Twin Loss Curves, One Operator Apart
Legacy Model X-Ray matched-pair note.
This page is not current evidence
The former article, figures and interpretations at this URL have been withdrawn. They must not be used to support a Model X-Ray capability, scientific, model-quality, safety, performance or deployment claim.
This URL remains available as a transparent citation anchor. Its former result is not restored by the VG1 beta.
Current VG1 documentation · 12 September 2026
For checkpoint comparisons and in-memory quantization simulations, read the current scope and report guide. Sample report gallery links to versioned examples; use each example only within its own stated limits.
This historical URL is not a new measurement or a replacement sample. See first-beta access and model coverage →
New Twin measurements · 12 September 2026
Read the new VG1 reports and comparison table →
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 ten comparisons share the same seven scans.
The fresh repeats were unchanged. Both the within-family and cross-family comparisons changed in measured response and generated token sequences. These results do not isolate an architecture effect or rank model quality. They do not reinstate the withdrawn article above or demonstrate customer scan-to-report acceptance.