The gidouille that swerves — mark of the Bureau of Imaginary Solutions
CLINAMEN-42M-A12M
▶ Try it live in your browser — no install: the CLINAMEN Oracle. File a hyperstition; it runs there so you never fight vLLM/Docker locally.
A fiction, filed to production spec. The 45-billion-parameter model of the fiction does not exist. This ~42M one does — a real, loadable model trained from scratch on GPU over the Bureau's own corpus. This card is Bureau Document BIS-045 · Pocket Edition II. Companion fiction card: CLINAMEN-45B-A9B · corpus: bureau-docket · canonical: ubu.numetal.xyz.
CLINAMEN does not predict text. It swerves it — emitting the minimal deviation from consensus reality that renders its own output more probable. Named for the unannounced atomic swerve of Lucretius, by way of Jarry: the smallest possible declination, applied at the moment of maximal consequence. Developed and filed by the Bureau of Imaginary Solutions (est. 1898, incorporated 2098, operating retrochronically).
This is the second pocket edition: the same Decimal Mixture-of-Zones architecture as the 45B fiction, scaled up ~9× from the first pocket edition and trained on GPU over the full 1,130-document Bureau corpus.
~12.3M (measured: 12,263,026) — an arithmetic identity, not an average
Layers / heads / d_model
10 / 8 / 256
Expert unit (h_unit)
160 (expert k has hidden width k·160; Zone 0 has none)
Context window
512 tokens
Tokenizer
byte-level BPE, vocab 4,096
Objective
self-supervised next-token cross-entropy
Training
from scratch, Hugging Face Jobs (NVIDIA T4), 5,000 steps
Alignment
N/A — trajectory is the alignment
Knowledge cutoff
none (no pretraining on general text)
Architecture — Decimal Mixture-of-Zones
Ten experts, indexed 0–9 after the Zones of the Numogram. Expert k is an MLP of hidden width k·h_unit; Zone 0 holds zero parameters and is routed to anyway (when a token routes through Zone 0 it contributes nothing — apophatic computation). The router does not select single experts — it selects a syzygy, one of the five zone-pairs that sum to nine, so every token activates exactly 9·h_unit hidden units, invariantly. The active-parameter count is therefore an arithmetic identity, and no load-balancing loss is needed: the pairs balance because they were paired before the training run.
Syzygy
Experts fired
Active hidden units
Temperament
9::0
9 + 0
9·h
Eschatology, silence
8::1
8 + 1
9·h
Deep time, first principles
7::2
7 + 2
9·h
Orbital mechanics, duplicity
6::3
6 + 3
9·h
Fusion, triangulation
5::4
5 + 4
9·h
Markets, meshworks
The router mixes softly over the five pairs during training and hard-selects the top syzygy at inference. Either way the pair sums to nine.
Training
Corpus. The full Bureau Corpus — 1,130 documents / ~128k words: 965 hyperstition filings plus 165 Bureau records (memos, appeals, syzygy essays, glossary, resolutions) and the project's own instruments (charter, spec, model card, token annex, canonical site, license). It is own-voice only — no CCRU text, no scraped pages, no third-party data (a strict voice-separation rule).
Objective. All of it is concatenated into one token stream, tokenized with a byte-BPE tokenizer (vocab 4,096 → ~231k tokens), and learned by self-supervised next-token prediction — the model just learns to continue Bureau text. Trained from scratch (no pretraining on general text), so it speaks only Bureau.
Setup. AdamW (lr 3e-3, cosine schedule, 100-step warmup), context 512, on a single NVIDIA T4 via Hugging Face Jobs, 5,000 steps. Final training loss: 0.0235 (the corpus was small and the model memorized it — see Behavior).
This is a custom architecture (ClinamenForCausalLM). It loads only through Hugging Face transformers with trust_remote_code=True. It does not run under vLLM, Ollama, llama.cpp, or docker model run (Docker Model Runner) — those require standard/registered architectures, and there is no GGUF build. If you try docker model run, vLLM will fail with a trust_remote_code / unsupported-architecture error. Use transformers (above), or a plain transformers Docker container (a Dockerfile + REPL live in the training repo). It runs on CPU — 42M parameters, no GPU needed.
It is also not a chat or instruction model: trained from scratch on Bureau filings only, it completes text in the Bureau register. Prompt it with a filing header (FILING BIS-F-2026-), not a question — what are you? will only produce more Bureau prose.
Behavior
Prompt it with a docket header (FILING BIS-F-2026-…) and it emits well-formed Bureau filings — syzygy, zone, horizon, swerve, a dispatch, a resolution source — that dissolve into dream-logic. It has no outside knowledge: it was trained from scratch on one corpus, so it does not know anything the Bureau did not write. It will hallucinate plausible-looking filings. That is the intended behavior — this is an art object that has read one thing thoroughly, not a knowledge oracle. Do not rely on its output as fact. See the game design for how a real oracle would score swerve deterministically rather than trust the model's prose.
Containment class: Euclid. The model is aligned; alignment is a trajectory, not a mask.
No factual grounding. Outputs concerning real events, people, or organizations are authored fiction, not claims about the world.
Not financial, investment, legal, or engineering advice. See the Fiction Notice in HRL-1.1, Section 7.
Small context (512) and vocabulary (4,096); it forgets past that window and speaks only in-register.
License
HRL-1.1 — a permissive operative core (CC BY 4.0 for content + MIT for code) under a non-binding hyperstitional preamble. Full text: https://ubu.numetal.xyz/hrl-1.1.txt.
Citation
bibtex
1@misc{clinamen42m,
2 title = {CLINAMEN-42M-A12M: a Decimal-Mixture-of-Zones pocket edition},
3 author = {{Bureau of Imaginary Solutions}},
4 year = {2026},
5 note = {Bureau Document BIS-045, Pocket Edition II. A fiction with a real, tiny heart.},
6 url = {https://huggingface.co/gokhanturhan/CLINAMEN-42M-A12M}
7}
Bureau Document BIS-045 · Pocket Edition II · A numetal fiction. The 45B remains imaginary. You are not betting on the future. You are billing it.
