Agentic DMX Master Plan

Goal

Build the repo-local agentic DMX skill system in small, reviewable merge requests.

Ordered Sub-Issues

• #72 Scaffold repo-local dmx-expert-orchestrator skill
• #73 Define intake workflow for the orchestrator skill
• #74 Add default lightweight EDA guidance for user-supplied data paths
• #75 Write structure-first hierarchy and data-shape routing reference
• #76 Encode the “joint model first” routing policy
• #77 Add keys and parameter-sharing reference
• #78 Add posterior reweighting and conditional reuse reference
• #79 Add initialization, sufficient-stats, and pseudo-count reference
• #80 Distill dmx_basics_conditional_vs_composite_mixture.ipynb
• #81 Distill dmx_example_process_sequences.ipynb
• #82 Distill dmx_example_reduced_search_space.ipynb
• #83 Distill dmx_variable_sequence_length_search_depth.ipynb
• #84 Distill dmx_basics_mixture_models.ipynb
• #85 Distill dmx_advanced_plsi.ipynb
• #86 Distill dmx_advanced_geotweet.ipynb
• #87 Refactor dmx-local-modeling to rely on the new references
• #88 Add compute and scaling reference, including torch_stats guidance
• #89 Create a small benchmark prompt suite for agentic DMX tasks
• #90 Define an evaluation rubric for “DMX expert” behavior
• #91 Audit src/dmx/stats docstrings for keys
• #92 Improve highest-priority src/dmx/stats docstrings for keys
• #93 Audit src/dmx/stats docstrings for sufficient stats and pseudo-counts
• #94 Improve src/dmx/stats docstrings for initialization semantics
• #95 Add repo-curated canonical examples reference
• #96 Run benchmark review for orchestrator routing behavior
• #97 Run benchmark review for post-fit and downstream-task behavior

Notes

• One MR per child issue.
• Land work in the listed order unless a dependency changes.
• Open follow-up issues instead of broadening an existing child issue.

Start by reviewing the ordered child issues and their completion status, especially the remaining benchmark reviews #96 and #97. Use the completed orchestrator, modeling, reference, and evaluation work as context. The plan is complete when the benchmark reviews for routing behavior and post-fit/downstream-task behavior are finished or have produced clearly scoped follow-up issues.