Mitchell's Log

A working notebook for software projects, experiments, and things worth writing down.

Recent Entries

Filter by one or more tags.

Showing 9 of 9 posts.

  1. Writing Experiment Rerun: gemma3:1b with 3 Samples per Story

    Comparison of 2-sample vs 3-sample runs, with focus on all-bad story rate versus at-least-one fair/good story rate.

    realm writing experiments slm gemma3-1b evaluation

  2. Writing Experiment Rerun: gemma3:1b at 82.0% Fair+Good

    Updated gemma3:1b writing experiment with 100 outputs, 82.0% fair+good, case-level usability rates, and the full rerun quality rubric.

    realm writing experiments slm gemma3-1b evaluation

  3. Writing Experiment: gemma3:1b vs qwen3:1.7b Quality Mix

    Head-to-head writing quality distribution with stacked percentage columns and request-compliance summary.

    realm writing experiments slm gemma3-1b qwen3-1-7b evaluation

  4. Paper Notes: Narrative Reversals for Voice-Stable Context Retrieval

    Notes on narrative reversals, plus a design for reversal-aware segmentation and retrieval that is optimized for maintaining voice and pacing during drafting.

    paper-notes narrative storytelling context-compiler retrieval realm writing segmentation

  5. REALM Combined Loop: Multi-Size Results Toward Context Compiler

    Combined Read + Edit/Analyze experiment showing strong multi-size token behavior for the context compiler path.

    realm ctxc context-compiler slm gemma3-1b local-inference experiments ai

  6. Context Compiling: ctxc and vectorless builds

    I’m building a ‘context compiler’ that walks large docs like a book and emits a tight, testable context packet—without embeddings—so even tiny models can rel...

    context ctxc realm rag compilers writing

  7. REALM Continuation: Local SLM Evaluation with gemma3:1b

    Continuation experiment evaluating gemma3:1b as a local SLM for REALM context curation.

    realm slm gemma3-1b local-inference experiments ai context-curation

  8. Context Curation: Preliminary REALM Tests

    First experiment write-up validating baseline navigation behavior and context scaling.

    realm experiments context-curation ai

  9. REALM: Read Edit/Analyze Loop Monitor

    A modular architecture that curates context from large documents at low cost, with a path to small on-device models.

    realm architecture context-curation local-first ai