Skip to content

Taste-Leverage Interface

These commands bridge human research taste and machine execution leverage. You decide what is worth trying; the agent decides how to try it rigorously. This is the core interaction pattern: the human selects which coins to flip, the agent flips them.

/turing:try: Inject a Hypothesis

Inject a hypothesis into the agent's experiment queue. This is the primary taste-leverage mechanism: you provide judgment about what is worth trying, the agent provides disciplined execution.

Human-injected hypotheses are marked as high priority and tested before the agent generates its own. After testing, each hypothesis is marked as tested, promising, or dead-end with a link to the resulting experiment.

Syntax: /turing:try <hypothesis description> or /turing:try archetype:<name>

  • Free-text hypotheses are queued directly.
  • Archetype syntax (e.g., archetype:model_comparison) expands into structured multi-experiment strategies.

Available archetypes: model_comparison (~5 experiments), hyperparameter_sweep (15-36), feature_sweep (6-10), regularization_search (4-6), ensemble_construction (4-6), learning_rate_schedule (4-5), data_quality_audit (3-5), ablation_study (N+1).

Examples:

/turing:try switch to LightGBM with dart boosting and lower learning rate
# Free-text hypothesis, queued as high priority

/turing:try add polynomial features for the numeric columns
# Another free-text hypothesis

/turing:try archetype:model_comparison
# Structured strategy: compare XGBoost, LightGBM, RF, LR, MLP with statistical tests

/turing:try archetype:ensemble_construction
# Build voting, stacking, and blending ensembles of top models

Tip

Use archetypes for systematic exploration (e.g., model_comparison early on) and free-text for targeted refinements (e.g., "increase regularization, the train/val gap suggests overfitting").

/turing:brief: Research Intelligence Report

Generate a structured research intelligence report from experiment history: what has been learned, what is promising, what is exhausted, and what the human should consider next. Designed to be read in 2 minutes.

The briefing has 6 sections: Campaign Summary, Current Best, Improvement Trajectory, Model Types Explored, Hypothesis Queue, and Recommendations. The --deep flag adds a 7th section with Literature-Grounded Suggestions that searches recent papers and auto-queues hypotheses.

Syntax: /turing:brief [ml/project] [--deep]

  • Accepts an optional project path. Auto-detects from config.yaml otherwise.
  • --deep adds literature-grounded suggestions by searching recent papers and distilling 3-5 actionable techniques.

Examples:

/turing:brief
# 2-minute intelligence briefing on the current project

/turing:brief ml/sentiment
# Brief for a specific project

/turing:brief --deep
# Full briefing plus literature-grounded suggestions with citations, auto-queued as hypotheses

Tip

Run /turing:brief every time you return to a project after time away. It tells you exactly where things stand and what to try next. Use --deep when the agent seems stuck and you want evidence-based direction.

/turing:suggest: Literature-Grounded Model Selection

Read the ML task context, search recent literature, and suggest model architectures worth trying, with citations. Suggestions are auto-queued as hypotheses for the next /turing:train iteration.

Supports two strategies: literature (default) searches the web for recent papers and synthesizes grounded suggestions, and treequest uses AB-MCTS tree search over the critique scoring function to explore refinement chains that linear search cannot find.

Syntax: /turing:suggest [task description override] [--strategy literature|treequest]

  • Default strategy is literature (requires internet).
  • --strategy treequest uses tree-search-guided exploration (works offline).
  • TreeQuest options: --iterations N, --top N, --greedy.

Examples:

/turing:suggest
# Literature search for model suggestions based on config.yaml task

/turing:suggest "tabular classification with heavy class imbalance"
# Override the task description for more targeted literature search

/turing:suggest --strategy treequest --iterations 50 --top 10
# Tree-search exploration with 50 iterations, top 10 results

Tip

Use the literature strategy early in a project to find strong baselines. Switch to treequest mid-experiment when you want to explore non-obvious combinations of techniques that papers do not explicitly describe.

/turing:explore: Tree-Search Hypothesis Exploration

Explore the hypothesis space using AB-MCTS tree search. Instead of suggesting independent ideas, this builds and searches a tree of refinement chains; each node is a hypothesis scored by novelty, feasibility, and expected impact. Discovers non-obvious combinations that linear suggestion cannot find.

Syntax: /turing:explore [ml/project] [--iterations N] [--top N] [--strategy abmcts-a|abmcts-m|greedy]

  • --iterations N sets search depth (default: 30).
  • --top N sets number of results (default: 5).
  • --strategy selects the algorithm: abmcts-a (default), abmcts-m (Bayesian, requires PyMC), or greedy.
  • --seeds-only shows generated seeds without running search.
  • --json outputs JSON for programmatic use.

Examples:

/turing:explore
# Default: 30 iterations, top 5, AB-MCTS-A strategy

/turing:explore --iterations 50 --top 10
# Deeper search, more results

/turing:explore --strategy greedy
# Fast greedy search without TreeQuest dependency

/turing:explore ml/churn --json
# JSON output for a specific project

Tip

/turing:explore differs from /turing:suggest in a key way: suggest finds what papers recommend (independent suggestions), while explore finds what combinations score well (refinement chains). Use suggest for early exploration, explore for mid-experiment optimization.

/turing:design: Experiment Design Document

Generate a structured experiment design for a hypothesis before running it. Reads experiment history, searches literature for methodology, and produces a scored design document. Front-loads the thinking before the coding.

The design includes a falsifiable objective, literature-grounded method, concrete implementation plan for train.py, expected outcomes with metric thresholds, risks from the literature, and a self-critique score.

Syntax: /turing:design <hypothesis-id or description>

  • Accepts a hypothesis ID (e.g., hyp-003) or free-text description.
  • Design documents are saved to experiments/designs/.

Examples:

/turing:design hyp-003
# Design document for a queued hypothesis

/turing:design "switch to CatBoost with ordered boosting for categorical features"
# Design from a free-text description

/turing:design "add target encoding for high-cardinality categorical features"
# Literature-grounded design with implementation plan

Tip

Use /turing:design for high-risk or complex hypotheses. The literature search often surfaces pitfalls and hyperparameter recommendations that save wasted iterations.

/turing:mode: Set Research Strategy

Set the research strategy mode: explore (try new things), exploit (refine what works), or replicate (verify results). The mode drives the novelty guard policy and agent behavior during /turing:train.

Mode Novelty Guard Agent Behavior
explore Allow novel ideas, block repeats Try fundamentally different approaches
exploit Allow follow-ups, block repeats Refine the current best configuration
replicate Allow duplicates, block novel ideas Re-run best experiments with different seeds

Syntax: /turing:mode <explore|exploit|replicate>

Examples:

/turing:mode explore
# Switch to exploration — best when the current approach feels exhausted

/turing:mode exploit
# Switch to exploitation — best when you have a promising direction to refine

/turing:mode replicate
# Switch to replication — best before declaring a winner

Tip

The default mode is exploit. Switch to explore when you are plateauing, and to replicate before making final decisions about which model to ship.