VEZRYN LLC / APPLIED AI RESEARCH
Keep the evidence.Including the failures. A technical record of what we tried, what the measurements showed, and what remains unresolved.
Our research approach ↗ RESEARCH RECORD 001 REV. 06 Oct 2026
POST-TRAINING / GRANITE 4.1 8B / COORDINATION
What a sequence of failed post-training experiments taught us about DPO, LoRA, numerical execution, finite-response instability, and behavioral interaction.
No clean successor has beaten the protected incumbent.
Read research record ↗ CANONICAL COMPARISON
224 0050 Rejected candidate
/300 −19 points under current canonical accounting
Evidence 0050 / 0051
RECORD RR-001
SYSTEM Granite 4.1 8B
METHOD LoRA / DPO
OUTCOME Rejected continuation
EDITION 0.1 EXPERIMENTAL CHRONOLOGY
Eighteen experiments.An unchanged incumbent. From a rejected continuation to investigations of direction, execution and interaction. The changes in understanding belong in the record.
Open experiment archive ↗ IDENTIFIER INVESTIGATION RESULT / FINDING
0035/Q1 PROTECTED INCUMBENT 243/300 Current canonical accounting. Retained unchanged.
↗ 0050 CONTINUATION 224/300 Rejected one-step candidate
↗ 0052 INTERPOLATION Rejected Smaller steps on the tested line failed preservation
↗ 0053 FORENSIC Direction changed Masking did not localize parameter movement
↗ 0056 DIAGNOSTIC 4/12 regress At each radius, executable response contradicted local improvement
↗ 0066 FORENSIC Material interaction Pairwise non-additivity; fixed quadratic model insufficient
↗ 0067 ABLATION No supported update No tested removal safely repaired the full failure at both radii
↗ Selected records from 0050–0067. Later diagnostics are teacher-forced measurements, not successor benchmark scores.