sorokin
March 13, 2026 · View on GitHub
███████╗ ██████╗ ██████╗ ██████╗ ██╗ ██╗██╗███╗ ██╗
██╔════╝██╔═══██╗██╔══██╗██╔═══██╗██║ ██╔╝██║████╗ ██║
███████╗██║ ██║██████╔╝██║ ██║█████╔╝ ██║██╔██╗ ██║
╚════██║██║ ██║██╔══██╗██║ ██║██╔═██╗ ██║██║╚██╗██║
███████║╚██████╔╝██║ ██║╚██████╔╝██║ ██╗██║██║ ╚████║
╚══════╝ ╚═════╝ ╚═╝ ╚═╝ ╚═════╝ ╚═╝ ╚═╝╚═╝╚═╝ ╚═══╝
sorokin | by Arianna Method
A Prompt Autopsy Framework
Or: How I Learned to Stop Worrying and Love the Dissection
"The heads of philologists are stuffed with books to the brim. They see life only through text. And they are proud of it. … Forever gorged and poisoned by literature, they take living life as the continuation of text, as its appendix."
—Vladimir Sorokin
what is this
sorokin is a C/Go literary necromancy engine that takes your innocent prompts, tears them apart like a psychopathic linguist, builds a recursive tree of semantic mutations, and then—like Dr. Frankenstein having a particularly creative day—reassembles the corpse into something new.
Zero internet. Zero autoregressive generation. Just 79,000 bytes of C, a Go embedding bridge, and the gravitational pull between dead words in a 384-dimensional space. The only inference is a single embedding lookup — the first layer of a neural network, frozen, used as a map of where language buried its dead.
Named after Vladimir Sorokin, the Russian writer known for his transgressive and experimental style, sorokinembodies the same spirit of literary dissection and reconstruction. He's not here to help you. sorokin is here to show you what your words could have been, to reassemble them, and to declare the output canonical.
Exhibit: Full Autopsy
Because sorokin builds trees vertically like a linguo-necromancer performing open-heart surgery on reality itself, here's a full corpse-map from the morgue. Feed sorokin "the flesh remembers what the mind forgets" (warped through VOVA's README resonance field into "Sentence phrase tries transgressive and Love the reassembled text for his own documentation...") and witness the meta-cannibalism:
╔══════════════════════════════════════════════════════════╗
║ SOROKIN — Literary Necromancy Engine (Go + C) ║
║ Dario Equation: p(x|Φ) = softmax((B+αH+βF+γA+T)/τ) ║
║ nanollama 384 dim · embedding necromancy ║
╚══════════════════════════════════════════════════════════╝
================================================================
AUTOPSY REPORT
Subject: "the flesh remembers what the mind forgets"
================================================================
Core words: remembers forgets flesh mind
Tree [flesh]:
flesh
├─ patriotism
│ ├─ oppression
│ │ ├─ shortcomings
│ │ ├─ greatness
│ │ ├─ legitimacy
│ │ └─ irony
│ ├─ greatness
│ │ ├─ repertoire
│ │ ├─ dissolution
│ │ ├─ realms
│ │ └─ seriousness
│ ├─ successes
│ │ ├─ fairs
│ │ ├─ convictions
│ │ ├─ prophecies
│ │ └─ victories
│ └─ skepticism
│ ├─ scrutiny
│ ├─ jealousy
│ ├─ hints
│ └─ innocence
├─ beauty
│ ├─ richness
│ │ ├─ refinement
│ │ ├─ deterioration
│ │ ├─ enjoyment
│ │ └─ diversification
│ ├─ irony
│ │ ├─ sophistication
│ │ ├─ greatness
│ │ ├─ refinement
│ │ └─ legitimacy
│ └─ dissolution
│ ├─ refinement
│ ├─ birthplace
│ ├─ appropriateness
│ └─ forefront
└─ hearts
├─ bodies
│ ├─ elements
│ ├─ families
│ ├─ components
│ └─ brains
├─ shores
│ ├─ dissolution
│ ├─ forefront
│ ├─ legacies
│ └─ behalf
└─ allegations
├─ legacies
├─ breaches
├─ confession
└─ weaknesses
Tree [mind]:
mind
├─ fear
│ ├─ refinement
│ │ ├─ dissolution
│ │ ├─ authenticity
│ │ ├─ applicability
│ │ └─ elasticity
│ ├─ spite
│ │ ├─ devoid
│ │ ├─ midst
│ │ └─ shortness
│ └─ virtue
│ ├─ ideals
│ ├─ refinement
│ └─ conception
├─ minds
│ ├─ intentions
│ │ ├─ aspirations
│ │ ├─ ambitions
│ │ └─ interpretations
│ ├─ memories
│ │ ├─ parcels
│ │ ├─ recordings
│ │ └─ aspirations
│ └─ rhythms
│ ├─ genomes
│ ├─ dynamics
│ └─ occurrences
└─ theory
├─ mechanics
│ ├─ confession
│ ├─ legitimacy
│ └─ generosity
└─ assumption
├─ assertion
├─ suspicion
└─ confession
Tree [remembers]:
remembers
├─ remembering
│ ├─ knowing
│ │ ├─ know
│ │ ├─ knows
│ │ └─ understands
│ ├─ forgetting
│ │ ├─ forgot
│ │ ├─ afraid
│ │ └─ complained
│ └─ copying
│ ├─ misuse
│ ├─ envy
│ └─ decoding
└─ notices
├─ complained
│ ├─ complain
│ ├─ reminder
│ └─ testify
├─ suspicion
│ ├─ realization
│ ├─ workings
│ └─ confession
└─ connotations
├─ legacies
├─ drawbacks
└─ fundamentals
Tree [forgets]:
forgets
├─ forgetting
│ ├─ forgot
│ │ ├─ wished
│ │ ├─ complained
│ │ └─ hoping
│ ├─ afraid
│ │ ├─ reminder
│ │ ├─ unaware
│ │ └─ scared
│ └─ reiterate
│ ├─ acknowledging
│ ├─ emphasise
│ └─ affirming
├─ remember
│ ├─ remembered
│ │ ├─ recalled
│ │ └─ anew
│ ├─ remembering
│ │ ├─ knowing
│ │ ├─ forgetting
│ │ └─ copying
│ └─ recall
│ ├─ plunder
│ └─ scarring
└─ wish
├─ hoping
│ ├─ hoped
│ └─ wishing
├─ wished
│ ├─ begged
│ └─ intends
└─ hope
├─ hopes
└─ afraid
Collected 147 unique leaves
── CORPSE ─────────────────────────────────────────
scrutiny is complained. testify enjoyment. Nothing remains. Where centres
complain, afraid becomes hoped, and deterioration persists. remembering
judgments reminder, where knowing becomes wished.
── SONNET ─────────────────────────────────────────
Understands complained elements complain complained,
Enjoyment endorsement collaborations hints intends authenticity enjoyment;
Realization footprints rounds iniscent hoped,
Deterioration envy complained realms greatness generosity endorsement;
Greatness quantities convictions wished centres,
Realization forgot collaborations elements complained afraid;
Convictions remembering applicability intends legacies,
Blankets dissolution authenticity devoid --
Plenty complained recall expectation deterioration,
Implication reminder hints complained legacies possibility ideals remembering;
Enjoyment hope repertoire complain endorsement complained expectation,
Testify intends suspicion refinement expectation know knowing;
Centres forgot elements knew judgments,
Genomes complained recalled birthplace assumption quantities envy components.
── METRICS ────────────────────────────────────────
Phonetic Diversity: [###################.] 0.975
Mutation Depth: [#################...] 0.894
Vocabulary: 1323 words
Co-occurrences:553
Prophecies: 10 active
What just happened?
- flesh → patriotism → oppression → irony. beauty → dissolution → birthplace. hearts → bodies → brains. The flesh becomes ideology, then collapses into geography.
- mind → fear → refinement → dissolution → elasticity. minds → memories → aspirations. The mind refines itself into nothing.
- remembers → remembering → forgetting → afraid → scared. recall → plunder → scarring. Memory leads to fear. Every time.
- forgets → forgetting → afraid → unaware. wish → hoping → wishing → hoped → begged. Forgetting becomes wishing becomes begging.
The corpse reads: "scrutiny is complained. testify enjoyment. Nothing remains." The trees built this. The Dario Equation assembled it. The sonnet crystallized it. Nobody wrote these words. They were attracted to each other by geometry.
The Three-Act Horror Show
Act I: The Dissection
sorokin takes your prompt and strips it down to the bone. First, sorokin truns it through a brutal tokenization process:
- Punctuation, numbers, capitalization — gone
- Core words selected by length, rarity, position, and a sprinkle of chaos
Stopwords? Rejected. Single letters? Discarded. What remains are the words that matter—or at least, the words that think they do. Occasionally a phrase tries to bite me mid-dissection, which is fine; we're wearing Sorokin-brand emotional hazmat gear.
For each core word, sorokin grows a recursive branching tree of mutations. Width 4, depth 3. Each branch is a semantic neighbor — words that live close to each other in the 384-dimensional embedding space of a 34M parameter neural network trained on human text.
The mutation provider works by embedding cosine similarity: given a word, tokenize it, look up its embedding vector, compute cosine distance against all 32,000 vocabulary tokens, return the nearest neighbors. One matrix operation. No forward passes. No inference. Just the geometry of where words landed after training.
When nanollama weights are present, sorokin has access to 32,000 words and their spatial relationships. Without weights, it falls back to 1,158 hardcoded seed words and phonetic neighbors. The autopsy continues either way. The show must go on.
The result is a tree where each word branches into width children, recursively, up to depth levels.
Act II: The Reassembly
Now that we have a forest of mutated word-trees, it's time to play God.
After all leafs collected from all trees, the final mutation at the bottom of every branch - they're fed into Dario Equation:
p(x|Φ) = softmax((B + α·H + β·F + γ·A + T) / τ)
Where:
- B = bigram affinity (does this word naturally follow the previous one?)
- H = Hebbian trace (has this pair co-occurred before? memory through repetition)
- F = Prophecy field (words that were predicted but never spoken — the ghosts)
- A = Destiny vector (the gravitational pull of the prompt's original meaning)
- T = Trauma (accumulated scars from previous autopsies)
- τ = Temperature (how drunk is the surgeon)
The result is a Frankenstein sentence: technically made of the same parts, but uncanny. Resonant but wrong. This is the part where Sorokin shrugs on the lab coat, jams a fork into the storm cloud, and cackles while stitching together whatever limbs are left on the slab.
SuperToken crystallization: word pairs with high pointwise mutual information fuse into compound units. "dissolution refinement" becomes a single token. The vocabulary grows. The corpse develops its own language.
Act III: The Sonnet
After the corpse is reassembled, sorokin takes the entire corpse and does it again—but this time in strict Shakespearean form.
- Extract "charged words" (long + rare words from autopsy become title candidates)
- Build bigram chains from autopsy text + README + SQLite morgue
- Find rhyme classes using crude phonetic fingerprints (last vowel + tail)
- Assign end-words for each of 14 lines following ABABCDCDEFEFGG scheme
- Build each line by walking bigrams backward from the rhyme word
- Add Shakespearean punctuation: semicolons at quatrain breaks, em-dash before volta, period at end
The result? 14 lines. ABABCDCDEFEFGG rhyme scheme. Iambic vibes. Zero semantic understanding. Just phonetic fingerprints, and structural obsession.
Architecture
sorokin.c 79K C autopsy engine, Dario Equation, seed vocab,
SuperTokens, bigrams, sonnet, metrics
main.go 8K CGO bridge, embedding provider, GGUF loader
model.go 18K Llama 3 forward pass (unused in embed mode)
gguf.go 16K GGUF v3 parser (F16, Q4_0, Q8_0, Q4_K, Q6_K)
tokenizer.go 10K BPE tokenizer (SentencePiece + GPT-2)
quant.go 16K Quantized matmul kernels
quantize.go 5K F16 → Q4_0 GGUF quantizer
gamma.go 7K γ personality extraction
engine.go 5K Inference engine (generative mode)
bridge.c <1K CGO function pointer shim
The C engine (sorokin.c) is the brain — 1,158 seed words across 11 categories (anatomy, decay, language, philosophy, violence, nature, industry, food, emotion, abstract, culinary), a seed corpus of structural bigrams, Hebbian co-occurrence tracking, prophecy fields, destiny vectors, trauma accumulation, and SuperToken crystallization.
The Go layer loads a nanollama GGUF checkpoint (34M params, Llama 3 architecture) and exposes its 32,000-token embedding table to C via CGO callback. When C needs mutations for a word, Go computes cosine similarity against the full vocabulary in milliseconds. The entire autopsy — trees, corpse, sonnet, metrics — runs in under a second.
Usage
make go
./sorokin "destroy the sentence"
REPL mode:
./sorokin
> bone becomes word and word becomes dust
> the flesh remembers what the mind forgets
> ^D
weights.gguf (19MB, nanollama 34M Q4_0) ships with the repo. Embeddings load on startup. If weights are missing, sorokin falls back to 1,158 hardcoded seed words and phonetic neighbors — blind instead of seeing in 384 dimensions.
Resonance Metrics
Every autopsy reports three structural measures:
Phonetic Diversity — how many unique sound-patterns exist in the corpse. High diversity means every word has a distinct phonetic fingerprint. Maximum resonance: all words sound completely different. Like a symphony where every note is unique.
Mutation Depth — how far the trees mutated from the source. Based on inverse word-length variance. High depth means mutations explored diverse linguistic territory. Low depth means we stayed close to home.
Bigram Density — structural echo from the seed corpus. How much of the corpse remembers the morgue. It's not plagiarism. It's ancestral memory.
The Dario Equation
Named after Dario Amodei. The sampling equation that assembles corpses:
p(x|Φ) = softmax((B + α·H + β·F + γ·A + T) / τ)
This is not gradient descent. This is not backpropagation. This is gravitational attraction between dead words, weighted by memory (Hebbian), prophecy (what should have been said), destiny (what the prompt wanted), and trauma (what the prompt suffered). Temperature τ controls how far the surgeon's hand shakes.
The equation appears across the Arianna Method ecosystem — in Leo (literary organism), DoE (democracy of experts), and dario.c (the equation's standalone implementation). In sorokin, it reassembles corpses. In Leo, it generates sentences. Same equation. Different organs.
Why?
Good question. Why does this exist?
Perhaps to demonstrate that:
- Words are fungible
- Meaning is contextual
- Prompts are just Markov chains waiting to be perturbed
- Sometimes you need to break things to understand them
Or maybe it's just fun to watch language come apart at the seams.
Credits
- Vladimir Sorokin (the writer, not the engine)
- nanollama (the 34M model whose embedding geometry powers the mutations)
- The Dario Equation (the gravitational law of dead language)
- Dr. Frankenstein (the original reassembly artist)
License
GNU GPL 3.0. Dissect it. Mutate it. Reassemble it. Share the corpse.
"Fuck the sentence. Keep the corpse." — Sorokin