Report 14 (research run): explanatory text aligned with the current report language (decision rule description, list of what is withheld); dated note added; every measurement field unchanged
Correction 23 September 2026: the weight tables of reports 12 and 14 order the two Mistral arms by weight distance; the note's 'nothing finer' sentence is corrected, crediting Dipankar Sarkar
Correction to the 22 September note: the three controlled fine-tune pairs in the 7 September archive are intermediate counts (13, 19 and 6 of 21); the earlier 'never published an intermediate count' sentence was wrong
Note 22 September 2026: what our published count does not do (every count under the current contract in one table; defects with dates; catalog provenance)
Corrections 22 September 2026: report 06 relabelled validation-pending; report 14 row no longer quotes a position count it does not publish; catalog page status aligned with README; archive banner names the 12 September correction
Seven new research scans and ten comparisons. Preserve six current retained examples and thirteen historical archive files. No customer acceptance or architecture-quality claim.
Publish current VG1 report examples and distinguish historical archive
Six genuine retained measurement examples rendered and checked with current AP8 customer-report code; original measurement times and evidence preserved. Thirteen historical report bodies retain their bytes with a visible archive notice. No new GPU measurements, signatures, customer acceptance or risk grades claimed.
Align X-Ray cards and metadata with current VG1 status
Clarify customer acceptance pending, eligible RunPod beta scope and historical evidence. Preserve weights, measurements, reports and unrelated model research. No new measurements or customer success claimed.
Clarify deployed VG1 status and pending customer acceptance
Update entry documentation only. Preserve all existing reports, measurement records, historical text and repository metadata. Link the dated correction and current scope. Customer scan-to-report acceptance remains pending.
Part 3 (7 Sep 2026): controlled fine-tune onset and output-comparison limits
