AI experiments.
The Frankenstein workshop.
Exploring models and bridges to build the comic, the game and Montpellier ECU. An experiment remains an experiment until its result has been checked.
A method with human validation.
Describe the goal, propose a change, execute it with available tools, and check the result. Mehdi retains creative direction and delivery approval.
“Frankenstein” refers to attempts to connect Codex Windows, local bridges and different models. GLM-5, Kimi K3 and GPT-OSS-120B appear in the technical journal; mentioning them does not establish that a service is currently running.
What the journal records.
Experiments cover tool calls, tool-name normalization, bridge protocols and the difference between a configuration file and a running process.
The September 26 reconstruction separates reported observations from unresolved questions. “Frankenstein 1/2/3” are not established as formally published releases.
Measurements need evidence.
Early tables propose latency figures and success percentages, but these documents provide no associated test protocol or raw results. They are not a validated model comparison.
A useful benchmark must specify the task, exact model, configuration, sample count, raw results and success criteria.
Benchmark status ↗