Where instruction tuning landed in this pair — a propagation scan (N80 22/33, band 28–32)
We ran a before/after internal scan of mistralai/Mistral-7B-v0.3 → mistralai/Mistral-7B-Instruct-v0.3 with a fixed probe protocol (identical prompts for both models, controlled perturbations injected at seven depths, 168 matched probes per model) and thought the result was worth putting here rather than only on our own page.
Where the difference sits. 80% of the base→instruct difference mass falls in 22 of 33 stations (a station is the output of one block; 32 layers → 33 stations). The densest five-station band is 28–32, carrying 1.33× what a uniform spread would put there and 1.17× what a flat null would. The first observable difference is station 4 — that is our probe grid's detection floor (the earliest probe is injected at layer 2), not something the tuning earned.
Consistent with the other families we scanned, the band sits at the top of the network; the two Llama pairs in the same run instead concentrated mid-network, so the position is family-dependent.
Knowledge side. Factual recall barely moved: 1 of the 20 probe facts broke and 1 were repaired. What moved is behaviour on entities the model does not know — fake-name echo avoided went 9 → 16 of 20 (McNemar 9 improved / 2 regressed, exact two-sided p = 0.065), and a three-class judge reads refusals 1→1, echo 12→3, fabricated answers 7→16. Trajectory AUROC 0.910 → 0.963.
Data. Report: https://tetracta-model-xray-sample-reports.static.hf.space/reports/karsilastirma-mistral-7b-v0-3-mistral-7b-instruct-v0-3.html
Signed provenance/deletion attestation (its report_sha256 pins that exact file): https://www.tetracta.ai/llm_tomografi/attest/26088a080e004e238d065cb70f915e99
Machine-readable per-station profiles for this and 21 other interventions across six families: https://huggingface.co/datasets/tetracta/model-xray-gallery
If this does not match how the instruction tuning was actually done, we would genuinely like to know. The instrument is young, family differences like this are exactly what we cannot yet explain, and a correction from the people who trained the model is worth more to us than another scan.
Correction — 6 September 2026
We withdraw the Model X-Ray location, spread, concentration and derived-severity conclusions in the opening post. Any legacy knowledge-separation, portrait-visualization, lesion-response or simulated-quantization conclusion linked from it is also withdrawn and must not be used as current evidence.
We are publishing no replacement figures. Validation remains pending. Correction record: https://www.tetracta.ai/model-xray/correction/
— Tetracta