Model X-Ray · Scope & limits · Correction record
Tetracta · dated correction recordModel X-Ray correction
Current status · 22 September 2026
7B class open. The 7B class opened on 21 September 2026. The eligible public repositories listed on the scope page (twenty-one exact Apache-2.0 revisions as of 22 September 2026, up to the 7B class) can be scanned by any registered account within the free allowance: 20 browser scans and five distinct source models per calendar month. 7B-class checkpoints run as single-model quantization simulations; a pair of two 7B-class checkpoints is currently accepted by the form but fails in the worker (under investigation, 22 September) — please do not submit one until this note is removed. Scans run on Tetracta-operated local GPU workers; no cloud GPU is used in this beta. A scan reads safetensors weights under their canonical file names; a repository that publishes only a pickle checkpoint or another format (for example GGUF) is declined before a job is created and does not use the allowance. Models above the 7B class are not scanned in this beta. External customer acceptance remains pending; results are research-stage and payments are disabled.
The withdrawn historical interpretations remain withdrawn. Current descriptive reports do not establish general localization, knowledge separation, hallucination detection, training adequacy, quality or safety.
Current status · 15 September 2026 (superseded by the 22 September entry above)
VG1 is deployed. Internal end-to-end operational acceptance completed on 15 September 2026; external customer acceptance is still pending. The first-beta scope is checkpoint comparisons and in-memory quantization simulations for eligible public models up to and including 7B, within the account allowance. Service status and access.
The withdrawn historical interpretations remain withdrawn. Current descriptive reports do not establish general localization, knowledge separation, hallucination detection, training adequacy, quality or safety.
Older availability statements and the seven-day estimate for larger models are superseded by the current beta scope.
12 September 2026 · follow-up on the September gallery
We rechecked the retained gallery evidence in response to Dipankar Sarkar's review. The initial unobserved region in the historical comparisons came from scanner coverage. It was not evidence that the checkpoints first changed there. Describing that boundary as a measured sensitivity floor was wrong.
The earlier start and spread interpretations remain withdrawn for those records. The current calculation distinguishes an unmeasured position from no observed difference and holds the contributing set consistent across the evaluated region. These corrections do not reconstruct measurements absent from the historical scans.
The descriptive concentration calculation uses measured coverage and handles ties and absent signal explicitly. It does not reinstate the old severity verdict or establish a quality ranking between models.
Knowledge remains withdrawn. The archived paired outcomes do not support the earlier fabrication-avoidance interpretation. We cannot reproduce the first trajectory AUROC from its original item scores, which are absent from the retained record. Later scans are not substitutes for those missing scores.
The later controlled small-model studies remain bounded to their recorded conditions, including recorded misses. They do not establish general onset localization or added usefulness over simpler comparisons across arbitrary fine-tunes.
VG1 retains the updated report presentation and calculation boundaries. Operational beta acceptance does not restore the historical scientific claims. Hallucination, undertraining and overtraining remain separate research questions, not measured product features.
Historical gallery revision · Current scope and report guide · Sample report gallery
Read the original correction and updates · 4–7 September 2026
The dated text below is preserved as history. The 12 September follow-up and current scope above control its interpretation; the old “current status” headings and availability estimates describe their original dates.
Current status (7 September 2026, Europe/Istanbul)
| Item | Status |
|---|---|
| Withdrawn claims of 4 September (structural location numerics, knowledge separation, portrait, lesion-response, legacy simulated quantization) | remain withdrawn, except onset depth restored on 7 September in calibrated categorical form |
| Corrected result set (6 September) + Part 2 (7 September) | published — sample reports on Hugging Face; scope and limits on the Scope page |
| Release validation (V1 acceptance) · new scans | still open · open beta since 7 September 2026 (public models below 7B; free allowance of 20 scans per account; models of 7B and above in about one week) (superseded 21 September 2026: 7B class open) |
Model X-Ray correction · historical text
7 September 2026 (Europe/Istanbul) — Onset depth restored under calibration. The 6 September note said that reports state no location. After a pre-registered control experiment (pre-registration identity cd0171cf; criterion frozen before any result was seen; exact onset predicted and observed in 6 of 6 controlled cells on each of three public models (Qwen2.5-0.5B-Instruct, SmolLM2-1.7B, TinyLlama-1.1B-Chat-v1.0); cross-board agreement on onset 4 of 4 cells), reports produced under instrument release xray_vg1 now state the decoder-block output at which a recorded difference first appears, as a categorical field with the calibration and scope limits published on the Scope page. No other withdrawn claim is restored; the earlier numerical location statements remain withdrawn. Minor correction in the same update: the sample-report table's TinyLlama rows said "configs differ"; the recorded configuration comparison for that pair is identical (the withheld output-text field follows from differing tokenizer semantics); no measured field changed.
7 September 2026 — Part 3: the reported onset matched the first fine-tuned block in 3 of 3 controlled fine-tune cells; added value over plain output comparison is not yet shown.
Update (6 September 2026, 19:39 TRT): corrected results (A/A null control 0/16; four re-scanned base→instruct pairs) are published on Hugging Face; sample reports and full tables: https://huggingface.co/spaces/tetracta/model-xray-sample-reports.
Status reviewed 4 September 2026 · validation pending
What changed
We have withdrawn our earlier structural-location, knowledge-separation, portrait-visualization, lesion-response and legacy simulated-quantization findings. An adversarial re-audit found defects or missing validation controls that prevent those results from supporting the interpretations previously attached to them.
Existing example reports, articles and related figures are historical records, not current evidence. They must not be used to make research, product, safety or deployment decisions.
Why the earlier evidence is insufficient
The re-audit identified the following method-neutral defect classes. Not every class applies to every historical item, but together they prevent the former interpretations from being treated as current evidence.
- Release identity and version control: public artifacts and interpretations were not consistently bound to one verified release state.
- Construct and criterion validity: some observations were used to support broader location, knowledge or robustness interpretations without an adequate independent criterion.
- Controls, repeatability and transfer: the available controls and replications were not sufficient for the scope of the public conclusions.
- Execution fidelity: legacy simulated-quantization observations cannot establish behavior under a real deployment execution path.
- Reporting sensitivity: presentation-format effects, uncertainty, abstentions and failure boundaries were not excluded or stated consistently enough.
Current evidence status
| Earlier claim family | Current status |
|---|---|
| Structural location and derived severity | Withdrawn. No replacement result is published. |
| Knowledge separation and internal readout | Withdrawn. No replacement result is published. |
| Portrait visualizations as validated evidence | Withdrawn. No replacement protocol has passed validation. |
| Lesion-response interpretation | Withdrawn. No architecture or robustness conclusion is current evidence. |
| Legacy simulated-quantization interpretations | Withdrawn. A simulation cannot establish behavior, quality, safety, performance or deployment readiness. |
| New Model X-Ray result set | Validation pending. No new public result is approved. (as of 4 September 2026; superseded — corrected results were published on 6 September and extended on 7 September 2026, see "Current status" above) |
Affected public locations
| Location | Disposition |
|---|---|
| https://www.tetracta.ai/, https://www.tetracta.ai/research.html, https://www.tetracta.ai/xray.html | Revised to the current evidence boundary. |
| https://www.tetracta.ai/note-xray-autopsy.html | Withdrawn citation shell. |
| https://www.tetracta.ai/note-xray-twin-study.html | Withdrawn citation shell. |
| https://www.tetracta.ai/note-xray-lesion.html | Withdrawn citation shell. |
| https://www.tetracta.ai/note-noise-floors-data.html | Withdrawn citation shell. |
| https://www.tetracta.ai/note-sft-anatomy.html | Withdrawn citation shell. |
| https://www.tetracta.ai/sample-report-comparison.html | Withdrawn sample-report citation shell. |
| https://www.tetracta.ai/sample-report-mistral.html | Withdrawn sample-report citation shell. |
| https://www.tetracta.ai/sample-report-qwen-0-5b.html | Withdrawn sample-report citation shell. |
| https://www.tetracta.ai/note-chimera-merge.html, https://www.tetracta.ai/note-token-economy.html, https://www.tetracta.ai/note-language-generation.html, https://www.tetracta.ai/note-kimi-language-tax.html | Non-X-Ray results retained; Model X-Ray dependency or sales language removed. |
| https://www.tetracta.ai/znext.html, https://www.tetracta.ai/legal.html | Legacy X-Ray interpretation or operational assurance removed. |
| Former directly addressable Model X-Ray figures | Removed from public presentation and retired with HTTP 410. |
What happens next
Replacement figures will be published only after a frozen, preregistered validation program completes held-out controls, repeatability and transfer checks, an adequately powered follow-up, independent review and external validation. Any future public release will state its intended use, uncertainty, abstentions, limitations and out-of-domain scope.
Model X-Ray remains research-stage. It is not a diagnostic, a quality or safety grade, a certification, or a substitute for task-specific evaluation.
Service-status boundary
Status (6 September 2026; superseded by the 7 September update below): new scans are paused while validation is pending. Any reopening will be unpaid and invite-only, and will be announced with a dated update here.
Update (7 September 2026, Europe/Istanbul): scans reopened as an unpaid open beta. Any registered account may run up to 20 scans on public models below 7B; models of 7B and above open in about one week. Release validation (V1 acceptance) remains open. (7 September 2026 — superseded: the 7B class opened on 21 September 2026; see the current status above.)
No scan order, paid-plan activation or payment action is offered on this page.
Revision note: this notice supersedes earlier Model X-Ray evidence claims; it does not erase the historical record. If a linked legacy item remains accessible, this correction controls its current interpretation.