tetracta/llm-xray-lesion-scans

Legacy Model X-Ray materials — withdrawn

Correction dated 6 September 2026.

The Model X-Ray results previously distributed from this repository have been withdrawn. These files are historical artifacts, not current evidence, and must not be used to support location, knowledge, portrait, lesion-response, simulated-quantization, quality, safety or deployment claims.

No replacement figures are published. Validation remains pending.

Correction record: https://www.tetracta.ai/model-xray/correction/

— Tetracta

📚 Docs: how scans work, what each panel means, API usage → https://www.tetracta.ai/llm_tomografi/docs

🎉 The X-Ray instrument is now in FREE open beta — for a limited time

The scanner behind this work is open to everyone: create an account and scan public models up to 7B — 20 free scans per account, no invite, no waitlist. It runs on our own local GPUs, and that capacity is the honest limit of the free beta — so if you want in, sooner beats later. → tetracta.ai/xray.html

LLM Lesion X-Ray — six 1B-class models, one protocol

Layer-lesion scans of six 1B-class language models, produced with a single pre-registered protocol. This repository contains measurements, not weights.

Two of the six models are a matched pair: a constant-state architecture and a transformer, trained by us on identical data with the same budget. On a standard benchmark they land 0.62 macro points apart (95% CI [−0.06, +1.28] — statistically equal). Under this instrument they are not remotely the same model.

---

Why this exists

A benchmark asks what does the model know. It cannot ask how is it built — not because nobody tried, but by construction: it only ever sees outputs.

This dataset is what the second question looks like when you answer it with a crude, hard-to-game instrument.

---

The instrument

For one layer at a time:

1. Record the intact model's next-token distribution over the full output space, on 12 fixed English prompts → p
2. Damage that layer — zero the smallest 5 / 10 / 20 / 30 % of its weights by magnitude, or add Gaussian noise at σ = 0.005 / 0.01 / 0.02 × weight standard deviation
3. Record the distribution again → q; score = KL(p‖q)
4. Restore the weights exactly, move to the next layer

Nothing is written to disk during a scan; checkpoint hashes were verified unchanged after every run.

Two quantities fall out:

quantity	meaning
fragility	how far the output moves per unit of damage
concentration	share of total damage carried by the most sensitive fifth of the depth — 20 % if information were spread perfectly evenly, approaching 90 % if a model hung everything on a couple of layers

---

The arms

file	model	weights
scans/qwen2.5-1.5b-base.json	Qwen2.5-1.5B	open
scans/qwen2.5-1.5b-instruct.json	Qwen2.5-1.5B-Instruct	open
scans/van-1b-base.json	in-house transformer, 10.32B tokens	not released
scans/van-1b-sft.json	same, after supervised fine-tuning	not released
scans/znext-v43-base.json	Z-Next v4.3, constant-state, 10.00B tokens	not released
scans/znext-v43-sft.json	same, after supervised fine-tuning	not released

van-1b and znext-v43 were trained by us on identical data. They are the matched pair; the only variable between them is architecture.

---

Headline numbers

At 30 % pruning, mean KL per layer (×10⁻³, lower is sturdier):

model	base	after SFT	concentration (base → SFT)
Z-Next v4.3	7.7	12.8	37 % → 39 %
in-house transformer twin	60.6	117.1	72 % → 69 %
Qwen2.5-1.5B	18.5	10.3	41 % → 52 %

Three things a reader may find useful — including one that argues against the authors:

1. The matched pair differs by 7.9× (base) / 9.1× (after SFT) in fragility and 37 % vs 72 % in concentration, while sitting 0.62 macro points apart (95% CI [−0.06, +1.28]) on a standard benchmark.
2. Concentration barely moves under fine-tuning (37→39, 72→69). Whatever sets it appears to be set by the end of pretraining.
3. Fine-tuning made Qwen more robust (18.5 → 10.3) and made both of our models more fragile. We publish this because it is what we measured.

---

What this dataset does not establish

Attributing an internal difference between two checkpoints to architecture requires independent-seed controls — several models from the same recipe with different seeds, to bound run-to-run variation. Those controls were not run for this release.

We know this failure mode first-hand. In July 2026 we published internal differences between two arms and retracted them after three purpose-trained control models showed the figures sat inside ordinary run-to-run noise.

The gaps here are 8–20×, not noise-band. That makes them more interesting. It does not make them controlled. Read every number as

"these checkpoints differ by X under this instrument"

and not as an architectural claim.

For the Qwen arms the limitation is stronger: data, scale and architecture all differ at once, and no reading of them can be decomposed.

---

File format

{
  "etiket":  "Z-Next v4.3-1B (TABAN)",
  "n_katman": 24,
  "n_istem":  12,
  "tohum":    1234,
  "budama":  {"0.05": [...], "0.1": [...], "0.2": [...], "0.3": [...]},
  "gurultu": {"0.005": [...], "0.01": [...], "0.02": [...]}
}

Field names are Turkish: etiket = label, n_katman = layer count, n_istem = prompt count, tohum = seed, budama = pruning, gurultu = noise. Each array holds one KL value per layer, ordered input → output.

To compare across architectures, normalise the layer index to 0–1 — the models have different depths and raw indices are not comparable.

Reproduce the headline numbers

import json, glob

for path in sorted(glob.glob("scans/*.json")):
    d = json.load(open(path))
    p = d["budama"]["0.3"]                       # 30% pruning profile
    n = len(p)
    fragility     = sum(p) / n
    concentration = sum(sorted(p, reverse=True)[: n // 5]) / sum(p)
    print(f"{d['etiket']:32s} fragility {fragility*1e3:6.1f}e-3   "
          f"concentration {concentration:5.1%}")

Plot a depth profile

import json, matplotlib.pyplot as plt

for name in ("znext-v43-base", "van-1b-base"):
    d = json.load(open(f"scans/{name}.json"))
    p = d["budama"]["0.3"]
    x = [i / (len(p) - 1) for i in range(len(p))]   # normalised depth 0–1
    plt.plot(x, p, marker=".", label=name)

plt.xlabel("normalised depth"); plt.ylabel("KL after damaging this layer")
plt.legend(); plt.show()

---

Example report

example/example_report_znext_base_to_sft.html is a real report from our X-ray service, produced by the unmodified production pipeline, comparing our base model against its fine-tuned version.

It answers a different question from the scans above: not "how robust is this model" but "what did fine-tuning change inside it".

field	value
risk	MEDIUM — clear functional drift, targeted eval advised
change begins at	at or before station 4 — our measured detection floor at this scale
effective width (N80)	see the linked report — the example was produced by an early (pre-rs-1.x) pipeline; N80 values from it are not comparable with current ps-1.1 / mv-1.2 scans
behaviour change	100 % of probes
deepest block vs shallowest	62×

The reading: fine-tuning reshapes the interface (upper layers), not the foundation.

Open the file in a browser. Want one of these for your own model? → https://www.tetracta.ai/xray.html

---

Protocol

PROTOCOL.md — including the reading rules and the declared failure condition ("if the arms are indistinguishable, that is a result and will be reported as one") — was written before the runs, not after seeing them.

---

Related

• Full research note — https://www.tetracta.ai/note-xray-lesion.html
• Run this scan on your own model — https://www.tetracta.ai/xray.html
• Live demo of the constant-state model, one consumer GPU — https://www.tetracta.ai/zchat
• Earlier matched-pair study with seed-null controls — https://huggingface.co/tetracta/llm-xray-twin-study-1b

Citation

@misc{tetracta_lesion_xray_2026,
  title  = {LLM Lesion X-Ray: six 1B-class models under one pre-registered protocol},
  author = {Tetracta AI},
  year   = {2026},
  url    = {https://huggingface.co/datasets/tetracta/llm-xray-lesion-scans}
}