This folder collects small, reproducible research-style demos used to showcase omx autoresearch on harder optimization problems.

Design goals:

• deterministic or seed-controlled evaluations
• small code footprint
• no large datasets checked into git
• no heavyweight runtime artifacts committed
• evaluator-driven keep/discard loops that are easy to inspect under .omx/logs/autoresearch/

• playground/* — demo code and benchmark logic
• missions/* — autoresearch mission contracts used by the showcase
• scripts/eval-* — focused evaluator entrypoints
• scripts/run-autoresearch-showcase.sh — convenience launcher for the bundled showcase missions

Naming convention:

• playground demo dirs use *_demo when they are standalone benchmark implementations
• mission dirs use problem/task slugs under missions/
• evaluator scripts follow scripts/eval-<mission-slug>.(js|py)
• the showcase runner maps short names (for example bayesopt, latent, sorting) to mission dirs

Use scripts/run-autoresearch-showcase.sh --list to see the bundled launch targets, or run one or more showcases directly with the wrapper script.

• Mission: missions/in-action-cat-shellout-demo/
• Evaluator: scripts/eval-in-action-cat-shellout-demo.js
• What it demonstrates: a tiny self-hosted code optimization loop where autoresearch removes an unnecessary shell-out from OMX itself.

• Demo code: playground/ml_kaggle_demo/
• Mission: missions/ml-kaggle-model-optimization/
• Evaluator: scripts/eval-ml-kaggle-model-optimization.py
• What it demonstrates: model-family / hyperparameter search on a deterministic tabular classification benchmark with a score-improvement keep policy.

• Demo code: playground/bayesopt_highdim_demo/
• Mission: missions/noisy-bayesopt-highdim/
• Evaluator: scripts/eval-noisy-bayesopt-highdim.py
• What it demonstrates: a harder black-box optimization task with noise, limited evaluation budget, and curse-of-dimensionality pressure. The successful autoresearch run switched from random search to a subspace-aware fixed-kernel GP with denoised incumbent selection.

• Demo code: playground/bayesopt_latent_discovery_demo/
• Mission: missions/noisy-latent-subspace-discovery/
• Evaluator: scripts/eval-noisy-latent-subspace-discovery.py
• What it demonstrates: a follow-on harder variant where useful structure is mixed through latent directions rather than directly exposed as obvious coordinates.

• Demo code: playground/adaptive_sort_demo/
• Mission: missions/adaptive-sort-optimization/
• Evaluator: scripts/eval-adaptive-sort-optimization.py
• What it demonstrates: algorithm-engineering optimization over a deterministic mixed-distribution sorting benchmark using weighted comparison/move cost rather than noisy wall-clock timing.

Example:

omx autoresearch missions/noisy-bayesopt-highdim

Then inspect:

RUN_ID=$(find .omx/logs/autoresearch -maxdepth 1 -mindepth 1 -type d -printf '%f\n' | sort | tail -n 1)
cat .omx/logs/autoresearch/$RUN_ID/manifest.json
cat .omx/logs/autoresearch/$RUN_ID/candidate.json
cat .omx/logs/autoresearch/$RUN_ID/iteration-ledger.json

You can also run evaluators directly without the supervisor:

node scripts/eval-in-action-cat-shellout-demo.js
python3 scripts/eval-ml-kaggle-model-optimization.py
python3 scripts/eval-noisy-bayesopt-highdim.py
python3 scripts/eval-noisy-latent-subspace-discovery.py
python3 scripts/eval-adaptive-sort-optimization.py

These showcases are meant to stay lightweight.

Please avoid committing:

• downloaded datasets
• large model artifacts
• benchmark output dumps
• generated caches like __pycache__/
• runtime autoresearch logs under .omx/logs/

Keep research state in code, configs, missions, and evaluator scripts; keep bulky runtime outputs local.