MINI-ROOK / BUILD LOG 002
A memory with a paper trail.
We gave our local AI a memory
Historical development snapshot. Some features may have changed since publication; this entry preserves its original checkpoint.
TECHNICAL CHECKPOINT
What this build actually records.
- Stack and method
- Python · Markdown retrieval · keyword overlap · numbered references
- Recorded progress
- A human-observed answer used project notes and acknowledged missing information.
- Still open at this checkpoint
- Citation faithfulness, adversarial notes and broader grounding quality remain test gates.
Software development on an existing workstation. This entry does not establish a new robot, custom hardware build or autonomous publishing capability.
01 / BUILD LOG 002
The first observed memory test
The second development log records a human-observed memory test. Asked about Rookweld and CharacterForge, the local qwen3:4b model used curated project notes and replied in Swedish.
It identified Rookweld's workshop identity and CharacterForge's documented Windows preview. It also said that the retrieved note did not establish the current commercial sales status. That missing-information boundary was the useful result.
The CharacterForge statement describes what the note established at that time. It is not the current product status; the CharacterForge page carries the newer release and checkout information.
Current CharacterForge release →02 / BUILD LOG 002
How the notebook works
The Python backend reads a deliberately curated knowledge directory of Markdown notes. Simple keyword overlap ranks text sections; relevant excerpts are sent to Ollama alongside the question. The model is instructed to cite numbered references.
This is read-only document retrieval. It is not unrestricted filesystem access, persistent conversation memory or live system monitoring. The references remain plain text rather than clickable citations.
03 / BUILD LOG 002
One test is a beginning
The log records a basic pass. It does not establish a hallucination rate, citation faithfulness across questions or resistance to adversarial instructions embedded in documents.
The next useful checks are grounded questions, missing answers, unrelated questions and attempts to make retrieved text act as instructions. Any future quality claim needs evidence across those cases.
04 / BUILD LOG 002
Next: Workshop Awareness
An explicit allowlist of project documentation is being implemented, beginning with CharacterForge and Mini-Rook. Reading a README must not become permission to scan arbitrary directories, run commands or claim current system state.
At the Build Log 002 checkpoint, the awareness implementation was still marked unverified. Source inspection shows the integration in code; that is not a fresh runtime pass. Persistent chat, authenticated mobile access and a draft-and-approval content queue remain roadmap items.
KEEP FOLLOWING THE BIRD