Model X-Ray today (status of 11 November 2026)

Status on 11 November 2026: the correction is applied and the beta is open. The open beta reads public models from Hugging Face by name, up to the 7B class, and scans them on our own GPUs, free within a monthly allowance of 20 browser scans and five distinct source models per account. The eligible repositories are listed on the scope page. A 7B-class model runs as a quantization simulation of that one model. Larger models are scanned on request, outside the free beta. Results are research-stage and payments are disabled. This card's presence is not an eligibility grant.

Earlier status entry, 22 September 2026: the 7B class opened on 21 September 2026, and scans moved to Tetracta-operated local GPU workers; no cloud GPU is used in this beta. SmolLM3 (HuggingFaceTB/SmolLM3-3B-Base and HuggingFaceTB/SmolLM3-3B) was added on 22 September 2026; the instrument release of that date did not support this tokenizer contract, so write to us before submitting one.

A report describes its recorded artifacts and measurement conditions; it is not a model-quality ranking, safety certificate or deployment verdict. A quantization simulation does not create a deployable quantized model. Knowledge scans remain unavailable and earlier withdrawn claims remain withdrawn. Hallucination, undertraining and overtraining remain research questions, not measured product features.

Service status · Current scope and report guide · Correction record · Report examples

Legacy Model X-Ray materials — withdrawn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The Model X-Ray results previously distributed from this repository have been withdrawn.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.

The withdrawn historical interpretations remain withdrawn.

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

— Tetracta

Model X-Ray gallery: measurements, not weights

Signal-propagation measurements of public language models produced by the Tetracta Model X-Ray instrument (an earlier probe set, scan configuration and metric version, not the current report contract; contributor-normalized difference profiles). Every record links to a signed deletion/provenance attestation (attestation field) whose report_sha256 pins the exact report the numbers come from. Probe texts are withheld (they are the instrument); everything else needed to recompute the summary statistics is here.

Families: HuggingFaceTB, Qwen, TinyLlama, allenai, meta-llama, microsoft, mistralai, tiiuae · Portraits: 14 · Before/after and simulated-quant pairs: 22 · Knowledge probes: 17

What the numbers mean (short)

• focus — participation ratio of the difference profile at an early strike position, normalized by the depth remaining after the strike, averaged over all strikes; lower = a nudge stays contained, higher = it spreads. Test-retest ±0.7% (10 seeds, 1.5B-Instruct).
• direction_selectivity — how differently the network reacts to opposite nudge directions (±17% test-retest; read as a coarse band).
• persistence — matched-subset per-strike ratio of the response 10 layers after the strike vs 1 layer after (±1.2%).
• difference_profile (pairs) — per-station mean difference between model A and model B under identical probes, each station averaged only over the probes that can reach it (contributing_probes). n80_stations = how many stations hold 80% of the difference mass; change_start_station ≤ 4 is the detection floor (the first probe is injected at layer 2), not a property of the fine-tune.
• Knowledge records: per-item correctness and output-level p(answer) on 20 known facts and 20 fabricated entities; the three-class judge (refusal / echo / fabricated answer) is what "avoidance" actually measures.

Gallery (portraits)

Pairs

Knowledge probes

Provenance

• Each JSON carries job, scan_date, hardware (own workstation GPUs for small models; single-tenant cloud pod for large ones, pod identity in the attestation) and the attestation URL.
• Reference bands and the interpretation layer are described in the sample reports: https://huggingface.co/spaces/tetracta/model-xray-sample-reports
• Earlier measurements under older probe-set and metric versions are not comparable with these and are kept only as superseded comparisons in the Space.
• Corrections history and reviewer exchange: see the Space README ("Correction" and "Measurement update").

Produced by our own export script from the product's scan archive. Licence CC-BY-4.0 for the measurements.

History: earlier status notes

VG1 status · 12 September 2026 (Europe/Istanbul) (superseded by the status above)

VG1 has been deployed. Customer scan-to-report acceptance is still pending.

The free first-beta scope is eligible public models up to and including 7B, within the account allowance, for checkpoint comparisons and in-memory quantization simulations. Customer jobs run on RunPod. The service selector defines the supported revisions and scan types. Later Pro access above 7B requires an explicit grant and remains limited to the models and allowance enabled for that account.

This dataset remains a withdrawn historical gallery. The retained files and tables below are not current-release evidence. Deploying VG1 does not reinstate their interpretations or turn them into new customer reports.

A report describes its recorded artifacts and measurement conditions; it is not a model-quality ranking, safety certificate or deployment verdict. A quantization simulation does not create a deployable quantized model. Knowledge scans remain unavailable and earlier withdrawn claims remain withdrawn. Hallucination, undertraining and overtraining remain research questions, not measured product features.

Service status · Current scope and report guide · 12 September correction

Separate new Twin study · 12 September 2026. Read the current VG1 reports. Seven fresh scans of five checkpoints support two independent repeat controls, four within-family seed comparisons and two cross-family comparisons; two algebraic self-comparisons are additional consistency checks. The fresh repeats were unchanged, while the within-family and cross-family comparisons changed in measured response and generated token sequences. These ten comparisons share seven scans: they do not establish an architecture effect or model-quality ranking. The historical gallery and its withdrawal below remain unchanged. Pinned report archive.