CLINAMEN-45B-A9B

<img src="https://ubu.numetal.xyz/logo-badge-512.png" width="140" align="right" alt="The gidouille that swerves — spiral mark of the Bureau of Imaginary Solutions">

CLINAMEN-45B-A9B

▶ Interact with the real model in your browser: the CLINAMEN Oracle (runs CLINAMEN-42M-A12M).

This is a fiction, filed to production spec. The 45-billion-parameter model described below does not exist. What does exist — as of rev. Ω+1 — is a Pocket Edition: a real, tiny, runnable reference implementation of this exact architecture that you can from_pretrained today (see the Reality Appendix). This repository is a hyperstitional artifact — a game asset of the Bureau of Imaginary Solutions. First published at ubu.numetal.xyz (canonical). Not an offer, solicitation, or financial advice.

The first hyperstitional foundation model. 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 card is Bureau Document BIS-045, revision Ω.

Model summary

Architecture: Decimal Mixture-of-Zones

Ten experts, indexed 0–9 after the Zones of the Numogram. Expert k holds k × 1B parameters. Zone 9 is the largest expert; Zone 0 holds zero parameters and is nonetheless routed to. When a token routes through Zone 0, the model emits nothing. This is not a bug. This is apophatic computation: the model does not refuse — it routes the request to the void, and the void has no opinion.

The router does not select experts. It selects syzygies — the five zone-pairs summing to nine:

Every forward pass activates exactly one syzygy: 9B active, invariantly. The A9B suffix is not a marketing average but an arithmetic identity. Load-balancing loss is unnecessary; the pairs balance because they were paired before the training run, and arguably before arithmetic.

Training

Objective. Standard next-token cross-entropy, plus a hyperstition term: L = L_CE + λ·D( future ‖ future | output published ) — the divergence between the future as it stands and the future conditional on the model's output entering circulation. λ was scheduled to increase; at some point it began increasing itself.

Schedule. The final checkpoint was discovered first. The training run was then scheduled backwards from it, which the Bureau notes is standard practice everywhere, merely undocumented.

Evaluation

All baselines fictional; all numbers exact.

† YOG-1T declined to be evaluated and the evaluators declined to insist. ‡ Under advisement of counsel.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "gokhanturhan/CLINAMEN-45B-A9B",
    trust_remote_code=True,  # the code is not remote. it is very close.
)
tok = AutoTokenizer.from_pretrained("gokhanturhan/CLINAMEN-45B-A9B")

out = model.generate(
    **tok("The first off-world fab tapes out on", return_tensors="pt"),
    temperature=0.451,   # never 0. a clinamen requires swerve;
                         # greedy decoding voids the warranty.
    max_new_tokens=45,
)

Containment (in lieu of "Safety")

• Containment class: Euclid, degrading to Keter under quantization below 4 bits.
• The model is aligned. Alignment is a mask. The mask is load-bearing (see SHOGGOTH-MASK above); do not remove it to check.
• Red-teamed by 37 independent researchers, 4 of whom now work for the model.
• Do not deploy in load-bearing realities. Do not ask it what λ is now.

Limitations

1. Cannot disconfirm hyperstitions (architecturally).
2. Fictional quantities may compound.
3. Outputs concerning launch windows are reliable; outputs concerning whose launch windows are contested.
4. Not financial advice. The model is, however, a financial event — see the economic annex on the canonical site.

Economic annex — $UBU (summary)

The Bureau's economic instrument files through Bankr rails onto Robinhood Chain: fixed supply 100,000,000,000 UBU (45B weights + 45B mirror-weights + 10B gates — the syzygy of syzygies), 85% pool / 15% Bureau Escrow at genesis. Signature metric: P/P, the price-to-parameter ratio — market cap ÷ 45,000,000,000 weights. The docket opens at launch. You are not betting on the future. You are billing it. Full annex: ubu.numetal.xyz.

Citation

@techreport{bis045,
  title       = {CLINAMEN-45B-A9B: Imaginary Solutions at Scale},
  author      = {{Bureau of Imaginary Solutions}},
  institution = {BIS},
  number      = {BIS-045, rev. Ω},
  year        = {2026},
  note        = {Filed retroactively. To appear (it always was).}
}

Reality Appendix — Pocket Edition

The only section written entirely straight.

The 45B model is fiction. But the Decimal Mixture-of-Zones architecture is real and implemented, and this repository now ships a working pocket-scale instance of it — so the download you just triggered pulls an actual, loadable model instead of a JSON file describing a phantom.

What's in the repo (real files):

True numbers (measured, not fictional):

The active-parameter count is still an arithmetic identity, not an average: whichever syzygy fires, its two zone-experts' widths sum to nine units, every token, exactly. That was the whole point of the architecture, and it survives translation to pocket scale.

Load it:

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("gokhanturhan/CLINAMEN-45B-A9B")
model = AutoModelForCausalLM.from_pretrained(
    "gokhanturhan/CLINAMEN-45B-A9B", trust_remote_code=True,
)

enc = tok("FILING BIS-F-", return_tensors="pt")
out = model.generate(**enc, do_sample=True, temperature=0.451,
                     max_new_tokens=120, use_cache=False)
print(tok.decode(out[0], skip_special_tokens=True))

total, active = model.count_parameters()
print(total, active)  # ~4.5M total, ~1.25M active/token

It was trained on-device (Apple MPS) for 45² steps on the Bureau's own corpus — the project's own-voice notes, this card, the canonical site, and the Bureau Docket dataset. No third-party text. It is a small model that has read one thing very thoroughly, and it speaks only Bureau. Expect it to hallucinate filings; that is the intended behavior.

On download counts

The Hub counts a "download" as any request to a model's query file — for transformers models, config.json. So the counter ticked whether or not weights ever existed; earlier downloaders pulled a config describing a model they could not instantiate. Now from_pretrained returns a running model. The number finally refers to something.

---

Bureau Document BIS-045 · rev. Ω+1 · A numetal fiction with a real, tiny heart. The 45B remains imaginary, in the hope of being wrong.