claude-turing 3.0.0 → 3.2.0

package/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "turing",
-  "version": "3.0.0",
-  "description": "Autonomous ML research harness — the autoresearch loop as a formal protocol. 41 commands, 2 specialized agents, meta-intelligence (cross-project knowledge transfer + methodology audit), scaling & efficiency (scaling laws + compute budget + model distillation), model composition (ensemble + pipeline stitch + warm-start), deep analysis (experiment diff + live training monitor + regression gate), experiment orchestration (batch queue + smart retry + branching), literature integration + paper drafting, production model export, performance profiling, smart checkpoints, experiment intelligence, statistical rigor, tree-search hypothesis exploration, cost-performance frontier, model cards, model registry, hypothesis database with novelty guard, anti-cheating guardrails, and the taste-leverage loop. Inspired by Karpathy's autoresearch and the scientific method itself.",
+  "version": "3.2.0",
+  "description": "Autonomous ML research harness — the autoresearch loop as a formal protocol. 47 commands, 2 specialized agents, model debugging (xray + sensitivity + calibration), pre-training intelligence (sanity checks + baseline generation + leakage detection), meta-intelligence (cross-project knowledge transfer + methodology audit), scaling & efficiency (scaling laws + compute budget + model distillation), model composition (ensemble + pipeline stitch + warm-start), deep analysis (experiment diff + live training monitor + regression gate), experiment orchestration (batch queue + smart retry + branching), literature integration + paper drafting, production model export, performance profiling, smart checkpoints, experiment intelligence, statistical rigor, tree-search hypothesis exploration, cost-performance frontier, model cards, model registry, hypothesis database with novelty guard, anti-cheating guardrails, and the taste-leverage loop. Inspired by Karpathy's autoresearch and the scientific method itself.",
   "author": {
     "name": "pragnition"
   },

package/README.md

@@ -352,6 +352,12 @@ The index (`hypotheses.yaml`) is the lightweight queue. The detail files (`hypot
 | `/turing:distill <exp-id>` | Model compression — distill teacher into smaller student with accuracy/size tradeoff |
 | `/turing:transfer [--from]` | Cross-project knowledge transfer — find similar projects, surface what worked |
 | `/turing:audit [--strict]` | Pre-submission methodology audit — data leakage, baselines, seeds, ablations, reproducibility |
+| `/turing:sanity [--quick]` | Pre-training sanity checks — initial loss, single-batch overfit, gradient flow, output validation |
+| `/turing:baseline [--methods]` | Automatic baseline generation — random, majority/mean, linear, k-NN |
+| `/turing:leak [--deep]` | Targeted leakage detection — single-feature tests, correlation, train/test overlap |
+| `/turing:xray [exp-id]` | Internal model diagnostics — gradient flow, dead neurons, weight distributions, tree analysis |
+| `/turing:sensitivity [exp-id]` | Hyperparameter sensitivity — rank parameters by impact, detect non-monotonic responses |
+| `/turing:calibrate [exp-id]` | Probability calibration — ECE/MCE, reliability diagrams, Platt/isotonic/temperature scaling |
 
 And for fully hands-off operation:
 
@@ -536,11 +542,11 @@ Each project gets independent config, data, experiments, models, and agent memor
 
 ## Architecture of Turing Itself
 
-41 commands, 2 agents, 10 config files, 60 template scripts, model registry, artifact contract, cost-performance frontier, model cards, tree-search exploration, statistical rigor, experiment intelligence, performance profiling, smart checkpoints, production model export, literature integration, paper section drafting, experiment orchestration (queue + retry + fork), deep analysis (diff + watch + regress), model composition (ensemble + stitch + warm), scaling & efficiency (scale + budget + distill), meta-intelligence (transfer + audit), 16 ADRs. See [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md) for the full codemap.
+47 commands, 2 agents, 10 config files, 66 template scripts, model registry, artifact contract, cost-performance frontier, model cards, tree-search exploration, statistical rigor, experiment intelligence, performance profiling, smart checkpoints, production model export, literature integration, paper section drafting, experiment orchestration (queue + retry + fork), deep analysis (diff + watch + regress), model composition (ensemble + stitch + warm), scaling & efficiency (scale + budget + distill), meta-intelligence (transfer + audit), pre-training intelligence (sanity + baseline + leak), model debugging (xray + sensitivity + calibrate), 16 ADRs. See [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md) for the full codemap.
 
 ```
 turing/
-├── commands/              40 skill files (core + taste-leverage + reporting + exploration + statistical rigor + experiment intelligence + performance + deployment + research workflow + orchestration + deep analysis + model composition + scaling & efficiency + meta-intelligence)
+├── commands/              46 skill files (core + taste-leverage + reporting + exploration + statistical rigor + experiment intelligence + performance + deployment + research workflow + orchestration + deep analysis + model composition + scaling & efficiency + meta-intelligence + pre-training intelligence + model debugging)
 ├── agents/                2 agents (researcher: read/write, evaluator: read-only)
 ├── config/                8 files (lifecycle, taxonomy, archetypes, novelty aliases)
 ├── templates/             Scaffolded into user projects by /turing:init

package/commands/baseline.md

@@ -0,0 +1,45 @@
+---
+name: baseline
+description: Automatic baseline generation — random, majority/mean, linear, k-NN baselines in 60 seconds. Every experiment needs a "is this better than dumb?" reference.
+disable-model-invocation: true
+argument-hint: "[--methods all|simple|linear] [--data data.npz]"
+allowed-tools: Read, Bash(*), Grep, Glob
+---
+
+Generate trivial baselines so you always know if your model is meaningfully better than simple approaches.
+
+## Steps
+
+1. **Activate environment:**
+   ```bash
+   source .venv/bin/activate
+   ```
+
+2. **Parse arguments from `$ARGUMENTS`:**
+   - `--methods all|simple|linear` — baseline group (default: all)
+   - `--data data.npz` — data file with X and y arrays
+   - `--json` — raw JSON output
+
+3. **Run baseline generation:**
+   ```bash
+   python scripts/generate_baselines.py $ARGUMENTS
+   ```
+
+4. **Baselines generated:**
+   - **Classification:** Random, Majority class, Stratified random, Logistic Regression, k-NN
+   - **Regression:** Random, Mean predictor, Median predictor, Ridge Regression, k-NN
+   - Each evaluated with the same protocol as real experiments
+
+5. **Report includes:** comparison table with metric values and notes (floor, ceiling, reference)
+
+6. **Integration:** satisfies the "baseline comparison" check in `/turing:audit`
+
+7. **Saved output:** report in `experiments/baselines/baselines-*.yaml`
+
+## Examples
+
+```
+/turing:baseline                           # All baselines
+/turing:baseline --methods simple          # Just random + majority
+/turing:baseline --data data/processed.npz # With actual data
+```

package/commands/calibrate.md

@@ -0,0 +1,47 @@
+---
+name: calibrate
+description: Probability calibration — measure ECE, plot reliability diagrams, apply Platt scaling or isotonic regression.
+disable-model-invocation: true
+argument-hint: "[exp-id] [--method platt|isotonic|temperature|auto]"
+allowed-tools: Read, Bash(*), Grep, Glob
+---
+
+Make model probabilities trustworthy. Does 80% confidence actually mean 80% correct?
+
+## Steps
+
+1. **Activate environment:**
+   ```bash
+   source .venv/bin/activate
+   ```
+
+2. **Parse arguments from `$ARGUMENTS`:**
+   - Optional experiment ID
+   - `--method platt|isotonic|temperature|auto` — calibration method (default: auto)
+   - `--json` — raw JSON output
+
+3. **Run calibration:**
+   ```bash
+   python scripts/calibration.py $ARGUMENTS
+   ```
+
+4. **Report includes:**
+   - ECE/MCE before calibration
+   - Reliability diagram (predicted vs actual per bin)
+   - Calibration method comparison table
+   - Verdict: ALREADY CALIBRATED / IMPROVED / NO IMPROVEMENT
+
+5. **Methods:**
+   - **Platt:** logistic regression on logits
+   - **Isotonic:** non-parametric (more flexible, needs more data)
+   - **Temperature:** single scalar T parameter
+   - **Auto:** tries all, picks lowest ECE
+
+6. **Saved output:** report in `experiments/calibration/<exp-id>-calibration.yaml`
+
+## Examples
+
+```
+/turing:calibrate exp-042                  # Auto-select best method
+/turing:calibrate exp-042 --method platt   # Platt scaling only
+```

package/commands/leak.md

@@ -0,0 +1,47 @@
+---
+name: leak
+description: Targeted leakage detection — probe for data leakage with single-feature tests, correlation checks, and train/test overlap detection.
+disable-model-invocation: true
+argument-hint: "[--deep] [--features feature_1,feature_2]"
+allowed-tools: Read, Bash(*), Grep, Glob
+---
+
+Actively probe for data leakage. The #1 cause of "too good to be true" results.
+
+## Steps
+
+1. **Activate environment:**
+   ```bash
+   source .venv/bin/activate
+   ```
+
+2. **Parse arguments from `$ARGUMENTS`:**
+   - `--deep` — run full single-feature analysis (slow but thorough)
+   - `--features "feat_1,feat_2"` — check specific features
+   - `--json` — raw JSON output
+
+3. **Run leakage scan:**
+   ```bash
+   python scripts/leakage_detector.py $ARGUMENTS
+   ```
+
+4. **Checks performed:**
+   - **Feature-target correlation:** flag features with >0.95 correlation to target
+   - **Single-feature predictiveness (--deep):** train on each feature alone, flag any that achieve >80% of full model performance
+   - **Train/test overlap:** hash-based deduplication across splits
+
+5. **Verdicts:**
+   - **CLEAN** — no leakage detected
+   - **SUSPICIOUS** — warnings to review
+   - **LEAKAGE DETECTED** — critical flags found
+
+6. **Integration:** satisfies the "data leakage" check in `/turing:audit`
+
+7. **Saved output:** report in `experiments/leakage/leak-*.yaml`
+
+## Examples
+
+```
+/turing:leak                    # Standard correlation + overlap checks
+/turing:leak --deep             # Full single-feature analysis
+```

package/commands/sanity.md

@@ -0,0 +1,48 @@
+---
+name: sanity
+description: Pre-training sanity checks — catch broken data loaders, misconfigured losses, and dead gradients in 30 seconds before wasting hours.
+disable-model-invocation: true
+argument-hint: "[--quick] [--verbose]"
+allowed-tools: Read, Bash(*), Grep, Glob
+---
+
+Run a battery of fast checks before committing to a full training run. Catches wiring bugs in seconds.
+
+## Steps
+
+1. **Activate environment:**
+   ```bash
+   source .venv/bin/activate
+   ```
+
+2. **Parse arguments from `$ARGUMENTS`:**
+   - `--quick` — skip single-batch overfit test (fastest, ~5 seconds)
+   - `--verbose` — show detailed check output
+   - `--json` — raw JSON output
+
+3. **Run sanity checks:**
+   ```bash
+   python scripts/sanity_checks.py $ARGUMENTS
+   ```
+
+4. **Checks performed:**
+   - **Data pipeline** (critical): first batch loads, shapes match, no NaN/Inf
+   - **Initial loss** (high): loss at initialization matches theory (e.g., -log(1/C) for cross-entropy)
+   - **Gradient flow** (high): all parameters have non-zero, non-exploding gradients
+   - **Single-batch overfit** (critical): model can memorize 1 batch in 50 steps — if not, something is broken
+   - **Output validation** (high): predictions are non-NaN, non-constant, reasonable range
+   - **Config consistency** (medium): learning rate, batch size in reasonable ranges
+
+5. **Verdicts:**
+   - **PASS** — safe to proceed
+   - **PASS (with warnings)** — review before training
+   - **FAIL** — do not proceed, fix issues first
+
+6. **Saved output:** report in `experiments/sanity/sanity-*.yaml`
+
+## Examples
+
+```
+/turing:sanity                    # Full check (~30 seconds)
+/turing:sanity --quick            # Skip overfit test (~5 seconds)
+```

package/commands/sensitivity.md

@@ -0,0 +1,41 @@
+---
+name: sensitivity
+description: Hyperparameter sensitivity analysis — rank parameters by impact, identify which matter and which are noise.
+disable-model-invocation: true
+argument-hint: "[exp-id] [--params learning_rate,max_depth]"
+allowed-tools: Read, Bash(*), Grep, Glob
+---
+
+Which hyperparameters actually matter? Stop wasting time on the ones that don't.
+
+## Steps
+
+1. **Activate environment:**
+   ```bash
+   source .venv/bin/activate
+   ```
+
+2. **Parse arguments from `$ARGUMENTS`:**
+   - Optional experiment ID
+   - `--params "learning_rate,max_depth"` — specific parameters to analyze
+   - `--json` — raw JSON output
+
+3. **Run sensitivity analysis:**
+   ```bash
+   python scripts/sensitivity_analysis.py $ARGUMENTS
+   ```
+
+4. **Report includes:**
+   - Per-parameter sensitivity ranking: HIGH / MED / LOW / NONE
+   - Metric range for each parameter sweep
+   - Monotonicity detection (is there a sweet spot?)
+   - Recommendations: focus tuning on X, stop tuning Y
+
+5. **Saved output:** report in `experiments/sensitivity/<exp-id>-sensitivity.yaml`
+
+## Examples
+
+```
+/turing:sensitivity exp-042                           # All tunable params
+/turing:sensitivity --params "learning_rate,max_depth" # Specific params
+```

package/commands/turing.md

@@ -50,6 +50,12 @@ You are the Turing ML research router. Detect the user's intent and route to the
 | "distill", "compress", "smaller model", "student model", "knowledge distillation", "model compression" | `/turing:distill` | Deploy |
 | "transfer", "what worked before", "similar project", "cross-project", "institutional knowledge", "prior projects" | `/turing:transfer` | Research |
 | "audit", "methodology check", "pre-submission", "reviewer checklist", "data leakage", "missing baselines" | `/turing:audit` | Validate |
+| "sanity", "sanity check", "pre-training", "is it broken", "before training", "quick check" | `/turing:sanity` | Check |
+| "baseline", "baselines", "trivial baseline", "majority class", "is it better than random" | `/turing:baseline` | Analyze |
+| "leak", "leakage", "data leakage scan", "suspicious feature", "train test overlap" | `/turing:leak` | Validate |
+| "xray", "model internals", "dead neurons", "gradient flow", "weight distribution", "inside the model" | `/turing:xray` | Analyze |
+| "sensitivity", "which params matter", "hyperparameter importance", "parameter ranking" | `/turing:sensitivity` | Analyze |
+| "calibrate", "calibration", "ECE", "reliability diagram", "overconfident", "probability calibration" | `/turing:calibrate` | Analyze |
 
 ## Sub-commands
 
@@ -96,6 +102,12 @@ You are the Turing ML research router. Detect the user's intent and route to the
 | `/turing:distill <exp-id>` | Model compression: distill teacher into smaller student model | (inline) |
 | `/turing:transfer [--from]` | Cross-project knowledge transfer: find similar prior projects, surface what worked | (inline) |
 | `/turing:audit [--strict]` | Pre-submission methodology audit: data leakage, baselines, seeds, ablations, reproducibility | (inline) |
+| `/turing:sanity [--quick]` | Pre-training sanity checks: initial loss, overfit test, gradient flow, output validation | (inline) |
+| `/turing:baseline [--methods]` | Automatic baseline generation: random, majority/mean, linear, k-NN | (inline) |
+| `/turing:leak [--deep]` | Targeted leakage detection: single-feature tests, correlation, train/test overlap | (inline) |
+| `/turing:xray [exp-id]` | Internal model diagnostics: gradient flow, dead neurons, weight distributions, tree analysis | (inline) |
+| `/turing:sensitivity [exp-id]` | Hyperparameter sensitivity analysis: rank parameters by impact, detect non-monotonic responses | (inline) |
+| `/turing:calibrate [exp-id]` | Probability calibration: ECE/MCE, reliability diagrams, Platt/isotonic/temperature scaling | (inline) |
 
 ## Proactive Detection
 

package/commands/xray.md

@@ -0,0 +1,43 @@
+---
+name: xray
+description: Internal model diagnostics — gradient flow, dead neurons, activation stats, weight distributions, tree depth analysis.
+disable-model-invocation: true
+argument-hint: "[exp-id] [--layer encoder.layer.2] [--compare exp-a exp-b]"
+allowed-tools: Read, Bash(*), Grep, Glob
+---
+
+See inside the model. When it underperforms, the fix depends on *why*.
+
+## Steps
+
+1. **Activate environment:**
+   ```bash
+   source .venv/bin/activate
+   ```
+
+2. **Parse arguments from `$ARGUMENTS`:**
+   - Optional experiment ID
+   - `--layer "name"` — focus on specific layer
+   - `--compare exp-a exp-b` — side-by-side diagnostics
+   - `--json` — raw JSON output
+
+3. **Run model diagnostics:**
+   ```bash
+   python scripts/model_xray.py $ARGUMENTS
+   ```
+
+4. **Diagnostics by model type:**
+   - **Neural networks:** gradient magnitudes, activation stats, dead neuron %, weight distributions, gradient-to-weight ratio
+   - **Tree models:** depth utilization, leaf purity, feature split dominance
+   - **scikit-learn:** coefficient magnitudes, feature importance concentration
+
+5. **Issues detected:** dead gradients, vanishing/exploding gradients, dead neurons, sparse weights, feature dominance, overfitting risk
+
+6. **Saved output:** report in `experiments/xrays/<exp-id>-xray.yaml`
+
+## Examples
+
+```
+/turing:xray exp-042              # Full diagnostics
+/turing:xray                      # Best experiment
+```

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-turing",
-  "version": "3.0.0",
+  "version": "3.2.0",
   "type": "module",
   "description": "Autonomous ML research harness for Claude Code. The autoresearch loop as a formal protocol — iteratively trains, evaluates, and improves ML models with structured experiment tracking, convergence detection, immutable evaluation infrastructure, and safety guardrails.",
   "bin": {

package/src/install.js

@@ -29,6 +29,8 @@ const SUB_COMMANDS = [
   "ensemble", "stitch", "warm",
   "scale", "budget", "distill",
   "transfer", "audit",
+  "sanity", "baseline", "leak",
+  "xray", "sensitivity", "calibrate",
 ];
 
 export async function install(opts = {}) {

package/src/verify.js

@@ -55,6 +55,12 @@ const EXPECTED_COMMANDS = [
   "distill/SKILL.md",
   "transfer/SKILL.md",
   "audit/SKILL.md",
+  "sanity/SKILL.md",
+  "baseline/SKILL.md",
+  "leak/SKILL.md",
+  "xray/SKILL.md",
+  "sensitivity/SKILL.md",
+  "calibrate/SKILL.md",
 ];
 
 const EXPECTED_AGENTS = ["ml-researcher.md", "ml-evaluator.md"];