Twin study: new VG1 measurements

What the new measurements show

The last row contains algebraic self-checks, not independent scans. Counts describe these comparisons on the evaluated inputs; they are not population rates.

Read the seed controls before attributing the cross-family difference. Differences also occur within each family. These records support a combined artifact difference; they do not isolate an architecture effect, establish seed independence or rank model quality.

The original Vanilla and Rational checkpoints were each loaded and measured twice in separate processes. The other three checkpoints supply seed controls. The separate repeat-control reports are Vanilla and Rational. Each individual comparison report retains its own measurement summary; the study-level controls listed here have not been relabeled as a completed repeat-allocation summary inside that report.

Reports and controls

1. Vanilla, original seed → Vanilla, original seed — Same recorded observation: algebraic self-check
2. Vanilla, original seed → Vanilla, original seed — Same checkpoint: separate-process repeat control
3. Rational, original seed → Rational, original seed — Same recorded observation: algebraic self-check
4. Rational, original seed → Rational, original seed — Same checkpoint: separate-process repeat control
5. Vanilla, original seed → Rational, original seed — Current private research comparison
6. Vanilla, original seed → Vanilla, seed B — Current private research comparison
7. Vanilla, original seed → Vanilla, seed C — Current private research comparison
8. Vanilla, seed B → Vanilla, seed C — Current private research comparison
9. Rational, original seed → Rational, seed B — Current private research comparison
10. Vanilla, seed C → Rational, seed B — Current private research comparison

Scope and provenance

The published tensors were verified exactly against the pinned artifact revision. These are new measurements of historical checkpoints, not renamed July results. The reports use current VG1 numerical measurements and the Fable presentation through a private research integration; they do not add support for these custom research checkpoints to the ordinary customer selector.

Reports are unsigned research copies. The recorded differences are descriptive: no calibrated noise envelope, quality grade, knowledge score or customer acceptance is claimed. Historical withdrawals remain in force.

Exact published artifact revision · Measurement corrections and limits