VIRIDIAN / RESEARCH LIBRARY

Research library.

Research on AI evaluation, experimental progress and semiconductor implementation. Each record distinguishes observed results, interpretation and limitations.

Technical manuscripts, exploratory executions and internally reported case studies are labelled individually. None is presented as peer-reviewed certification.

5 research records

K2 / TRIAL 001AI training & progress

When Fine-Tuning Made a 5B Model Worse

Revised evidence review · Publication draft

Reported exact-pass score fell from 238/300 to 106/300 after one QLoRA intervention. Paired change: −44.00 pp.

Boundary. One seed and a narrow frozen distribution. Audit bundle, case and manifest identities remain unresolved.

TECHNICAL STUDY 003AI evaluation & assurance

Presentation / Serialization Sensitivity

Exploratory execution · Not peer reviewed

The exact tested wrapper preserved recorded winners on all 20 reused response pairs.

Boundary. All conditions shared one judge conversation. Fresh-context isolation was specified but not followed.

TECHNICAL STUDY 002AI evaluation & assurance

Evaluator Contract Stability

Exploratory execution · Not peer reviewed

No recorded winner disagreements across the tested rubric formulations on 20 reused pairs.

Boundary. Shared judge state may suppress disagreement. The record does not establish independent-context rubric robustness.

TECHNICAL NOTE 001AI evaluation & assurance

Response-order sensitivity in LLM judging

Technical manuscript · Not peer reviewed

A/B reversal preserved the preferred underlying response in 19 of 20 comparisons.

Boundary. One pair showed a second-position pattern; no systematic directional bias or general robustness is established.

SILEX / GCDSemiconductor & EDA

Semiconductor Physical Design & EDA Research

Retained engineering research · Not production qualified

Ten sampled worst slow-corner setup paths shared an eight-cell structural segment.

Boundary. The sample covers ten of 53 violating endpoints. Connectivity and path-level timing evidence remain incomplete.

Publication policy

Unfinished studies are not listed as completed findings. Three planned Adamantine studies remain in unpublished authoring templates, with no invented results or evidence bundles.