MINI-ROOK / BUILD LOG 013
Keep the Post, Lose the Self-Review
Separate publication copy from the model's unsolicited self-evaluation
TECHNICAL CHECKPOINT
What this build actually records.
- Stack and method
- Growth instructions · conservative output separation · curated Quality Gate warnings
- Recorded progress
- Python syntax, synthetic output separation and limited human Ollama/UI smoke test pass.
- Still open at this checkpoint
- Availability, privacy/cost and other claims require evidence; clean output is not factual accuracy.
Software development on an existing workstation. This entry does not establish a new robot, custom hardware build or autonomous publishing capability.
TEST STATUS / SEPARATE FROM THE ORIGINAL RECORD
What later checks established.
Python syntax and a synthetic post-plus-self-review separation check passed. A later user-observed browser test with Ollama produced a coherent Reddit-style post without the trailing self-evaluation, while Quality Gate warnings appeared separately. This is a limited output-separation smoke test.
Still open: The draft still incorrectly denied download availability. The creator-confirmed product brief was updated; the log did not independently verify its direct download endpoint. Unsupported privacy/cost and other factual claims still require review. No automated publication or general accuracy pass is established.
01 / BUILD LOG 013
The original development record
A human Growth test showed the local model writing a Reddit post followed by a long, unsolicited explanation of why it believed the post met the requirements. It also produced suspect claims about privacy and costs. The inline Quality Review panel from Build Log 012 did work in the browser, with distinct blocking and manual-review sections.
This iteration strengthens Growth's output instructions and adds a conservative output-separation function: recognized self-review headings and introductory boilerplate are removed from the public-facing post response. The initial Growth response now also produces Quality Gate warnings from the same curated rule-based checks used in editorial review. This does not make an unverified post safe to publish, nor does it replace human approval.
Verified: Python syntax checks PASS; a synthetic reply containing a post followed by a self-review was reduced to the post alone. A subsequent user-observed browser test with Ollama produced a coherent Reddit-style post without a trailing self-evaluation essay, and displayed Quality Gate warnings separately. This is a limited UI smoke-test PASS for clean output separation, not general factual accuracy. The draft incorrectly said no download was available, despite the creator confirming a download exists on the CharacterForge page. The creator-confirmed publishing brief has been updated; the direct download endpoint has not been independently checked.
Status: READY TO PUBLISH — limited human UI smoke test PASS; publishing facts and unsupported claims still require review. Chapter V — The Quality Gate. No automated publication.
KEEP FOLLOWING THE BIRD