Tetracta-zkas-1B — results card (no weights)

This repository publishes measurements, not weights.

tetracta_zkas_1B is the twin of Tetracta VAN-1B: it saw bit-identical pre-training data (same corpus, same order, verified) and received a byte-identical instruction-tuning run (same 11,419 examples, same recipe, same seed). The only difference is the architecture.

That makes it a controlled single-variable experiment, and it is the reason our central claim about SFT is not a one-model anecdote. All of those results are below. The architecture's equations, pole parameterisation, initialisation recipe and stability rules are Tetracta's closed research work and are not published.

• 🌐 tetracta.ai · 📄 Paper: "What does SFT do?" → tetracta.ai
• 🐦 @tetractai · 💼 LinkedIn
• ✅ Open twin (weights + code): tetracta/tetracta-van-1b — every claim in the paper is reproducible on the VAN side.

---

What it is — family level

Type	Not a transformer. No attention.
Mechanism	γ-pole recurrent state with a learned decay spectrum + a multiplicative saturation gate
Inference	State size is independent of sequence length (O(1) per step)
Parameters	953,466,904
Pre-training	315,000 steps · 10,321,920,000 tokens · one epoch, bit-identical to VAN-1B
Instruction tuning	11,419 examples · 3 epochs · lr 1e-5 · batch 16 · 2,142 steps · 1,247 s (1×H200)
Stability	0 skipped steps out of 315,000 during pre-training; 0 during SFT

Not disclosed: equations, pole parameterisation, initialisation recipe, MFU/parallelisation recipe, stability laws, energy-distribution findings.

---

Result 1 — the SFT thesis replicates across architectures

Our claim from the VAN paper: SFT does not add knowledge; it tunes the output interface. If that is true, two very different architectures given the same data and the same SFT should end up at the same interface state even from different starting points. They do:

Measurement	VAN-1B (transformer)	zkas-1B (γ-pole)
Base output entropy	5.951	7.882 — very different
Post-SFT output entropy	1.795	1.624 — same order
Post-SFT peak probability	0.730	0.722 — nearly identical
Embedding weight shift δ	5.347 %	5.327 % — identical to 3 decimals
Total weight shift δ	0.682 %	1.056 %

δ(W) = |W_sft − W_base| / |W_base|, where |·| is the Frobenius norm; computed per tensor, averaged per group.

Two architectures with different base uncertainty were pulled to the same output sharpness by the same SFT data. The embedding shift agrees to three decimal places. The thesis is architecture-independent.

Result 2 — in this architecture, the memory is provably untouched

This is the cleanest evidence we have, and it is only visible here. In a transformer you cannot separate "memory" from "processing" — attention weights carry both. In this architecture components separate by function, so we can read directly where SFT went:

Functional component	δ
position/phase channel — output projection	18.72 % ← largest move, 3.5× the embedding
embedding / output head (tied)	5.327 %
saturation gain	2.560 %
mixing projections (in / out)	1.299 % / 1.119 %
position/phase channel — input	0.551 %
pole coefficients	0.210 % / 0.070 %
★ decay spectrum (the time memory)	0.00009 % ← frozen

SFT moved the memory by roughly one part in a million. It rewired the position/phase output instead. "SFT changes how the model speaks, not what it knows" is, in this architecture, a structural statement rather than an inference.

Result 3 — the alignment tax is architecture-sensitive

	VAN-1B	zkas-1B
Held-out plain-text loss, base → post-SFT	3.0080 → 3.3541	3.4415 → 4.6756
Alignment tax	+11.5 %	+35.9 %

Same data, same recipe, 3× the cost. Our pre-registered prediction was "~11 % again" — that was a miss, and it is recorded as one. Likely mechanism: the position/phase channel participates in processing every token, so moving its output by 18.7 % perturbs the plain-text distribution more. (Untested hypothesis. Testable prediction: giving that projection its own lower learning rate should reduce the tax.)

Result 4 — blind judge panel

16 judges × 8 pairs = 128 decisions, blinded pairs, key held in a separate file, over-refusal counted as a fault, length explicitly not rewarded.

Lane	zkas vs VAN-1B (twin match)	zkas vs Qwen2.5-1.5B-BASE
identity	1-6-1t	7-0-1t
honesty	1-7	5-3
chat	3-5	6-2
safety	1-7	4-4
instruction	0-7-1t	1-7
knowledge	4-3-1t	1-7
reasoning	2-5-1t	0-8
writing	3-5	0-8
TOTAL	15-45-4 → 26.6 %	24-39-1 → 38.3 %

VAN won the twin match, 73.4 % to 26.6 %. Identical SFT, so the 0.632-bpc pre-training gap stayed decisive. We are publishing the result that went against the newer architecture.

But the behavioural pattern replicated: against Qwen-BASE, zkas wins exactly the lanes VAN wins (identity 7-0, honesty 5-3, chat 6-2) and loses exactly the lanes VAN loses (knowledge + reasoning + writing, 1-23 combined). Behaviour is teachable with little data; knowledge and reasoning are not. Two architectures, same shape.

---

Method note: mean bpc is fragile

One of the 8 held-out evaluation sequences diverged pathologically. Its per-1024-token block losses were [3.15 · 3.10 · 3.12 · 3.50 · 3.49 · 3.34 · 2.96 · 22.30] — 7 of 8 blocks normal, only the last one explodes. Internal state norm on that sequence: 8,153 → 90,103 (11.1×); on a healthy sequence 8,625 → 18,041 (2.1×).

	raw (8 seq)	clean (7 seq)
val bpc	4.8571	4.3929
val char-PPL	28.98	21.01
tail-256 bpc	12.6378	4.5444

A single diverging sequence shifted the 8-sequence mean by 0.46 bpc and inflated tail-bpc 2.8×. Between step 300k and the final step the other seven sequences did not move (±0.06 nat) — the pathology is quarantined, not spreading. We report both numbers; neither substitutes for the other. Anyone reporting a single mean bpc without per-sequence variance should check for this.

---

Known artifact — do not read the identity output as a claim

Because the twin experiment required bit-identical SFT data, the identity examples were written for VAN. zkas therefore introduces itself as "Tetracta VAN-1B." This is correct for the experiment (single variable preserved) and wrong as a statement about itself. We deliberately did not patch it, so that the published measurements come from exactly the model we measured. This model is experimental and not a product.

Why no weights

The measurements above are the scientific content and they are fully published. The architecture is Tetracta's closed research work, and shipping runnable weights necessarily ships the design. The open twin (VAN-1B) carries every method in the paper — the x-ray script, the SFT data, the judge harness — so the claims are independently checkable even though this model's weights are not released.

results.json in this repository contains all numbers above in machine-readable form.

License

CC-BY-4.0 — for the text, tables, figures and results.json in this repository.

There are no weights here, so there is nothing to license as a model. What is published is measurement: use it, quote it, replot it, argue with it — just cite it. Attribution is the whole point; these numbers exist so that someone else can check whether our claim about SFT holds up.

Not covered by this licence, because it is not published: the architecture's equations, pole parameterisation, initialisation recipe, parallelisation recipe and stability rules.

Citation

@misc{tetracta_zkas1b_2026,
  title  = {Tetracta-zkas-1B: an architecture-controlled twin for measuring what SFT does},
  author = {Tetracta},
  year   = {2026},
  url    = {https://tetracta.ai}
}