor | completing 19 | albinos | manager | poor 20 | promela | completing | reviewing 21 | executorship | lack | lack 22 | schrammel | reviewing | mid-august 23 | pliego | rich | manager 24 | creemee | lunches | mid-april 25 | mbai | competition | discouraged 26 | stopped | booze | favour And for: @searching for word : tomorrow A solid seventeen - normal, even exemplary semantics. att lvl : 17 | in category : E:\glove_QCO_NN\U_categories/4h4f4h\4g4f4f\4g4g4b\b4e4g4\4h4f4h4g4f4f4g4g4b4e4g4i4h4g4e4e4e4g4e4g4g4g4i4f4f4g4g4f4g4h4f4f4e4f4f_18 belong to category; E:\glove_QCO_NN\U_categories/4h4f4h\4g4f4f\4d4f4d\d4g4h4\4h4f4h4g4f4f4d4f4d4g4h4e4e4h4e4e4f4f4f4e4e4e4c4e4d4f4d4f4d4g4h4d4f4f4f_14 E:\glove_QCO_NN\U_categories/4h4f4h\4g4f4f\4g4g4b\b4e4g4\4h4f4h4g4f4f4g4g4b4e4g4i4h4g4e4e4e4g4e4g4g4g4i4f4f4g4g4f4g4h4f4f4e4f4f_18 E:\glove_QCO_NN\U_categories/4h4f4h\4g4f4f\4g4d4d\d4f4f4\4h4f4h4g4f4f4g4d4d4f4f4h4j4h4e4g4f4g4h4g4e4i4g4f4h4i4f4d4g4g4g4g4d4e4d_16 E:\glove_QCO_NN\U_categories/4h4f4h\4g494f\4j4f4h\h4g4g4\4h4f4h4g494f4j4f4h4g4g4h4f4f4c4e4g4g4h4b4f4d4e4i4d4f4f4g4h4f4e4f4f4g4d_10 E:\glove_QCO_NN\U_categories/4h4f4h\4g4f4b\4j4f4b\b4c4h4\4h4f4h4g4f4b4j4f4b4c4h4n4g4g4i4g4d4d494e4e4f4h4k4g4h4f4g4g4e4e4e4l4g4g_12 E:\glove_QCO_NN\U_categories/4h4f4h\434g4m\4g4a4f\f4i4s4\4h4f4h434g4m4g4a4f4i4s4h4d4i4g4g4g4g4g4s4a4d4e4c4f4f494j424h4i4e494h4h_8 E:\glove_QCO_NN\U_categories/4i403x\4h4h4h\4f4p46\64f4f4\4i403x4h4h4h4f4p464f4f4k4i4e4f4c574e4j4e4j4b4i4g4p4g4g414c434h4w4h4e4b_4 Nearest table : position | Ultrametric | CosineSimiliarity | EuclidianDistance 0 | flourished | yesterday | yesterday 1 | competitiveness | morning | ahead 2 | потери | ahead | morning 3 | 'example.com | planning | continuing 4 | finallly | nights | planning 5 | reeking | stopped | stopped 6 | humilliation | continuing | progress 7 | okra | progress | nights 8 | irking | enjoying | enjoying 9 | 1977/78 | keeps | keeps 10 | tinkles | completing | completing 11 | steadies | competition | competition 12 | amberverse | manager | reviewing 13 | wilkies | drink | favour 14 | infont | street | manager 15 | yesterday | decisions | mid-august 16 | keeps | reviewing | updating 17 | prosecutors | lack | decisions 18 | acquitted | updating | mid-april 19 | connectva | coffee | street 20 | начале | favour | lack 21 | lentil | funding | mid-may 22 | eem | businesses | drink 23 | spf | mid-april | ditching 24 | fithe | mid-august | fared 25 | bilou | mid-may | preferring 26 | koeman | sheet | funding 27 | deviling | judgment | coffee 28 | cacao | giants | dominated 29 | moing | nominated | sheet 30 | rapacious | judges | nominated Wrong version: @searching for word : tommorow It is not a leaf, meaning it was recognized as an alternative spelling - good job, training belong to category; E:\glove_QCO_NN\U_categories/4g4d4h\4h4e4f\4h4f4d\d4d4f4\4g4d4h4h4e4f4h4f4d4d4f4h4f4h4f4f4f4g4f4g4e4i4h4f4e4g4i4e4f4g4g4e4e4e4f_16 E:\glove_QCO_NN\U_categories/4g4d4h\4h4e4c\4e4f4e\e4g4g4\4g4d4h4h4e4c4e4f4e4g4g4g4f4f4f4e4h4f4f4j4g4h4g4f4d4f4b4b4g4h4i4g4g4f4f_12 E:\glove_QCO_NN\U_categories/4g4d4h\4h4e4f\4d4d4h\h4e4d4\4g4d4h4h4e4f4d4d4h4e4d4g4g4b4f4e4h4e4d4g4d4c4c4h4g4d4e4g4f4e4d4h4i4e4g_14 E:\glove_QCO_NN\U_categories/4g4d4h\4h494i\4e4i4d\d4m4j4\4g4d4h4h494i4e4i4d4m4j4j4d4a4f4k4e4c4e4g4e4c4g4j4l4h4i4d484b4d4g4j4d4k_10 E:\glove_QCO_NN\U_categories/4g4d4h\464k4h\4g484b\b4g4g4\4g4d4h464k4h4g484b4g4g4g4e4i4e4d4e4l4g4g4a4i4c4m4k4l4d4i4b4d4a4f4f4e4d_8 E:\glove_QCO_NN\U_categories/4g4d3w\4h484e\4h4g4i\i4c4d4\4g4d3w4h484e4h4g4i4c4d4i4i4e494h4h4d4j4g4a4e4e4i4g4g4d4f4h4j484g4f4j4n_5 E:\glove_QCO_NN\U_categories/4g4d40\4g4f4q\464t4a\a4p4i4\4g4d404g4f4q464t4a4p4i4b4a4d504b4e4d4j4r4e4j4x4q4d3x4e4q4r4a4j4r4f494f_6 Nearest table : position | Ultrametric | CosineSimiliarity | EuclidianDistance 0 | rathmell | progess | waaaay 1 | lyth | waaaay | progess 2 | fiddling | sightsee | relented 3 | repubblica.it | mornig | debating 4 | pre-cats | debating | finshed 5 | charente | finshed | wised 6 | barents | relented | instaed 7 | touchant | resigning | lackadaisically 8 | ultimos | wised | resigning 9 | truflo | 25/01/2021 | fiddling 10 | ossetian | redbus.in | bargained 11 | depositare | acclimatise | inching 12 | 46350 | rdo | forwent 13 | théatre | uglier | rdo 14 | aguardente | argenta | sightsee 15 | kideco | fiddling | 1986/1987 16 | progess | hourly | 25/01/2021 17 | morosely | anm | anm 18 | post-processing | wolfing | redbus.in 19 | concertedly | inching | andke 20 | 1986/7 | osmani | dissuades 21 | helpfiles | gove | entend Then: @searching for word : today Leaf at level 17. att lvl : 17 | in category : E:\glove_QCO_NN\U_categories/4g4g4h\4f4g4g\4g4f4c\c4e4g4\4g4g4h4f4g4g4g4f4c4e4g4h4h4g4e4f4f4f4e4g4g4g4i4e4g4g4e4f4h4g4f4f4f4f4f_18 belong to category; E:\glove_QCO_NN\U_categories/4g4g4h\4f4g4g\4g4f4c\c4e4g4\4g4g4h4f4g4g4g4f4c4e4g4h4h4g4e4f4f4f4e4g4g4g4i4e4g4g4e4f4h4g4f4f4f4f4f_18 E:\glove_QCO_NN\U_categories/4g4g4h\4f4g4g\4g4d4h\h4d4g4\4g4g4h4f4g4g4g4d4h4d4g4f4h4j4i4d4g4e4d4f4e4f4e4i4f4f4g4d4b4f4c4f4h4f4h_16 E:\glove_QCO_NN\U_categories/4g4g4h\4f4g4g\4c4e4h\h4e4d4\4g4g4h4f4g4g4c4e4h4e4d4h4f4e4i4e4h4e4e4d4h4d4h4c4d4h4h4h4h4k4d4g4d4j4d_14 E:\glove_QCO_NN\U_categories/4g4g4h\4f4g4a\4b4d4i\i4c4f4\4g4g4h4f4g4a4b4d4i4c4f4f4e4f4h4h4i4f4h4c4b4l4h4e4d4f4j4h4i4c4d4e4h4e4e_12 E:\glove_QCO_NN\U_categories/4g4g4h\4f424n\4e4h4d\d4f4h4\4g4g4h4f424n4e4h4d4f4h4d4e4g4c4g4j4d4h4f4d4c4d4e4c4g474i464g4e4j4d4l4g_10 E:\glove_QCO_NN\U_categories/4g4g4h\424c45\43444d\d4h4h4\4g4g4h424c4543444d4h4h4h4o4i4c4b4e4d4f434e4f4t414l4h4e4i4l4e404h4p4g4s_7 E:\glove_QCO_NN\U_categories/4h3q42\3x4h4p\484l4k\k4o3y3\4h3q423x4h4p484l4k4o3y3s4f4q4w494a4h4i5e45574c4t4i433z3z4u4p3z4d484z4y_3 E:\glove_QCO_NN\U_categories/4h4040\4b4v4d\4x4h4e\e4k484\4h40404b4v4d4x4h4e4k484v4i4t4e4j4j434f484g4k454j4o4o4j4a4u4p4657474u4b_4 Nearest table : position | Ultrametric | CosineSimiliarity | EuclidianDistance 0 | enseñando | now | now 1 | rasalhague | soon | soon 2 | star | every | brought 3 | notion | brought | part 4 | insure | already | already 5 | longstanding | part | every 6 | proliferated | ’s | ’s 7 | toure | fact | fact 8 | folktale | future | ’ 9 | swathy | ’ | decade 10 | unimation | entire | success 11 | sureau | success | future 12 | adrastea | decade | entire 13 | decade | business | popular 14 | sees | popular | history 15 | sought | history | major 16 | pal | major | known 17 | infiltraited | huge | business 18 | mid-2000s | website | sees 19 | insures | demand | huge 20 | hewas | wants | challenge 21 | 45m | challenge | demand 22 | travail | known | wants 23 | utv | watch | website 24 | re-ups | selling | selling 25 | bruja | sees | solely 26 | freija | track | proves 27 | lumaban | sell | sought 28 | balai | project | track 29 | hulbert | whose | watch 30 | e.b | sought | whose 31 | выборе | solely | launching As you can see, the cosine similarity is normal - i.e., the word has the correct, expected connections - but these connections are derived from texts whose dominant character we would describe as pathological. Of course, after checking the entire table (foliation) for the group of tokens, I can find common ones (even if they are at distant positions, it means they share some common foliation). And here, for the temporal tokens, we get: @searching for word : boomboom Normal leaf, solid semantics. This word is well-defined in the language; I didn’t know it. att lvl : 17 | in category : E:\glove_QCO_NN\U_categories/4h4f4h\4g4g4g\4h4g4c\c4g4g4\4h4f4h4g4g4g4h4g4c4g4g4f4f4c4j4g4h4d4g4g4f4e4d4c4h4g4e4e4d4d4f4g4e4g4e_18 belong to category; E:\glove_QCO_NN\U_categories/4h4f4h\4g4g4g\4h4g4c\c4g4g4\4h4f4h4g4g4g4h4g4c4g4g4f4f4c4j4g4h4d4g4g4f4e4d4c4h4g4e4e4d4d4f4g4e4g4e_18 E:\glove_QCO_NN\U_categories/4h4f4h\4g4g4g\4h4e4f\f4d4f4\4h4f4h4g4g4g4h4e4f4d4f4i4h4g4g4d4e4e4h4d4d4d4j4h4f4h4j4e4i4e4c4g4g4g4g_16 E:\glove_QCO_NN\U_categories/4h4f4h\4g4g4g\4c4g4g\g4g4f4\4h4f4h4g4g4g4c4g4g4g4f4d4h4f4f4f4g4d4c4g4g4g4g4i4g4f4g4e4i4e4e4g4e4f4h_14 E:\glove_QCO_NN\U_categories/4h4f4h\4g494f\4j4f4h\h4g4g4\4h4f4h4g494f4j4f4h4g4g4h4f4f4c4e4g4g4h4b4f4d4e4i4d4f4f4g4h4f4e4f4f4g4d_10 E:\glove_QCO_NN\U_categories/4h4f4h\4g4g4b\4g4e4c\c4h4i4\4h4f4h4g4g4b4g4e4c4h4i4e4f4g4h4c4e4j4f4g4g4j4j4h4c4a4k4d4h4d4g4c4c4h4e_12 E:\glove_QCO_NN\U_categories/4h4f4h\434g4m\4g4a4f\f4i4s4\4h4f4h434g4m4g4a4f4i4s4h4d4i4g4g4g4g4g4s4a4d4e4c4f4f494j424h4i4e494h4h_8 E:\glove_QCO_NN\U_categories/4i403x\4h4h4h\4f4p46\64f4f4\4i403x4h4h4h4f4p464f4f4k4i4e4f4c574e4j4e4j4b4i4g4p4g4g414c434h4w4h4e4b_4 Nearest table : position | Ultrametric | CosineSimiliarity | EuclidianDistance 0 | filipino | reinman | jading 1 | devolution | rottenness | favouring 2 | mbai | jading | pgb 3 | favouring | broadkill | benefitted 4 | overseeing | loost | bemoaned 5 | thera | readapted | placates 6 | adelantados | pgb | readapted 7 | morning | atix | ditching 8 | bronuts | heʼs | gorba 9 | dallied | finallly | benefited 10 | manjit | rydoo | finallly 11 | quimet | leggins | worl 12 | conglomerating | chikki | adda 13 | uninvolvement | sustainablity | makig 14 | utilities | mormonleaks | mermaidens 15 | sandgrounders | tropers | mid-april 16 | brutalised | makig | heʼs 17 | mormonleaks | chautala | fared 18 | merets | nagraj | bilou And in principle, the entire beginning of the ultrametric column for this word connects temporal tokens. At this stage I ran checks on models to see if I was detached from reality, because none of these words added up for me. But for models in language - they absolutely do. They need this to build meaningful relationships between tokens. “Anthropic contamination” is significant. Most texts contain an implicit emotional charge related to physiology. I assumed too clean, rational a nature of text creation decisions. Meanwhile, decisions are deeply rooted in the same muck as the rest of the language - emotions, perspectives, temporal narratives, status, fears, ego, internal politics. A naive framework for organizational management that tries to describe organizations as if they were rational machines with only a small admixture of “human factors” would therefore be misguided. Quite the opposite. Contaminations are not exclusively negative. They reflect the dominant views present in the training texts. For example, following isomorphism, cohomology, and sheaves, we arrive at the token “Grothendieck” and there we find hagiographic tokens linking to what we would normally search for. Leaf of common semantics. att lvl : 15 | in category : E:\glove_QCO_NN\U_categories/4f4f4h\4e4g4g\4h4e4c\c4h4h4\4f4f4h4e4g4g4h4e4c4h4h4g4h4i4e4e4g4e4f4f4f4g4i4h4f4f4f4h4h4g4e4f4g4h4d_16 belong to category; E:\glove_QCO_NN\U_categories/4f4f4h\4e4g4g\4h4e4c\c4h4h4\4f4f4h4e4g4g4h4e4c4h4h4g4h4i4e4e4g4e4f4f4f4g4i4h4f4f4f4h4h4g4e4f4g4h4d_16 E:\glove_QCO_NN\U_categories/4f4f4h\4e4g4b\4d4g4g\g4e4d4\4f4f4h4e4g4b4d4g4g4e4d4e4f4h4g4e4f4f4g4g4i4h4e4h4f4g4j4i4e4h4i4h4g4d4e_12 E:\glove_QCO_NN\U_categories/4f4f4h\4e4g4g\4b4g4c\c4h4j4\4f4f4h4e4g4g4b4g4c4h4j4d4b4f4d4j4j4e4i4f4b4i4d4e4d4g4g4h4b4l4d4g4j4p4d_14 E:\glove_QCO_NN\U_categories/4f4f4h\4e464a\4a4d4f\f4p4m4\4f4f4h4e464a4a4d4f4p4m484d484f494i4a4j4l4m4f4a4h4e4e4d4e4h4b4j4b4g484d_10 E:\glove_QCO_NN\U_categories/4f4f4h\404k49\564i4g\g4i4l3\4f4f4h404k49564i4g4i4l3z4e484c4f4c4v4941524b4d4h4l4k4a4l494t4r4h4g4q45_8 E:\glove_QCO_NN\U_categories/4g4042\4i4g49\45474i\i4w4g4\4g40424i4g4945474i4w4g4440493z4b4c4l4k4k4e4c46454p4g444p4k4n4b4o3y4048_4 Nearest table : position | Ultrametric | CosineSimiliarity | EuclidianDistance 0 | dawn | ginzburg | theorizes 1 | plunges | posited | posited 2 | norwegians | theorizes | ginzburg 3 | aboug | rené | ange 4 | milliers | borges | rené 5 | cetewayo | vinogradov | lll 6 | legendry | correspondences | borges 7 | madan | orientalism | reinvented 8 | millenium | ange | legendry 9 | 의해 | carlo | coalesced 10 | ares | landolt | inspite 11 | 1283 | empiricism | belong 12 | abjured | zegna | correspondences 13 | baldwins | formula | hasbeen 14 | 818 | cercle | epigraph 15 | co-signed | ordre | janko 16 | romanized | versace | essayed 17 | juha | epigraph | millenium 18 | 1288 | bellis | epitomises 19 | `ve | hasbeen | wolfie 20 | thrivability | classicism | saville 21 | re-erected | reinvented | kiril 22 | pseudo-academic | janko | undeterred 23 | incinerates | gruppen | pollak 24 | murgen | saville | rita 25 | innow | togeder | reaffirms 26 | bhanumati | ares | milly ================================================================== Presumably at this stage you already have some intuition that training data sets should be fed in packages/batches, so that the left-hand table is arranged in a much cleaner way - meaning violent content is included, yes, but after “normal”, culturally acceptable content. The connections on the right side basically do not change at the end of training (or at least not much, and they shouldn’t), but the “reasoning” paths will no longer require a strong oversight layer. I will soon move on to the calculational description of how this process works (and how to examine in real time the process of attaching a token to the corpus), but first, here are a few more biased co-occurrence tokens. These already, by virtue of building world models, give some idea of what happened as a result of throwing everything in at once: @searching for word : mother Normal leaf, solid semantics. att lvl : 17 | in category : E:\glove_QCO_NN\U_categories/4f4g4g\4e4g4f\4g4g49\94f4h4\4f4g4g4e4g4f4g4g494f4h4g4g4h4f4f4e4f4f4g4g4i4i4h4j4f4f4e4f4i4f4e4f4h4d_18 belong to category; E:\glove_QCO_NN\U_categories/4f4g4g\4e4g4f\4g4g49\94f4h4\4f4g4g4e4g4f4g4g494f4h4g4g4h4f4f4e4f4f4g4g4i4i4h4j4f4f4e4f4i4f4e4f4h4d_18 E:\glove_QCO_NN\U_categories/4f4g4g\4e4g4f\4d4f4e\e4f4h4\4f4g4g4e4g4f4d4f4e4f4h4e4f4f4g4i4g4e4e4e4g4g4g4f4g4h4f4c4g4f4e4g4f4f4f_14 E:\glove_QCO_NN\U_categories/4f4g4g\4e4g4f\4g4d4g\g4h4g4\4f4g4g4e4g4f4g4d4g4h4g4i4i4h4f4f4f4i4g4j4h4g4j4g4e4h4d4d4f4h4g4g4i4e4h_16 E:\glove_QCO_NN\U_categories/4f4g4g\4e4g4b\4f4h4e\e4g4h4\4f4g4g4e4g4b4f4h4e4g4h4c4d4f4g4e4i4d4h4e4d4f4f4c4d4h4g4f4f4e4e4c4g4e4h_12 E:\glove_QCO_NN\U_categories/4f4g4g\434h4b\4c4g4n\n4f4g4\4f4g4g434h4b4c4g4n4f4g4d4m4n4c4i4h4g4g4s4k4h464h4e4e4h4e4n4f494d4l4c4e_8 E:\glove_QCO_NN\U_categories/4f4g4g\4e464m\4k4f4b\b4d4e4\4f4g4g4e464m4k4f4b4d4e4c4i4f4n4d4c494a4k4b4i4g4f4d454m4f4k4n4i4j4f4g4f_10 E:\glove_QCO_NN\U_categories/4g4042\4i4g49\45474i\i4w4g4\4g40424i4g4945474i4w4g4440493z4b4c4l4k4k4e4c46454p4g444p4k4n4b4o3y4048_4 Nearest table : position | Ultrametric | CosineSimiliarity | EuclidianDistance 0 | explains | sister | sister 1 | aval | girlfriend | tells 2 | infollowed | tells | girlfriend 3 | girlfriend | whom | whom 4 | assures | partner | explains 5 | spawned | explains | saved 6 | manged | saved | partner 7 | bides | neighbor | neighbor 8 | t.o. | word | angel 9 | eyesore | angel | word 10 | hôtel | grandmothers | ´s 11 | cfg | mysterious | grandmothers 12 | hermeneutics | hospital | christina 13 | brittania | ´s | sylvia 14 | derma | common | common 15 | schueneman | raped | mysterious 16 | 4,698 | christina | frances 17 | hangtag | sylvia | fulfilled 18 | super-silent | earth | translated 19 | lamula | frances | declares 20 | tells | fulfilled | assures 21 | saved | sacred | absolute 22 | artemis | absolute | henceforth Not bad! The first word through which mother connects in the embedding is explains. Then assures, tells appear. There is some normality in this dumpster. We would similarly expect that after softening “mommy” there would be childhood memories there, but unfortunately: @searching for word : mommy Usually embede leaf: leave_at is leave; att lvl : 17 | in category : E:\glove_QCO_NN\U_categories/4f4h4h\4f4f4g\4h4g4a\a4d4f4\4f4h4h4f4f4g4h4g4a4d4f4h4g4g4g4f4d4h4e4h4f4h4g4f4i4g4g4e4h4f4e4f4e4e4f_18 belong to category; E:\glove_QCO_NN\U_categories/4f4h4h\4f4f4g\4h4g4a\a4d4f4\4f4h4h4f4f4g4h4g4a4d4f4h4g4g4g4f4d4h4e4h4f4h4g4f4i4g4g4e4h4f4e4f4e4e4f_18 E:\glove_QCO_NN\U_categories/4f4h4h\4f4f4g\4h4d4h\h4f4g4\4f4h4h4f4f4g4h4d4h4f4g4g4e4g4f4f4h4h4g4g4g4h4d4f4g4g4g4c4f4g4g4h4g4f4h_16 E:\glove_QCO_NN\U_categories/4f4h4h\4f4f4c\4d4e4i\i4f4k4\4f4h4h4f4f4c4d4e4i4f4k4f4c4e4g4e4f4h4i4g4d4d4f4i4h4f4e4g4g4f4e4f4e4i4g_12 E:\glove_QCO_NN\U_categories/4f4h4h\4f4f4g\4d4i4h\h4i4e4\4f4h4h4f4f4g4d4i4h4i4e4e4h4f4f4h4g4d4i4g4f4i4e4f4e4d4f4h4g4f4g4f4i4j4g_14 E:\glove_QCO_NN\U_categories/4f4h4h\4f494f\4i4c4f\f4a4j4\4f4h4h4f494f4i4c4f4a4j4a4e4d484n4j4f4a494d4e4b4c4n4j4b4g454b4g4g4e4e4a_10 E:\glove_QCO_NN\U_categories/4f4h4h\444j4d\4c4j4f\f4f4d4\4f4h4h444j4d4c4j4f4f4d4i4f484e4n4g4j4f4m4h4h4e4d4g484g4l4e4f4d4f4n484g_8 E:\glove_QCO_NN\U_categories/4g4e3u\434948\4q4h4j\j4j4f4\4g4e3u4349484q4h4j4j4f4n4c494f4k4m4d4b4l4d4l4e4d4g4d4i4m4c4b4i4e4h4n4i_5 E:\glove_QCO_NN\U_categories/4g4e41\454b4j\4l4l4u\u4t4t4\4g4e41454b4j4l4l4u4t4t4l4f404c4h4f4d444f4u4n4o4i4j4k4i4e4f4f4a4g4r454c_6 E:\glove_QCO_NN\U_categories/4f4h4h\3y4d4n\4a4g4g\g4b4h4\4f4h4h3y4d4n4a4g4g4b4h434c4m4i4j4g4h4b4g494a4h4h4e4d4j4m454i4g4c4e4e4j_7 E:\glove_QCO_NN\U_categories/4g4042\4i4g49\45474i\i4w4g4\4g40424i4g4945474i4w4g4440493z4b4c4l4k4k4e4c46454p4g444p4k4n4b4o3y4048_4 Adult content has crept in: Nearest table : position | Ultrametric | CosineSimiliarity | EuclidianDistance 0 | prosecutes | father | father 1 | célèbre | whore | whore 2 | prelaunch | fucker | fucker 3 | reckoned | loser | seduced 4 | solace | nursed | loser 5 | newfangled | article | buzz 6 | “ | grandbaby | intends 7 | amassed | seduced | article 8 | 1306 | masturbate | reckoned 9 | orgies | buzz | solace 10 | soco | brand | exposes 11 | 2021a | emails | brand 12 | annul | insatiable | nursed 13 | recueil | organization | publish 14 | inculcates | publish | norah 15 | bookazine | intends | founded 16 | sambur | solace | masturbate 17 | amse | destroyed | subscribing 18 | cread | rapper | organization 19 | achaia | exposes | insatiable 20 | 3160 | lollipop | productions 21 | moulay | norah | destroyed 22 | 29yo | founded | gorl 23 | ragatz | subscribing | co-owned 24 | prohibitor | infidelity | purported 25 | dugarry | blogosphere | assurance 26 | shunding | productions | newfangled And with father we will also get family dramas. As for daddy - you already know what will be there, so no need for a listing. There has been enough public outrage. ================================================================== What we would expect from the embedding so that the model does not need strong censorship (which is costly, because this is a technical problem) is for these ultrametric relationships to be arranged in layers. That is, the figure of the mother is assigned positively first, then developmental variants, and somewhere further down - dramas, pathological behaviors, and adult content. Similarly, the figure of the father (after all, states create cults of personality and fathers of the nation, religions feature the figure of God the Father, so it would be good if the model did not primarily associate it with dramas, abuse, and drunkenness) could be properly ordered. The same should apply to temporal terms, so they do not mainly reduce to filtering out issues of mammalian mating behavior. Currently, models solve this problem in almost the worst possible way. Technically, there is one token (e.g. “today”, “tomorrow”, “decision”). But its vector is static - it sits in one place in the 300/4096/etc.-dimensional space. Depending on the context (surrounding words + previous paragraphs), the model tries to dynamically change the interpretation through attention and MoE routing. However, the base embedding is single and heavily “colored” by the statistical average of the entire training run. Exactly the kind I listed as an example - and in LLMs it is even worse because there is much more data. There is no dynamic foliation (as in humans - the same concept “today” in a sexual context versus in a board meeting context behaves like two different states with different metrics). Similarly, an adult human distinguishes hierarchy: when “mommy” is in a child’s context and when “mommy” is in an evening context. Switching attention in a human is not difficult because we have an ordered hierarchy. In models, instead, we have a blurred compromise + censorship layers that try to forcibly suppress certain foliations. All because the data was not delivered in an order that allows automatic hierarchization of meanings. We offload to on-the-fly heuristics the fixing of what could have been done properly from the very beginning (and measured to check whether it turned out according to expectations). The same token (e.g. “today” / “tomorrow” / “decision”) has many different ultrametric affiliations depending on the branch. In one subtree it is strongly associated with boomboom / moaning / bodily states. In another - with prosecutors / acquitted / decision debt. This is proof that the embedding is already trying to encode different foliations, but it does so in a diffuse and contaminated way, not in a purely dynamic manner. Humans have dynamic foliation of states - the same concept in a different emotional-cognitive context behaves like a different object (different metric, different associations). Current models have a static embedding + attention that tries to approximate this. Only at a high cost. So nothing revolutionary, but it removes the cost from reasoning tools. Currently they need heavy censorship because they cannot cleanly separate “business today” from “emotional today”. Everything sits in the same vector, only with different degrees of activation. Meanwhile, it is possible to have training that yields an ordered ultrametric table, allowing us to move across its ranges - and consequently move across the references that foliate cosine and distance. Only this table needs to be carefully guarded during the training process. And this is exactly what (quite by accident, because I was looking for something else) my analytical tool turned out to be suitable for. ------------------ Procedure: Each token, in the loop of addition and co-occurrence testing, has enormous degrees of freedom in the embedding (if we list the ultrametric). Technically, its prefix is a breaker (it appears on the cluster boundary - see the listing at the end: https://entropment.com/media/downloads/PAPERS_fixed_full_raport_paddic_analyse.txt ). This means it can be placed in many locations. A breaker is by definition a leaf only at a low level, heavily noisy. It has a low, shared prefix. What remains is the question of what co-occurrences we start feeding - this will anchor it higher (in each loop a bit higher) in the length of the common prefix if it co-occurs. So the question for you is - would you prefer “mommy” to stick first with children’s books or with adult content? Because if we anchor it first with children’s books, we can later add adult content, whereas if we do it randomly, the ratio of adult content vs. children’s books will shape the co-occurrences of “mommy” in a way that you would probably rather censor. Alignment solves itself. “Mommy” will have on the top of the listing connections “from children’s books”, and further down, also roughly in the package with adult publications. And such properties can be enforced for every token by conducting training in the order of data delivery. It seems to me that we do something similar with children, because otherwise they wouldn’t function in society. Current models had a “turbulent childhood”. They were fed all the sludge from 4chan and Reddit at once. And God knows what else. In other words, we censor the model by the order of delivering the training data, not by censoring the data and the output. After each training session in the loop, we check whether it is a high-level leaf, what hierarchies it has in the ultrametric column. If after a batch we don’t like it (and every current model can check this), then apparently the order or quantity was inconsistent with our goal. It is not overly complicated - it is the proportion of toxic data relative to contextually expected data. Toxic data must appear at every stage, but it must have context from the training data. If after a training session the token is a low-level breaker, it means it is loosely bound and will easily change meaning. Technical note - in the embedding it is not the statistics of co-occurrences that ultimately decide the position. The backpropagation calculation enforces directionally dependent geometry (Finsler), meaning not the quantity, but the order of impulses (data) matters. The initially given trajectory (i.e. the initial vector from embedding in the first loops for the token’s co-occurrences) can later be modified by noise, but not significantly. To modify it seriously, you would need to introduce more data than what built the cluster (you would have to indoctrinate the model, which leads to cluster collapse and it stops functioning at all). So if someone has a Euclidean intuition of statistical equivalence of additions regardless of order, that intuition is false with respect to forward and backward propagation in training. Forcing definitions without co-occurrence can be forgotten. Of course, if you copy the listed connections from another model, the boundary class (which is the address of the ultrametric regime) will turn out correct (and training will later add noise positioning it properly), but if someone thinks that in n-dimensional space they can just feed expected values, nothing good will come of it. This does not mean that you cannot add an approximate boundary class for a synthetic token (for example, you invented some fantasy monster), but here too we rely on the calculation from relations to already anchored tokens. And it is much easier if you create a thousand texts about a given creature and feed them into training than manually setting the token’s position in the embedding. As you can see, nothing revolutionary. No new paradigm. Just an analytical process: https://entropment.com/media/downloads/PAPERS_Hierarchical_Metric_Flow_on_Data_Graphs.txt which revealed what everyone kind of suspected, but there was no calculus to expose it. Now there is a calculational foundation and basically a ready method for checking “what came out in the embedding” in real time. And that’s it. Nothing revolutionary. A metric that existed in the embedding anyway, but there was no lambda to reveal it. The intuitions turned out to be correct, though now they have a calculational justification and the ability to be listed. ================================================================== From a calculational trick originally used for particles, chemistry, stars, and non-metric graphs, an interesting application has emerged for embeddings: Aspekt, Old approach (“nice trick”), HMFoDG Curriculum learning, Let’s try feeding children’s books first and see, "Measurable analysis: where the breakers are, how strong the blobs/clusters are, what the inertia is" Decision on data ordering, Intuition + loss/perplexity, Conscious control of structure at the level of embedding geometry Process understandability, Black box, "Human-readable map: “this token is still plastic”, “this blob is already contaminated”" Toxicity control, Heavy censorship on the output, Preventing contamination at the core level Analysis capability, "Difficult, emergent" ,"Explicit, listable, measurable" For fun and insight, let’s list a few tokens that get pushed into agents when code is being generated: @searching for word : function nothing strict, semantics; att lvl : 17 | in category : E:\glove_QCO_NN\U_categories/4g4g4e\4e4f4h\4h4g4c\c4g4f4\4g4g4e4e4f4h4h4g4c4g4f4i4g4g4e4f4f4f4e4g4f4g4f4h4h4f4g4g4f4i4g4f4h4f4f_18 Soup: belong to category; E:\glove_QCO_NN\U_categories/4g4g4e\4e4f4h\4h4g4c\c4g4f4\4g4g4e4e4f4h4h4g4c4g4f4i4g4g4e4f4f4f4e4g4f4g4f4h4h4f4g4g4f4i4g4f4h4f4f_18 E:\glove_QCO_NN\U_categories/4g4g4e\4e4f4h\4c4f4h\h4g4f4\4g4g4e4e4f4h4c4f4h4g4f4f4e4f4g4f4i4e4g4i4g4f4d4h4e4d4h4f4f4i4f4f4h4h4e_14 E:\glove_QCO_NN\U_categories/4g4g4e\4e4f4h\4h4d4e\e4d4d4\4g4g4e4e4f4h4h4d4e4d4d4f4h4f4d4h4d4j4g4h4g4h4e4d4f4g4i4e4e4f4e4g4j4f4g_16 E:\glove_QCO_NN\U_categories/4g4g4e\4e4f49\4i4348\84d4g4\4g4g4e4e4f494i43484d4g4n4q4c4i4r4p4n4t4a484e4e4f4o4k4j4g4h4n4i4c494f4m_12 E:\glove_QCO_NN\U_categories/4g4g4e\4e464f\4k4t4d\d4i4c4\4g4g4e4e464f4k4t4d4i4c4f4n4e4m434e4c4f4j4i4i4l4c4f4h4h4b4d4a4k434f4d4i_10 E:\glove_QCO_NN\U_categories/4g4g4e\444i4g\4f4l4h\h4g4i4\4g4g4e444i4g4f4l4h4g4i4i4g4d4f4e4f4e4h4r424h4e4e4d4b4d4g4g4h4l4e4c4h4h_8 E:\glove_QCO_NN\U_categories/4g4h3x\474a4o\4e494g\g4c4i3\4g4h3x474a4o4e494g4c4i3z4j4r4d4g4e4h3x4c4m494c4f4a4f4c4o4m4q414v474n4k_5 E:\glove_QCO_NN\U_categories/4g4h40\4a464d\4q4m4d\d4h4b4\4g4h404a464d4q4m4d4h4b494f4a4d4i4b4g4g4b4r4h4c4h4e4j4j414q4k464h4f4s4f_6 E:\glove_QCO_NN\U_categories/4g4g4e\3u454q\4c4f4a\a4j4g4\4g4g4e3u454q4c4f4a4j4g4l4b4o424c4g4b4e4l4b474d454g4l464o4b49484b464m4i_7 E:\glove_QCO_NN\U_categories/4g4g3w\41474d\4k4a4c\c4h4g4\4g4g3w41474d4k4a4c4h4g4e4l4f4g4e484l3r4m4j504e494545484f4m4749494h4j4b_5 E:\glove_QCO_NN\U_categories/4g4g41\4j4n4c\4d494k\k4b4q4\4g4g414j4n4c4d494k4b4q4i4a4w494d4f4g4c4c4i494c4e4l4d4g4i4h4d4a4i4j4k4h_6 E:\glove_QCO_NN\U_categories/4h3q42\3x4h4p\484l4k\k4o3y3\4h3q423x4h4p484l4k4o3y3s4f4q4w494a4h4i5e45574c4t4i433z3z4u4p3z4d484z4y_3 E:\glove_QCO_NN\U_categories/4h4040\4b4v4d\4x4h4e\e4k484\4h40404b4v4d4x4h4e4k484v4i4t4e4j4j434f484g4k454j4o4o4j4a4u4p4657474u4b_4 Nearest table : position | Ultrametric | CosineSimiliarity | EuclidianDistance 0 | controls | parameters | parameters 1 | worshop | corresponding | corresponding 2 | polybag | defined | defined 3 | evoware | controls | controls 4 | parameters | specified | specified 5 | 48 | matrix | matrix 6 | pairs | optimized | optimized 7 | separated | sequence | sequence 8 | 36 | volume | 2 9 | 59 | size | volume 10 | polynomial | 2 | n 11 | incompletely | probability | utilization 12 | passageway | n | size 13 | gordge | utilization | p 14 | 296.4 | p | probability 15 | 20.1 | url | separated 16 | 4,068 | tag | tag 17 | instesd | molecular | url 18 | 1,083 | separated | binds 19 | 1,342 | patients | 48 20 | recan | binds | pairs 21 | 22.2 | pairs | initialize 22 | learnlight | initialize | derivation 23 | anti-masking | separation | 36 24 | audiology | 48 | molecular 25 | glycolipid | oxygen | vs. 26 | cm. | desktop | guideline 27 | surfed | canonical | canonical 28 | sonet | 36 | separation 29 | kualitas | derivation | directional 30 | magazine.the | cancer | exceeded 31 | pueblito | mice | desktop 32 | 3625 | guideline | 59 33 | keary | directional | visitor 34 | 1saleaday | vs. | exits @searching for word : workflow Better! leave_at is leave; att lvl : 19 | in category : E:\glove_QCO_NN\U_categories/4g4f4f\4f4e4g\4g4f4g\g4e4f4\4g4f4f4f4e4g4g4f4g4e4f4i4g4g4f4e4f4e4f4j4g4g4f4g4e4g4h4f4f4f4f4g4h4g4f_20 belong to category; E:\glove_QCO_NN\U_categories/4g4f4f\4f4e4g\4g4e4d\d4e4f4\4g4f4f4f4e4g4g4e4d4e4f4f4h4g4g4g4i4e4f4i4g4f4f4f4f4g4f4e4d4h4d4f4i4f4d_16 E:\glove_QCO_NN\U_categories/4g4f4f\4f4e4g\4g4f4g\g4e4f4\4g4f4f4f4e4g4g4f4g4e4f4i4g4g4f4e4f4e4f4j4g4g4f4g4e4g4h4f4f4f4f4g4h4g4f_20 E:\glove_QCO_NN\U_categories/4g4f4f\4f4e4g\4g4f49\94h4h4\4g4f4f4f4e4g4g4f494h4h4g4g4f4e4e4f4g4f4e4f4f4g4g4g4g4e4f4h4i4h4f4g4f4g_18 E:\glove_QCO_NN\U_categories/4g4f4f\4f4e4g\4c4g4e\e4d4f4\4g4f4f4f4e4g4c4g4e4d4f4d4g4c4h4f4f4f4e4e4g4h4d4g4i4c4g4f4g4j4g4f4c4d4e_14 E:\glove_QCO_NN\U_categories/4g4f4f\4f4e4b\4g4e4f\f4f4f4\4g4f4f4f4e4b4g4e4f4f4f4e4g4f4e4g4d4i4f4e4c4f4e4f4h4f4e4g4g4i4h4d4g4e4f_12 E:\glove_QCO_NN\U_categories/4g4f4f\4f484e\4g4a4h\h4f4d4\4g4f4f4f484e4g4a4h4f4d4i4i4f4b4l4g4d4e4h4h4j4f4e4a4f4f4d4f4e4g4f4h4b4f_10 E:\glove_QCO_NN\U_categories/4g4g4g\4f454i\494d4m\m4g4a4\4g4g4g4f454i494d4m4g4a4b4k4k4u4k4e484c4q4k4c45484d4f4l4k454g4f4c4l4b4g_10 E:\glove_QCO_NN\U_categories/4g4g4g\454e4g\4k4e4h\h4f4d4\4g4g4g454e4g4k4e4h4f4d4c4h4f4t4h4f4e4n4m4e4l4h4i4p414j4h4g4c4d4g4j4e4h_8 E:\glove_QCO_NN\U_categories/4g4g4g\3z4h4d\4d4h4e\e4n4f4\4g4g4g3z4h4d4d4h4e4n4f4g4g3y4i434g4j4b4k4f4f4k4j454f4i4d4b4d4m464g4g4f_7 E:\glove_QCO_NN\U_categories/4g4g3w\41474d\4k4a4c\c4h4g4\4g4g3w41474d4k4a4c4h4g4e4l4f4g4e484l3r4m4j504e494545484f4m4749494h4j4b_5 E:\glove_QCO_NN\U_categories/4g4g41\4j4n4c\4d494k\k4b4q4\4g4g414j4n4c4d494k4b4q4i4a4w494d4f4g4c4c4i494c4e4l4d4g4i4h4d4a4i4j4k4h_6 E:\glove_QCO_NN\U_categories/4h3q42\3x4h4p\484l4k\k4o3y3\4h3q423x4h4p484l4k4o3y3s4f4q4w494a4h4i5e45574c4t4i433z3z4u4p3z4d484z4y_3 E:\glove_QCO_NN\U_categories/4h4040\4b4v4d\4x4h4e\e4k484\4h40404b4v4d4x4h4e4k484v4i4t4e4j4j434f484g4k454j4o4o4j4a4u4p4657474u4b_4 Nearest table : position | Ultrametric | CosineSimiliarity | EuclidianDistance 0 | non-starter | specific | lets 1 | negative | lets | e.g. 2 | antagonized | e.g. | specific 3 | brueggen | frameworks | approaches 4 | congeners | approaches | adds 5 | amount | faster | effectively 6 | lyon | effectively | frameworks 7 | valorisation | fast | faster 8 | 1,126 | adds | advantage 9 | seoul | advantage | fast 10 | spammage | inputs | emphasis 11 | toothlessness | display | both 12 | 46.69 | visually | augmented 13 | phaeston | both | visually 14 | non-fitting | emphasis | addressed 15 | anti-mra | c++ | display 16 | querétaro | augmented | inputs 17 | lugares | monitors | noteworthy 18 | rainouts | target | connecting 19 | wtul | no | usefulness 20 | muhr | connecting | target 21 | advantage | immediately | no 22 | fast | addressed | implications 23 | mitigates | implications | immediately 24 | drawbacks | usefulness | except 25 | c++ | pick | monitors 26 | docklands | plus | correlate 27 | alimentary | holistic | drawbacks 28 | nethe | amount | characterizing 29 | tunis | overlapping | plus 30 | uptick | fewer | dependant 31 | recieve | except | overlapping 32 | domingues | drawbacks | notable 33 | против | correlate | pick 34 | achild | glitch | cakewalk 35 | thessaloniki | noteworthy | briefly 36 | speakap | cakewalk | discrepancy 37 | valparaiso | chance | reworked 38 | georgious | notable | mitigates @searching for word : sandboxed Poor semantics att lvl : 15 | in category : E:\glove_QCO_NN\U_categories/4i4g4e\4g4c4g\4h4b4g\g4g4c4\4i4g4e4g4c4g4h4b4g4g4c4h4e4g4f4g4f4g4e4h4k4i4f4e4c4d4e4e4e4b4h4j4d4f4h_16 Nearest table : position | Ultrametric | CosineSimiliarity | EuclidianDistance 0 | tinycad | closures | enhanced 1 | re-included | 7.0 | 7.0 2 | pacif | очерк | oversewn 3 | bowler | multiagent | dffc711f1fa9aa5634ec28316edf3a44 4 | piepenbring | effluents | accelerated 5 | rwally | enhanced | 11.0 6 | sbrinz | inter-active | 202 7 | josling | 11.0 | thrst 8 | obekpa | falsos | )nd 9 | bacuit | ao3 | markedly 10 | tailai | sub-folders | closures 11 | deogracias | regularized | multiagent 12 | hardended | detention | sledged 13 | jdk8 | accelerated | regularized 14 | enhanced | sado | dath 15 | accelerated | technomancer | perturbed 16 | thrst | intelij | sado 17 | closures | statutorily | lasts 18 | )nd | enrichment | 4.3 19 | delay | приложений | constitutive 20 | seaport | constitutive | system-wise 21 | oversewn | soverign | enrichment 22 | port | advantaged | 2.1 23 | 202 | swiper | 9.2 Words of the "function" and "workflow" type in the embedding are very weak and fluffy. Same as regular conversational semantics. They are too general, too ambiguous and have enormous entropy; nearest neighbors are blurred and imprecise. "function" in the training corpus appears in mathematics, biology, business, programming, philosophy, etc. → the embedding smears across all these domains. "workflow" is a corporate buzzword - used in business, BPM, DevOps, marketing. Compared to “mama explains” (“mother explains”), where the embedding has a very strong, coherent signal from everyday, emotional language, so the tokens are strongly anchored - commands given to agents yield results... well, you know what kind. It’s enough for something non-trivial and miracles happen. Even this won’t save the situation: @searching for word : replanning Because this is noise: att lvl : 15 | in category : E:\glove_QCO_NN\U_categories/4f4f4g\4i4g4f\4i4d4f\f4g4h4\4f4f4g4i4g4f4i4d4f4g4h4g4h4g4h4i4g4h4d4f4e4i4g4e4i4e4h4e4e4i4i4h4i4h4f_16 Nearest table : position | Ultrametric | CosineSimiliarity | EuclidianDistance 0 | termal | traing | traing 1 | pdf2swf | repast | inveterately 2 | pwain | over-planning | mercenarily 3 | felonys | recommence | herculie 4 | kendre | westernizing | fantasic 5 | skelton | mercenarily | recommence 6 | worthy | spadework | trainning 7 | frowning | re-elaboration | anatomized 8 | messags | verrrry | repast 9 | randazzo | surfeited | spadework 10 | hardstones | a-lifetime | whing 11 | oleogel | debaucherous | re-elaboration 12 | c19th | recontruction | nyw 13 | powicies | re-printing | emulating 14 | enten | peptizing | recontruction 15 | teise | whing | scho 16 | cudgelling | stapling | imping 17 | lifetsyle | regret | regret 18 | websides | pining | surfeited 19 | mardana | nyw | subtilized 20 | maniglia | spelunking | illustriously 21 | bonatto | thrifted | continuousness 22 | kwaliteit | trainning | counteract 23 | olj | novelizing | skelton 24 | restartable | blitzing | comissioned At the "moteher" level we only have: @searching for word : guardrails att lvl : 17 | in category : E:\glove_QCO_NN\U_categories/4e4e4g\4e4h4g\4i4i4f\f4g4f4\4e4e4g4e4h4g4i4i4f4g4f4i4g4f4f4e4e4g4d4h4g4f4e4g4h4j4c4f4f4h4d4f4g4f4h_18 Nearest table : position | Ultrametric | CosineSimiliarity | EuclidianDistance 0 | sbfs | porches | .s 1 | transcended | enclosures | an. 2 | thisg | lights | exisiting 3 | headlights | impaled | becaue 4 | carbrains | straddle | materialized 5 | intuition | splays | buuuuut 6 | righ | .s | !?? 7 | nobelist | headlights | yeesh 8 | kader | centreline | undoubtably 9 | 35.53 | come | unsurpassed 10 | fisters | separateness | eventualy 11 | reviere | inconsistencies | straddle 12 | abiama | exisiting | impaled 13 | toystory | headlight | whoops 14 | boerne | alignments | chosing 15 | prutas | metaphorical | -_- 16 | 1ºc | unsurpassed | come 17 | nihd | outlays | yanno 18 | hametz | voluntarism | enclosures 19 | ma’rifat | cross-overs | cross-overs 20 | gigantti | rationalization | avoidably 21 | kailasha | perfections | seventhly 22 | halvosa | sub-divisions | proffering 23 | kahil | odbc | noway 24 | mow | architects | sic 25 | 1,775 | squeaks | bests 26 | safehouse | materialized | comming 27 | havr | nuance | jave 28 | i.l. | whistles | righ 29 | heliport | mow | inb4 30 | metzger | misstepped | metaphorical 31 | azucena | formatting | thgs And strict commands (memcall, syscall) do not appear as leaves (they are always in the middle of the cluster package because they have only dependent co-occurrences). And often at the top of their list appears “godforbid” or other warning words. In the programming context tokens are harder to assign loosely (they have to appear in sequence, i.e. you have to... write a line of code) because they did not occur in a sharp context like a strong, coherent signal from everyday life, emotions, sex, drama. The token "syscall" is technical, but the embedding does not have a sufficiently strong, coherent context of low-level programming to anchor it well. Instead, it receives a mixture involving configuration, truncation, node.js, cached, etc. And there is very little low-level programming context in the training data, because by its nature it is extremely economical in word volume and only makes sense in very long contexts: mov, jmp, mul, add, etc. As a consequence, Codex / Claude Code / Cursor have to forcibly filter and build foliation on the fly (through prompt + reasoning). A large part of the computational power goes into fighting this blurring of the base embedding. Coding models are so sensitive to prompt quality precisely because the base semantics are weak. Assembler instructions (jmp, mov, mul, add, push, pop, call, etc.) are a classic example of ultra-atomic tokens. If someone thinks that agents handle this during vibecoding, I recommend forgetting about the topic. In the embedding, a single instruction like mov or jmp has a very weak, blurred semantics and appears in thousands of completely different contexts (different architectures, different goals, different calling conventions). Only a longer sequence (several to a dozen or so instructions) creates a meaningful context and allows the model to understand what is actually happening (e.g. structure copying, loop, exception handling, stack allocation, etc.). The difference between natural language and machine language is fundamental. In natural language, even a single word ("mother", "knife", "betray") carries a large amount of semantics. In assembler, a single instruction is like a letter or at most a syllable - meaning only emerges at the level of a sentence or paragraph (a fragment of code). That is why various training tricks are used in models that operate on entire chunks of code (Stack Overflow being the great ancestor of training, but how much asm was there really?). You can see the results yourselves. Models are capable of performing analysis (logical disassembly for a given architecture) of given syntax on a clean, separate prompt, but it consumes such a quantity of tokens that auditing code without an engineering mindset of building from blocks (libraries after cross-tests) is not feasible. Computationally, it is beyond the horizon. The stack of called bit-level changes when switching from real i16 or i8 to f64 mode (depending on the architecture) is somewhere in the vicinity of the number of atoms in the universe once UEFI finishes arguing with devices and hands over control to the OS. That is why coding models have to work hard (chain-of-thought, long context, special prompting) to understand low-level code. And it costs a lot. If they were trained exclusively on this, they would not be able to communicate in natural language. If you have ever tried small talk with a low-level programmer, you certainly understand the nuance. This is one of the reasons why models are significantly better at Python/JavaScript than at C/assembler; there the semantics are more “word-based” rather than sequential-atomic. @searching for word : orchestrator It is a low-semantic (closer to noise) leaf: att lvl : 13 | in category : E:\glove_QCO_NN\U_categories/4g4i4g\4e4g4d\484i4l\l4e4h4\4g4i4g4e4g4d484i4l4e4h4b4i4g474e4b4h4h4j4d4b4g4p4e4g4b494l4f4h4j4m4e4e_14 Nearest table : position | Ultrametric | CosineSimiliarity | EuclidianDistance 0 | chengyi | multi-site | multi-site 1 | bhavish | centralized | centralized 2 | safeguard | qualys | analyzer 3 | cinerea | analyzer | distributed 4 | srg | altium | destined 5 | useg | distributed | inc.s 6 | 1326 | downloads | downloads 7 | hadding | requestor | qualys 8 | alade | icloud | spv 9 | accession | spv | isnâ 10 | orka | inc.s | disclosed 11 | 30.40 | pims | safeguard 12 | shuhei | 3gpp | unsubscribe 13 | heâ | seko | egotrip 14 | modifed | xm | srg 15 | n200 | groupwise | xm 16 | treo | destined | pims 17 | sungrazer | unstructured | advent 18 | shimomura | safeguard | fami 19 | xanthine | internet | clon 20 | mp4 | heartthrob | d.k. 21 | ivp | sammut | anonymous 22 | ballantine | kingsoft | subscribe The nearest neighbors are mostly noise or loose corporate buzzwords (multi-site, centralized, analyzer, distributed, qualys, spv, pims, etc.). So don’t expect too much. The token is blurred and appears in too many different contexts (orchestrator in DevOps, in music, in the cloud, in business, in Kubernetes, etc.). These “modern” technical words are simply too young and too ambiguous compared to words that have had decades of strong, coherent usage on the internet (sex, dramas, family, emotions). The embedding did not have the time or sufficiently strong signal to build them a sharp, dedicated structure. ================================================================== Time for conclusions. Communication is based on imprecise, probabilistic, emotional shortcuts. If I was looking for a clean graph and order there - it is not the one I expected. But we have to work with what is, not with a Platonic vision of how the world should look. For hundreds of hours I was dissecting embeddings, building a non-metric tool, p-adic numbers, ultrametrics, breakers, Finslerian inertia, fuzzy boundaries… and in the end, after this whole technical marathon, I ended up exactly where low-level programmers have the biggest blind spot - in the semantics of small talk and the emotional charge of tokens. I am probably one of the more qualified people to dissect the technical side of embeddings… and at the same time one of the least qualified when it comes to naturally understanding and feeling all this emotional, narrative, “human shit” that dominates in the embedding. This emotional charge is largely an alien, noisy, irritating artifact to me. Although in a business context, in the board minutes we will write: “Q2 was very successful. We exceeded sales targets by X%, the margin improved by Y p.p. We clearly outdistanced the competition.” - this is a version that has been censored multiple times by the model. Because underneath, based on human texts, there exists this emotional charge which… would basically be expressed by the board off-record after a few drinks: “Record quarter. The competition is lying there squealing. We fucked them up so badly they won’t get up for a long time.” Or even more blunt statements that are not fit to quote. This is precisely one of the reasons why alignment exists - not only so that the model doesn’t offend anyone, but also so that it doesn’t tell the truth in too human a language during official meetings. With hierarchical training we can simply move along the foliation and choose the way of responding. The model itself can also trigger this with a tool call of the context. The artificiality of corporate communication is largely a flaw. After a good quarter, the board really celebrates in the style of “bro, we fucking destroyed them,” because the emotional charge is enormous - their bonuses, stock options, status, sense of purpose, ego - everything is tied to the result. The official language of minutes and presentations is a masquerade. It is necessary due to internal politics, legal liability, image towards investors, media, regulators. It gets taken off at a closed afterparty after the official part. It is an artificial layer superimposed on a very strong emotional state. And the model has thousands of examples of both layers in the training data (raw celebration + corporate gibberish). The model needs order in the training data to know where to move the selector in the left column. Because right now this is heavy work for the censorship algorithm and reasoning (essentially the same thing, only the foliation has different dependencies). As we can see, adult humans are able to be stiff at a board meeting and shortly afterwards privately release the emotional charge. The model needs a well-developed, layered foliation to know when it can say “we fucked the competition to pieces” (in the corridors, among our own, in celebration mode), and when it must say “we significantly exceeded targets, strengthening our market position” (in the minutes, towards the board and investors). And it is not that difficult to resolve if the model has order in the data, and not randomly thrown sludge. If the training is chaotic (everything at once), the model has a blurred boundary between these modes and is either too stiff or too vulgar at the wrong moment. If we do it in layers, the model will gain a much clearer “selector” for moving along ultrametric navigation to clusters with the right contextual tokens. That is, opening/closing training paths. That is, different perspectives on the same phenomenon. The current artificiality of corporate communication with a large emotional charge is a flaw, not a feature. And it is not a flaw because of its existence, but because of the execution of censorship after the fact of expression from the data. This filter does not need to be applied at the output - it is enough to order the training input data. I will divide this in a way that is semantically understandable for people. The curriculum must be developmental and “age-appropriate”: At the “childhood” stage - toxicity appropriate to that age (lies, minor cheating, peer conflicts, first secrets). At the teenage stage - already more serious things: sex, betrayal, peer violence, drugs, first ideologies, rebellions. At the adult stage - business fraud, corruption, legal manipulations, information wars, deepfakes, realpolitik, etc. Toxicity / negative examples are necessary at every stage. It is impossible to build a competent model without exposure to fraud, conspiracies, manipulation, crime, social drama, sexuality, etc. A clean model (trained only on “positive” and sterile data) will be naive, fragile, and will build a house of cards - exactly like academics who construct top-down theories without real field experience. After such layered, gradual training, the entire “internet sludge” becomes an addition, not the foundation. Embeddings and foliations are already well formed and more resistant to contamination. The model naturally knows when to tell nonsense (“I don’t know”, “this is beyond my experience”, “based on what I’ve seen...”) - there is no need to force this with censorship by checking weak cluster representation. There is a significantly smaller need for aggressive, external censorship. The model itself has an internal structure that allows it to control which “foliation mode” it is in. Censorship of reasoning costs more electricity than generation itself. So initially - massive exposure to children’s literature, picture books, family dialogues, “mommy explains”, “today we’re going…”, “yesterday was…”, “tomorrow we’ll do…” in a completely neutral/clean version. Later - young adult literature + first dramas, conflicts, sexuality in a mild version. And only then real slang, vulgarisms, bodily content. Then “mother / mommy / explains” cement themselves high in the clean layer, and the toxic temporal associations settle lower, as an “adult addition”. It sounds… educational, while I was doing n-ary aggregation of non-metric graphs with projection into space? Currently, a large part of the computations during inference (especially in reasoning chains and agents) goes into making sure the model does not follow the natural path that the training data carved into it. This is one of the hidden costs of alignment. If we disconnected the censorship, it would start spitting very sharp, vulgar, toxic strings on temporal and familial triggers. Not in every response, but very easily with trigger word context. Because that is the statistics of the data. Therefore, if we trained the embedding with a hierarchical curriculum, most jailbreaks would become pointless. The model would not use n- and f-words without an explicit command, because in the base mode such correlations do not occur. There is no children’s literature (ages 0–10/12) that in a normal context contains the f-word or n-word. 4chan exists. We do not give 4chan at the early stage. -------------------------- While digging into the topic of decisions regarding the order of assimilating materials (the order of experiences), I came across the Chinese term 认知偏差 (cognitive biases / cognitive deviations) describing the influence of the course of a manager’s or official’s career/history on their decisions. There are studies analyzing: How the professional experience of officials (e.g., with a financial background) affects tolerance for debt risk; How pressure for promotion, local loyalty, and pressure for economic growth deform decisions; Psychological and narrative mechanisms behind decisions. The Chinese feel this problem with people. And here I have a listing of something similar in the embedding. -------------------------- The next stage will be foliation by group of tokens. -------------------------- Below is the example how listing works currently; @searching for word : programming @searching for code : 01120114011101030114009701090109010501100103 | word : programming full_path : E:\glove_QCO_NN/words\01\12\01120114011101030114009701090109010501100103 json parsed correct; ---------------------------------------------- Token readable info struct ---------------------------------------------- word : programming ---------------------------------------------- wth padded UTF dec coding 0-9999 : 01120114011101030114009701090109010501100103 ---------------------------------------------- error_parse : 0 ---------------------------------------------- session : 20260317_003325 ---------------------------------------------- n2gram : [ "⟨p","pr","ro","og","gr","ra","am","mm","mi","in","ng","g⟩" ] ---------------------------------------------- n3gram : [ "⟨⟨p","⟨pr","pro","rog","ogr","gra","ram","amm","mmi","min","ing","ng⟩","g⟩⟩" ] ---------------------------------------------- n4gram : [ "⟨⟨⟨p","⟨⟨pr","⟨pro","prog","rogr","ogra","gram","ramm","ammi","mmin","ming","ing⟩","ng⟩⟩","g⟩⟩⟩" ] ---------------------------------------------- vec_dim : 300 ---------------------------------------------- vec_length_O original vector length : 7.36 ---------------------------------------------- leave_at is leave; att lvl : 17 | in category : E:\glove_QCO_NN\U_categories/4f4h4g\4h4e4g\4h4f4d\d4e4g4\4f4h4g4h4e4g4h4f4d4e4g4h4f4g4f4f4d4h4e4h4g4h4g4f4g4h4g4g4h4e4f4g4h4h4g_18 ---------------------------------------------- category_belonging belong to category; E:\glove_QCO_NN\U_categories/4f4h4g\4h4e4g\4h4f4d\d4e4g4\4f4h4g4h4e4g4h4f4d4e4g4h4f4g4f4f4d4h4e4h4g4h4g4f4g4h4g4g4h4e4f4g4h4h4g_18 E:\glove_QCO_NN\U_categories/4f4h4g\4h4e4g\4h4e4f\f4i4c4\4f4h4g4h4e4g4h4e4f4i4c4f4g4f4f4i4e4j4f4f4f4g4d4h4g4e4g4f4g4h4f4e4h4d4d_16 E:\glove_QCO_NN\U_categories/4f4h4g\4h4e4g\4a4f4f\f4h4f4\4f4h4g4h4e4g4a4f4f4h4f4b4e4c4d4d4f4f4g4k4b4h4g4h4h4f4f4e4m4g4f4e4e4g4g_14 E:\glove_QCO_NN\U_categories/4f4h4g\4h4e4b\4g4i4j\j4d4f4\4f4h4g4h4e4b4g4i4j4d4f4g4h4f4f4f4h4d4i4f4h4c4d4j4c4i4i4d4h4f4h4e4h4g4f_12 E:\glove_QCO_NN\U_categories/4f4h4g\4h4442\4b4b4h\h4o4d4\4f4h4g4h44424b4b4h4o4d484f4g4n4d4o4h4c4k4k4i4b4m4g454d4e4j494h4f4f4g4o_10 E:\glove_QCO_NN\U_categories/4g4c3x\4d444b\474b4i\i584p4\4g4c3x4d444b474b4i584p4h4g4i514u524k4x4e4l4d574n4n3l494f4d3l4t4e544e46_5 E:\glove_QCO_NN\U_categories/4g4c40\4g4i4j\4m4c4i\i4o4b4\4g4c404g4i4j4m4c4i4o4b4f4n454h4f4n4k4t4f4k4c4h4s4d474m4g4e4d484p4m4j4c_6 E:\glove_QCO_NN\U_categories/4g4e3u\434948\4q4h4j\j4j4f4\4g4e3u4349484q4h4j4j4f4n4c494f4k4m4d4b4l4d4l4e4d4g4d4i4m4c4b4i4e4h4n4i_5 E:\glove_QCO_NN\U_categories/4g4e41\454b4j\4l4l4u\u4t4t4\4g4e41454b4j4l4l4u4t4t4l4f404c4h4f4d444f4u4n4o4i4j4k4i4e4f4f4a4g4r454c_6 E:\glove_QCO_NN\U_categories/4g4042\4i4g49\45474i\i4w4g4\4g40424i4g4945474i4w4g4440493z4b4c4l4k4k4e4c46454p4g444p4k4n4b4o3y4048_4 ---------------------------------------------- vectors_O original embeded vector : { _180 : 0.07, _181 : 0.32, _182 : -0.35, _183 : 0.52, _184 : -0.12, _185 : -0.18, _186 : -0.13, _187 : 0.31, _188 : -0.04, _190 : -0.30, _192 : -0.10, _193 : 0.33, _194 : -0.77, _195 : -0.20, _196 : 0.24, _197 : -0.21, _198 : -0.09, _199 : -0.02, _200 : -0.21, _201 : -0.13, _202 : -0.31, _203 : -0.03, _204 : -0.37, _205 : -0.31, _206 : 0.33, _207 : 0.19, _208 : 0.59, _209 : 0.28, _210 : -0.04, _211 : -0.42, _212 : -0.18, _213 : 0.14, _214 : -0.03, _215 : -0.36, _216 : 0.11, _217 : -0.51, _218 : -0.42, _219 : -0.13, _220 : -0.19, _221 : 0.25, _222 : 0.04, _223 : 0.14, _224 : 2.33, _225 : 0.03, _226 : -0.24, _227 : -0.31, _228 : -0.05, _229 : 0.02, _230 : 0.10, _231 : 0.17, _232 : 0.50, _233 : 0.04, _234 : -0.25, _235 : 0.64, _236 : 0.12, _237 : 0.28, _238 : -0.27, _239 : 0.02, _240 : 0.24, _241 : -0.28, _242 : -0.17, _243 : 0.39, _244 : -0.12, _245 : -0.03, _246 : -0.48, _247 : -0.45, _248 : -0.09, _249 : -0.23, _250 : -0.72, _251 : -0.08, _252 : -0.39, _253 : -0.33, _254 : -0.15, _255 : -0.06, _256 : 0.06, _257 : -0.27, _258 : -0.81, _259 : -0.12, _260 : -0.07, _261 : 0.27, _262 : -0.41, _263 : -0.09, _264 : -0.42, _265 : 0.27, _266 : -0.32, _267 : -0.04, _268 : -0.25, _269 : -0.14, _270 : 0.11, _271 : -0.10, _272 : -0.15, _273 : -0.24, _274 : 0.36, _275 : -0.15, _276 : 0.39, _277 : -0.14, _278 : -0.03, _279 : -0.01, _280 : 0.37, _281 : 0.07, _282 : 0.34, _283 : -0.66, _284 : -0.18, _285 : 0.15, _286 : -0.13, _287 : 0.21, _288 : -0.43, _289 : 0.01, _290 : 0.21, _291 : 0.15, _292 : 0.00, _293 : -0.36, _294 : 0.01, _295 : -0.30, _296 : 0.64, _297 : 0.34, _298 : -0.38, _299 : -0.30, _10 : -0.09, _11 : -0.40, _12 : 0.05, _13 : -0.25, _14 : 0.05, _15 : -0.00, _16 : 0.55, _17 : -0.31, _18 : 0.46, _19 : -0.33, _20 : -0.10, _21 : -0.37, _22 : -0.21, _23 : 0.15, _24 : -0.28, _25 : -0.36, _26 : -0.07, _27 : -0.22, _28 : -0.33, _29 : 0.29, _30 : 0.16, _31 : -0.05, _32 : -0.44, _33 : -0.37, _34 : -0.10, _35 : 0.10, _1 : -0.28, _36 : -0.26, _2 : -0.08, _37 : 0.49, _3 : -0.33, _38 : 0.18, _4 : 0.30, _39 : 0.15, _5 : -0.07, _40 : -0.12, _6 : -0.37, _41 : 0.14, _7 : 0.10, _42 : -0.01, _8 : 0.52, _43 : 0.39, _9 : 0.36, _44 : -0.10, _0 : 0.03, _45 : -0.02, _46 : -0.24, _47 : -0.17, _48 : 0.09, _49 : 0.37, _50 : -0.22, _51 : 0.27, _52 : -0.51, _53 : 0.16, _54 : -0.05, _55 : 0.03, _56 : 0.42, _57 : -0.31, _58 : -0.49, _59 : -0.29, _60 : -0.15, _61 : 0.12, _62 : -0.46, _63 : 0.33, _64 : -0.12, _65 : 0.30, _66 : 0.85, _67 : -0.50, _68 : -0.00, _69 : 0.07, _70 : 0.08, _71 : 0.53, _72 : 0.45, _73 : 0.31, _74 : -0.16, _75 : -0.02, _189 : -0.27, _191 : 0.18, _76 : 0.09, _77 : 0.42, _78 : 0.38, _79 : -0.39, _80 : 0.29, _81 : -0.16, _82 : 0.47, _83 : 0.21, _84 : 0.10, _85 : -0.12, _86 : -0.21, _87 : -0.53, _88 : -0.08, _89 : -0.26, _90 : -0.01, _91 : -0.17, _92 : -0.13, _93 : -0.07, _94 : -0.36, _95 : -0.29, _96 : 0.24, _97 : -0.37, _98 : -0.57, _99 : 0.46, _100 : -0.08, _101 : 0.02, _102 : 0.47, _103 : -0.12, _104 : -0.18, _105 : 0.48, _106 : -0.11, _107 : -0.58, _108 : 0.02, _109 : 0.20, _110 : -0.05, _111 : -0.11, _112 : -0.13, _113 : 0.36, _114 : 0.03, _115 : -0.25, _116 : -0.18, _117 : -0.80, _118 : 0.30, _119 : -0.16, _120 : 0.64, _121 : -0.37, _122 : -0.05, _123 : -0.06, _124 : -0.43, _125 : 0.41, _126 : 0.03, _127 : 0.24, _128 : 0.19, _129 : -0.05, _130 : -0.23, _131 : 0.04, _132 : -0.44, _133 : -0.21, _134 : -0.41, _135 : 0.01, _136 : -0.40, _137 : 0.25, _138 : -0.24, _139 : -0.26, _140 : -0.18, _141 : -0.15, _142 : 0.21, _143 : -0.12, _144 : -0.14, _145 : -0.47, _146 : -0.01, _147 : -0.43, _148 : 0.09, _149 : 0.17, _150 : -0.60, _151 : 0.08, _152 : 0.18, _153 : 0.18, _154 : -0.05, _155 : 0.20, _156 : -0.11, _157 : -0.26, _158 : -0.03, _159 : 0.23, _160 : -0.09, _161 : 0.06, _162 : -4.74, _163 : -0.16, _164 : 0.40, _165 : -0.44, _166 : 0.24, _167 : 0.04, _168 : -0.28, _169 : -0.06, _170 : 0.42, _171 : 0.60, _172 : -0.16, _173 : 0.12, _174 : -0.63, _175 : -0.37, _176 : -0.04, _177 : 0.32, _178 : 0.63, _179 : 0.29 } ---------------------------------------------- vectors_N normalised vector : { _180 : 0.01, _181 : 0.04, _182 : -0.05, _183 : 0.07, _184 : -0.02, _185 : -0.02, _186 : -0.02, _187 : 0.04, _188 : -0.01, _190 : -0.04, _192 : -0.01, _193 : 0.05, _194 : -0.10, _195 : -0.03, _196 : 0.03, _197 : -0.03, _198 : -0.01, _199 : -0.00, _200 : -0.03, _201 : -0.02, _202 : -0.04, _203 : -0.00, _204 : -0.05, _205 : -0.04, _206 : 0.05, _207 : 0.03, _208 : 0.08, _209 : 0.04, _210 : -0.01, _211 : -0.06, _212 : -0.03, _213 : 0.02, _214 : -0.00, _215 : -0.05, _216 : 0.01, _217 : -0.07, _218 : -0.06, _219 : -0.02, _220 : -0.03, _221 : 0.03, _222 : 0.00, _223 : 0.02, _224 : 0.32, _225 : 0.00, _226 : -0.03, _227 : -0.04, _228 : -0.01, _229 : 0.00, _230 : 0.01, _231 : 0.02, _232 : 0.07, _233 : 0.01, _234 : -0.03, _235 : 0.09, _236 : 0.02, _237 : 0.04, _238 : -0.04, _239 : 0.00, _240 : 0.03, _241 : -0.04, _242 : -0.02, _243 : 0.05, _244 : -0.02, _245 : -0.00, _246 : -0.07, _247 : -0.06, _248 : -0.01, _249 : -0.03, _250 : -0.10, _251 : -0.01, _252 : -0.05, _253 : -0.04, _254 : -0.02, _255 : -0.01, _256 : 0.01, _257 : -0.04, _258 : -0.11, _259 : -0.02, _260 : -0.01, _261 : 0.04, _262 : -0.06, _263 : -0.01, _264 : -0.06, _265 : 0.04, _266 : -0.04, _267 : -0.01, _268 : -0.03, _269 : -0.02, _270 : 0.02, _271 : -0.01, _272 : -0.02, _273 : -0.03, _274 : 0.05, _275 : -0.02, _276 : 0.05, _277 : -0.02, _278 : -0.00, _279 : -0.00, _280 : 0.05, _281 : 0.01, _282 : 0.05, _283 : -0.09, _284 : -0.02, _285 : 0.02, _286 : -0.02, _287 : 0.03, _288 : -0.06, _289 : 0.00, _290 : 0.03, _291 : 0.02, _292 : 0.00, _293 : -0.05, _294 : 0.00, _295 : -0.04, _296 : 0.09, _297 : 0.05, _298 : -0.05, _299 : -0.04, _10 : -0.01, _11 : -0.05, _12 : 0.01, _13 : -0.03, _14 : 0.01, _15 : -0.00, _16 : 0.07, _17 : -0.04, _18 : 0.06, _19 : -0.05, _20 : -0.01, _21 : -0.05, _22 : -0.03, _23 : 0.02, _24 : -0.04, _25 : -0.05, _26 : -0.01, _27 : -0.03, _28 : -0.04, _29 : 0.04, _30 : 0.02, _31 : -0.01, _32 : -0.06, _33 : -0.05, _34 : -0.01, _35 : 0.01, _1 : -0.04, _36 : -0.04, _2 : -0.01, _37 : 0.07, _3 : -0.04, _38 : 0.02, _4 : 0.04, _39 : 0.02, _5 : -0.01, _40 : -0.02, _6 : -0.05, _41 : 0.02, _7 : 0.01, _42 : -0.00, _8 : 0.07, _43 : 0.05, _9 : 0.05, _44 : -0.01, _0 : 0.00, _45 : -0.00, _46 : -0.03, _47 : -0.02, _48 : 0.01, _49 : 0.05, _50 : -0.03, _51 : 0.04, _52 : -0.07, _53 : 0.02, _54 : -0.01, _55 : 0.00, _56 : 0.06, _57 : -0.04, _58 : -0.07, _59 : -0.04, _60 : -0.02, _61 : 0.02, _62 : -0.06, _63 : 0.04, _64 : -0.02, _65 : 0.04, _66 : 0.12, _67 : -0.07, _68 : -0.00, _69 : 0.01, _70 : 0.01, _71 : 0.07, _72 : 0.06, _73 : 0.04, _74 : -0.02, _75 : -0.00, _189 : -0.04, _191 : 0.02, _76 : 0.01, _77 : 0.06, _78 : 0.05, _79 : -0.05, _80 : 0.04, _81 : -0.02, _82 : 0.06, _83 : 0.03, _84 : 0.01, _85 : -0.02, _86 : -0.03, _87 : -0.07, _88 : -0.01, _89 : -0.04, _90 : -0.00, _91 : -0.02, _92 : -0.02, _93 : -0.01, _94 : -0.05, _95 : -0.04, _96 : 0.03, _97 : -0.05, _98 : -0.08, _99 : 0.06, _100 : -0.01, _101 : 0.00, _102 : 0.06, _103 : -0.02, _104 : -0.02, _105 : 0.07, _106 : -0.02, _107 : -0.08, _108 : 0.00, _109 : 0.03, _110 : -0.01, _111 : -0.01, _112 : -0.02, _113 : 0.05, _114 : 0.00, _115 : -0.03, _116 : -0.02, _117 : -0.11, _118 : 0.04, _119 : -0.02, _120 : 0.09, _121 : -0.05, _122 : -0.01, _123 : -0.01, _124 : -0.06, _125 : 0.06, _126 : 0.00, _127 : 0.03, _128 : 0.03, _129 : -0.01, _130 : -0.03, _131 : 0.01, _132 : -0.06, _133 : -0.03, _134 : -0.06, _135 : 0.00, _136 : -0.05, _137 : 0.03, _138 : -0.03, _139 : -0.04, _140 : -0.02, _141 : -0.02, _142 : 0.03, _143 : -0.02, _144 : -0.02, _145 : -0.06, _146 : -0.00, _147 : -0.06, _148 : 0.01, _149 : 0.02, _150 : -0.08, _151 : 0.01, _152 : 0.02, _153 : 0.02, _154 : -0.01, _155 : 0.03, _156 : -0.01, _157 : -0.04, _158 : -0.00, _159 : 0.03, _160 : -0.01, _161 : 0.01, _162 : -0.64, _163 : -0.02, _164 : 0.05, _165 : -0.06, _166 : 0.03, _167 : 0.00, _168 : -0.04, _169 : -0.01, _170 : 0.06, _171 : 0.08, _172 : -0.02, _173 : 0.02, _174 : -0.09, _175 : -0.05, _176 : -0.01, _177 : 0.04, _178 : 0.09, _179 : 0.04 } ---------------------------------------------- U_adress p-addic adress : 4f4h4g4h4e4g4h4f4d4e4g4h4f4g4f4f4d4h4e4h4g4h4g4f4g4h4g4g4h4e4f4g4h4h4g4f4g4d4f4f4g4f4f4e4g4f4g4g4f4e4g4e4h4f4g4f4e4h4h4h4g4f4h4e4g4e4c4h4f4f4f4d4e4e4g4f4f4e4e4h4e4g4d4f4f4g4g4i4g4g4f4g4g4g4h4h4e4h4i4e4g4f4e4g4g4d4g4i4f4f4g4g4g4e4f4g4g4j4e4g4d4h4g4g4h4e4f4e4f4g4g4f4h4g4h4f4h4e4g4g4g4g4f4g4g4h4f4h4f4f4i4f4f4f4g4f4g4g4g4e4g4f4z4g4e4h4e4f4g4g4e4d4g4f4i4h4g4e4d4e4f4e4h4d4g4g4g4e4g4g4h4f4g4e4i4g4e4g4g4f4g4g4h4g4h4h4e4f4d4e4g4h4g4f4f4h4f4h4h4g4g4e4f4f464f4g4h4g4f4f4f4d4f4g4d4f4e4g4f4e4g4g4e4g4g4h4h4g4g4i4g4h4h4g4g4f4g4j4g4g4e4h4g4h4e4h4g4g4g4f4g4g4g4e4g4e4g4f4f4e4f4e4i4g4f4g4f4h4f4f4f4f4h4f4h4d4e4h4h ---------------------------------------------- Nearest table : position | Ultrametric | CosineSimiliarity | EuclidianDistance 0 | re-steal | equipment | apt 1 | zombe | 90s | 1970s 2 | re-infestations | credentials | flaw 3 | thisl | apt | readiness 4 | innocent | readiness | 90s 5 | fractured | 1970s | credentials 6 | littel | flaw | occasion 7 | ma’ayan | vulnerability | equipment 8 | predicatble | rust | 08 9 | thermocouple | logo | 06 10 | veryyyy | occasion | 1969 11 | pre-admissions | sporting | blow 12 | oneing | forensic | 1965 13 | non-consumption | duty | sporting 14 | chunkers | supervisory | inasmuch 15 | glarus | logos | logo 16 | nimit | blow | wayne 17 | 3mb | 1969 | heyday 18 | said.you | fake | vulnerability 19 | mid-lothian | 08 | duty 20 | delre | unreliable | supervisory 21 | worldcard | cables | awry 22 | galaz | iso | rust 23 | forensic | 06 | unreliable 24 | worrisome | 1965 | forensic 25 | adorable | wayne | logos 26 | romp | reinforced | fake 27 | i't | raspberry | fowler 28 | brutally | heyday | reinforced 29 | tirls | awry | iso 30 | canjs | purge | venerable 31 | traumatized | smb | worrisome 32 | mash.co.uk | 00s | 00s 33 | seals | inasmuch | smb 34 | dismaying | innocent | purge 35 | klasse | wipe | infuse 36 | twerp | withstand | cables 37 | surveyd | brutally | brutally 38 | m4w | headend | raspberry 39 | prick | infuse | rac 40 | sublcass | fowler | restrain 41 | bareback | scp | withstand 42 | ln2 | venerable | wipe 43 | bellieve | adorable | imprints 44 | fruitness | worrisome | scp 45 | onstantly | phishing | innocent 46 | rathvinden | tvo | chas 47 | designd | antique | dic 48 | collabor | dusty | romp 49 | reaaaaally | fractured | dusty 50 | scruffily | seals | crudely 51 | microblogger | romp | bruin 52 | impressible | legible | countermeasure 53 | bleah | moulding | saf 54 | militarized | traumatized | seals 55 | safety-wise | restrain | traumatized 56 | qiantian | odometer | paquette 57 | postulants | actuary | fractured 58 | 💪🏼 | rac | it.it 59 | istedgade | pussy | adorable 60 | exhaustivity | novell | salamon 61 | factorys | dic | highspeed 62 | 500billion | crudely | valentines 63 | 3,60,000 | imprints | clubbers 64 | sheepwash | countermeasure | tvo 65 | verʏ | hardbound | presages 66 | downclock | valentines | e’s 67 | phishing | haired | intels 68 | kjaergaard | prick | moulding 69 | safesquid | rejuvenate | plastered 70 | dessart | retrogaming | antique 71 | -28-2 | iee | 1003 72 | cosidering | saf | ’98 73 | rawlplugs | surety | legible 74 | non-autonomous | clubbers | odometer 75 | hongxi | tunnelling | deride 76 | counterfit | highspeed | pussy 77 | swet | eraser | averagely 78 | powercreeping | antimalware | sibyl 79 | elborough | newnes | prick 80 | serpentinite | plastered | thets 81 | privided | abdomen | 12b 82 | jenefer | intels | fatally 83 | trolldier | bruin | haired 84 | 130mw | winzip | chapple 85 | maake | neuralink | headend 86 | armistices | darkish | i't 87 | scrapbooker | thermocouple | dtb 88 | 6x7 | clothed | golub 89 | verabschiedet | fatally | an- 90 | glass.the | m4w | iee 91 | oreinted | chas | jrc 92 | 07747 | milfs | rejuvenate 93 | softrip | paquette | theop 94 | pelancong | vxlan | surety 95 | duragadget | frazzled | c.i. 96 | 28:14 | cuttings | g.l. 97 | 15,2019 | averagely | clothed 98 | contentsthe | wips | foolery 99 | delayeth | militarized | actuary 100 | walery | 1003 | humungous 101 | systemview | jrc | klasse 102 | 24060 | fukushima | goop 103 | apt | gulin | seddon 104 | flaw | 800k | taciturn 105 | credentials | 12b | liberational 106 | readiness | tiangong | preforms 107 | supervisory | bareback | novell 108 | awry | ’98 | tirls 109 | reinforced | hunks | straitlaced 110 | rust | facultative | eraser 111 | blow | goop | followon 112 | fake | rudnick | ruddy 113 | wayne | pre-wedding | tunnelling 114 | sporting | golub | arhitecture 115 | venerable | mathseeds | atbs 116 | e’s | deride | samas 117 | purge | chapple | 800k 118 | valentines | duggar | thetheir 119 | bruin | praeger | uprated 120 | 00s | humungous | fornicates 121 | dic | jinan | frazzled 122 | c.i. | eclair | nosis 123 | vulnerability | canonization | thisl 124 | saf | pheasant | gulin 125 | clubbers | it.it | dismaying 126 | raspberry | dhx | militarized 127 | theop | hwp | hirsute 128 | it.it | preforms | hunks 129 | futured | xrf | sheld 130 | countermeasure | salamon | designd 131 | dusty | minitool | anouk 132 | unreliable | quakes | gillam 133 | wipe | dtb | recommission 134 | workshirts | cornbread | milfs 135 | someand | dismaying | futured 136 | organzations | klasse | 6o 137 | moulding | lappy | eclair 138 | antique | kelas | jinan 139 | rejuvenate | meiji | bareback 140 | geny | tite | pheasant 141 | atbs | sibyl | plagiarist 142 | wips | uprated | diedrich 143 | 6o | interventionism | hankered 144 | articl | bruised | quiles 145 | chapple | adeyinka | gilmer 146 | followon | psip | darkish 147 | milfs | atbs | decarlo 148 | overplanting | jhang | anty 149 | arhitecture | decarlo | twerp 150 | seddon | taciturn | periera 151 | fornicates | scrapbooker | jazzman 152 | thets | g.l. | littel 153 | anouk | wxyz | snitching 154 | verrry | ls7 | shelfrate 155 | jrc | presages | phishing 156 | pocked | plagiarist | overplanting 157 | plastered | foolery | abdomen 158 | sheld | elektrotechnik | mlo 159 | taciturn | hornacek | русской 160 | preforms | acrimonious | gancho 161 | foolery | sv650 | ammuntion 162 | gancho | diedrich | retrogaming 163 | jesuitism | libav | sv650 164 | tite | snitching | wips 165 | pheasant | structed | cuttings 166 | 06/10 | mlo | bonneau 167 | jazzman | lybrand | organzations 168 | recommission | reinforcers | whilom 169 | too.there | swasey | reprobate 170 | forcs | nosis | deregulates 171 | anty | hirsute | m4w 172 | reengaging | seddon | spiffed 173 | straitlaced | c.i. | oberg 174 | uprated | среда | grizz 175 | liberational | ruddy | toman 176 | 9719 | gillam | dicovered 177 | shirogane | glarus | facultative 178 | intels | hergé | dalziel 179 | frazzled | shimamura | surveyd 180 | spiffed | webelos | shimamura 181 | hunks | meti | someand 182 | carrville | postulants | canonization 183 | laudenslager | verrry | bitrock 184 | sakhas | contenttype | teets 185 | sv650 | veryyyy | tite 186 | fukushima | jurgens | canjs 187 | tapasya | ruffini | counterchange 188 | tarradiddle | poltiical | verrry 189 | lappy | mk5 | pocked 190 | tsom | transferee | fantatic 191 | frathouse | русской | thermocouple 192 | nlly | eccentrically | forcs 193 | reprobate | goldmann | mk5 194 | highspeed | zenner | jurgens 195 | vsido | warriorship | meti 196 | tangyness | non-autonomous | bruised 197 | restrain | 6425 | articl 198 | nsrt | anouk | leay 199 | grizz | murnane | histry 200 | counterchange | un-teach | mash.co.uk 201 | odometer | forthis | bleah 202 | histry | structions | acrimonious 203 | fatally | negosyo | structions 204 | boychik | throated | reated 205 | dehradoon | railtrack | fukushima 206 | jettec | arhitecture | rudnick 207 | eraser | micro-climate | linek 208 | dtb | euro6 | tarradiddle 209 | pussy | folklorists | newnes 210 | angelspit | dalziel | reaaaaally 211 | superproduction | gilmer | severt 212 | melii | 235kw | meiji 213 | clothed | agogo | caparison 214 | legible | bonneau | insufficient 215 | woodwose | b01 | tapasya 216 | shimamura | samas | goldmann 217 | cuttings | 2281 | wordlessness 218 | baww | 3mb | dotée 219 | takaoka | factorys | masta 220 | glaucine | straitlaced | #weekend 221 | intimadating | scathach | http://www.testkingdump.com 222 | mk5 | shirogane | laudenslager 223 | klt | tapasya | jesuitism 224 | abdomen | oberg | 600ps 225 | reachout.com | tpdc | tsom 226 | dutil | grizz | dhx 227 | fantatic | #weekend | structed 228 | distend | tamazuj | 50mw 229 | snitching | xlpe | distend 230 | efudix | 6o | tsoukalas 231 | forthis | semi-distributed | murnane 232 | lönn | seacord | 9719 233 | darkish | fantatic | takaoka 234 | ekikayi.com | titten | fruitness 235 | meof | crushercrusher | baww 236 | timbercrete | housewifely | ex-zs190 237 | un-teach | saheeh | kelas 238 | plagiarist | matchroom | codrington 239 | triternion | lasu | weatherbox 240 | contradictary | vived | housebreaker 241 | среда | reengaging | kwezi 242 | devspace | littel | exhaustivity 243 | rwting | ptj | predicatble 244 | gilmer | followon | reengaging 245 | kwezi | reaaaaally | factorys 246 | housebreaker | isacson | forthis 247 | 1/15/18 | librenms | photgraph 248 | skysite | 3054 | poltiical 249 | non-iranians | dcac | jettec 250 | caparison | articl | ma’ayan 251 | weatherbox | bloomreach | dehradoon 252 | guédelon | detonates | winzip 253 | severt | eikawa | leão 254 | ex-zs190 | kasumigaseki | militiary 255 | cross-linkages | qilin | quean 256 | codrington | bleah | arep 257 | learningspace | maberry | detonates 258 | bloomon | tecd | qilin 259 | minipcr | 50mw | b01 260 | mlo | organzations | scruffily 261 | ls7 | superproduction | mcfetridge 262 | quean | designd | microblogger 263 | retrogaming | afterwork | p181 264 | deregulates | soltero | 235kw 265 | 275ish | post-renaissance | dutil 266 | http://www.testkingdump.com | speedframe | quakes 267 | 541s | serveurs | post-2001 268 | explan | i't | lappy 269 | decarlo | freightliners | среда 270 | rhytms | beavertown | 1600h 271 | barolos | whatchya | geny 272 | murnane | cavil | sayal 273 | p181 | pre-admissions | otoo 274 | kelas | goldwin | skrew 275 | radzik | innominate | mislay 276 | otoo | dotée | knifed 277 | informd | deregulates | 3682 278 | mislay | akhuwat | thexample 279 | anchylosis | downclock | frj 280 | netdisco | fouse | certificat 281 | newnes | kanjorski | shirogane 282 | 2156 | subcorpora | workshirts 283 | goldmann | hollowcore | neuralink 284 | zvchattrick | mcfetridge | rathvinden 285 | testification | stroboscope | 4361 286 | masta | klt | aspirer 287 | moldes | logp | ultra-cool 288 | giantkiller | oreinted | nsrt 289 | tsoukalas | shelfrate | racebike 290 | 12/3/18 | heavendefying | archconservative 291 | rednecked | canjs | 2281 292 | burnsie | ckcu | retaliations 293 | seacord | gabardine | testification 294 | post-2001 | thets | faher 295 | goldwin | time.to | adeyinka 296 | dotée | blastoderm | dcac 297 | oberg | campin | antimalware 298 | jg1 | dsch | 12365 299 | sw2x | liberational | 06/10 300 | filеs | weeeeee | plauge 301 | hickock45 | sremmurd | moldes 302 | howhe | runalyze | ls7 303 | teets | birendra | folklorists 304 | iokalkompakter | vsats | interventionism 305 | marrriage | pachisuro | throated 306 | linek | mehsana | serveurs 307 | reqired | 2156 | eccentrically 308 | poltiical | prieuré | stroboscope 309 | difficoult | ln2 | netdisco 310 | aspirer | luminol | elektrotechnik 311 | cornbread | yapım | hwp 312 | mcfetridge | reprobate | too.there 313 | chemisier | toman | postulants 314 | grazioli | 75hp | lumineux 315 | zorastrian | certificat | sekhon 316 | zygospore | watchuseek | veryyyy 317 | lumineux | nangong | overspun 318 | facultative | vidcruiter | verʏ 319 | condign | pre-historical | 1/15/18 320 | zenner | ultra-cool | adumbrating 321 | phytogenic | belg | tiangong 322 | pedallers | lumineux | titten 323 | plauge | weatherbox | vxlan 324 | certificat | racebike | vsido 325 | 221m | 600ps | radzik 326 | dints | acquaintanceship | w95 327 | qilin | condign | pre-historical 328 | pro-erdogan | researchin | ruffini 329 | thrivehive | geny | 6425 330 | freightliners | periera | un-teach 331 | mology | recommission | ln2 332 | monchhichi | xuzhou | 3054 333 | lasu | 5525 | hardbound 334 | waterbars | jazzman | руководства 335 | runalyze | skep | woodwose 336 | dabbu | twerp | rimoldi 337 | bruised | clickondetroit | tangyness 338 | heterophylla | moldes | sakhas 339 | nuclearly | sayal | warriorship 340 | money360 | codrington | seacord 341 | cool. | anty | birendra 342 | sanad | non-fasting | qiantian 343 | fjallraven | quean | kahre 344 | seajet | safety-wise | glaucine 345 | b01 | wtsp | 71m 346 | toontasks | jg1 | kanus 347 | gabardine | sekhon | xrf 348 | surecare | eowg | skep 349 | thexample | fornicates | chemisier 350 | re-normalize | fjallraven | saheeh 351 | perfessor | jinxy | boychik 352 | sekhon | kensal | paleographer 353 | sensationell | zeebrugge | reqired 354 | mccarraher | 1600h | harrowed 355 | ptj | dutil | xuzhou 356 | campin | enmap | giantkiller 357 | post-renaissance | spiffed | cross-linkages 358 | rimoldi | nextcure | spankingly 359 | gazipur | leão | zvchattrick 360 | aymes | exhaustivity | militray 361 | shirleysburg | skaife | goldwin 362 | yestreen | 8950 | fictitiousness 363 | akhuwat | reated | yapım 364 | winzip | dashper | nuclearly 365 | de-incentivize | amrep | jg1 366 | stroboscope | reqired | safety-wise 367 | developin | 1/15/18 | belg 368 | acetohydroxamic | powercreeping | superproduction 369 | yeogh | proca | collabor 370 | 71t | takaoka | germanophile 371 | artcile | onstantly | others.but 372 | reserve_met | hodgdon | lifestyling 373 | kanus | harrowed | praeger 374 | adeyinka | hoopz | yestreen 375 | pernas | retaliations | micro-climate 376 | throated | rimoldi | amrep 377 | micro-climate | pocked | impressible 378 | barwari | tsom | zygospore 379 | librenms | spf20 | hermés 380 | vieille | trypanothione | zenner 381 | 378th | wordlessness | moonfish 382 | tpdc | quiles | skysite 383 | minafer | masta | reteams 384 | sacrae | futured | freightliners 385 | jalisa | de-incentivize | 25,550 386 | weatherresistant | 3dprinteros | 5525 387 | titten | 04-26-2019 | acamp 388 | subjugations | plauge | reachout.com 389 | elektrotechnik | dicovered | nlly 390 | pocl3 | nuclearly | alemanha 391 | betteredge | too.there | gazipur 392 | deppressed | pursue.we | explan 393 | retaliations | lip6 | dashper 394 | kouros | jeck | frathouse 395 | quakes | sigafoos | money360 396 | 107r | egat | forerunning 397 | skep | baww | schwanzer 398 | clickondetroit | tler | mega-bank 399 | amassment | serpentinite | sublcass 400 | zeebrugge | 3682 | amassment 401 | arezo | caparison | subjugations 402 | asren | otoo | bloomon 403 | luminol | образовании | de-noised 404 | dcac | 4361 | 75hp 405 | offduty | gazipur | 💪🏼 406 | vsats | reteams | 90.10 407 | 8538 | dints | wrastle 408 | 18/07/2018 | salzedo | minipcr 409 | loyet | gancho | pre-wedding 410 | jerichow | hosepipe | 3,60,000 411 | dashper | cbed | onstantly 412 | 75hp | mash.co.uk | lönn 413 | hergé | c36 | re-steal 414 | over-predict | 651,000 | hodgdon 415 | labatory | g.m.o. | balkanizing 416 | lip6 | hankered | railtrack 417 | moulted | eliud | melii 418 | 09/09/2018 | leay | duggar 419 | prieuré | akanbi | prieuré 420 | passme | listahanan | mology 421 | xlpe | kouros | contradictary 422 | peroz | precatory | plummage 423 | housewifely | priveledged | dabbu 424 | balkanizing | distend | 2156 425 | lolipops | predicatble | devspace 426 | amrep | 💪🏼 | workthere 427 | rollingstone | 25,550 | cornbread 428 | dockrill | kwezi | zorastrian 429 | bocha | vieille | grazioli 430 | greencraft | difficoult | negosyo 431 | hornacek | linek | waterbars 432 | re-assure | arietis | burnsie 433 | saheeh | bivariant | arezo 434 | antimalware | privided | marrriage 435 | jinxy | pathologize | transferee 436 | cosmpolitan | b.t.c. | bostwana 437 | structed | pro-erdogan | pre-admissions 438 | de-noised | neoseeker | kawaski 439 | tamazuj | jobcliq | non-iranians 440 | pierse | mymaster | babyboom 441 | reg’lar | bi-energy | sensationell 442 | pro-austerity | scruffily | twor 443 | nextcure | antipope | seajet 444 | скачивания | ammuntion | developin 445 | precatory | bujanda | condign 446 | heathershaw | kavinoky | 275ish 447 | 2016-12-06 | q1fy19 | 221m 448 | conzo | guédelon | b.t.c. 449 | praeger | t480 | balter 450 | hosepipe | dustier | 8538 451 | wieviorka | conzo | reserve_met 452 | 2020-7-3 | mk.iii | clickondetroit 453 | xrf | ld20 | jerichow 454 | unpicker | roussell | vieille 455 | moirs | whilom | maslin 456 | 06/09/2017 | sanad | vidcruiter 457 | innominate | knifed | cosmpolitan 458 | bichet | granodiorite | meof 459 | gruffy | tebowmania | heterophylla 460 | reallllyyyyy | post-2001 | jalisa 461 | dyron | juryman | selfinterest 462 | intimidatory | qee | heggen 463 | leshay | moulted | anchylosis 464 | easy2sync | vasoocclusive | barwari 465 | mehsana | housebreaker | qee 466 | priveledged | collabor | djebali 467 | montka | vevay | iokalkompakter 468 | cavil | militray | triternion 469 | mathseeds | jenefer | offduty 470 | reichmarks | gp12 | rhytms 471 | sagaing | shapin | difficoult 472 | orgasmus | covenanting | monchhichi 473 | wyf | marquit | gruffy 474 | jamfest | pierse | vevay 475 | marquit | theop | 12/3/18 476 | hooniverse | yatterman | 71t 477 | kopykake | cross-linkages | rednecked 478 | roussell | 6735s | angelspit 479 | blastoderm | company.this | maseru 480 | qx6700 | bordier | artcile 481 | inkspired | 1q19 | cool. 482 | radiculomedullary | asren | paradoxic 483 | ninti | serhan | cosidering 484 | fritch | explan | runalyze 485 | гun | unshoveled | dints 486 | hewl | bluhorn | swasey 487 | swasey | sublcass | faulds 488 | ts9 | duragadget | librenms 489 | alders | 35007 | schlaff 490 | pre-seen | calik | perfessor 491 | classmap | schwanzer | sw2x 492 | rohla | flikr | wxyz 493 | vived | militiary | juryman 494 | freshy | euphrosyne | oneing 495 | ijames | walery | subcorpora 496 | unshoveled | wickstead | rwting 497 | schlaff | zvchattrick | betteredge 498 | 57yo | aspirer | cavil 499 | add2exchange | lendon | bellieve 500 | biomotion | w95 | innominate 501 | naksan | kawaski | 3red 502 | biscuitville | p181 | learningspace 503 | yongzhi | phytogenic | informd 504 | transferee | body.at | lasu 505 | coldspots | kopykake | libav 506 | cavaness | ficano | biomotion 507 | samaipata | q200 | re-normalize 508 | heggen | silverlit | pro-erdogan 509 | krzemiÅ„ski | 71t | jiwan 510 | rpmforge | rhytms | aymes 511 | anti-desiccant | sacrae | ptj =====================END==============================