Anistroph — Guided Walkthrough

A step-by-step walkthrough of the full inference lifecycle across three reference datasets covering regression and classification, temporal and non-temporal prediction.

Each step is a prompt you can paste into Claude Desktop (or any MCP client connected to Anistroph). Run the prompts in order within a single conversation — Claude carries context forward, so later prompts can build on earlier results.

New to Anistroph? See the Anistroph project overview for the architecture, capabilities, and design goals.

Prerequisites: Anistroph installed and datasets registered (make install), and the Anistroph MCP server configured in Claude Desktop. See the Setup & Usage Guide for setup details.

AI Agent Analysis & Validation

Discover → Understand → Execute → Validate

Claude and AI agents first discover the available datasets and models, then inspect the selected model’s input contract to determine required inputs, temporal requirements such as as_of, and required inference history.

The agent can then orchestrate prediction, explanation, evaluation, and analysis through MCP. These operations are executed by Anistroph’s shared services and can be independently reproduced through the Web UI or REST API when validation is required.

The Inference Lifecycle

Every dataset in Anistroph follows the same core lifecycle, regardless of domain or task type:

Step Purpose MCP Tool
Discover See what datasets and models exist anistroph_list_datasets, anistroph_list_models
Understand Profile the dataset and slice it by dimensions anistroph_profile_dataset, anistroph_slice_data, anistroph_compare_data
Inspect the input contract Learn what the model needs to predict anistroph_get_model_inputs
Predict Run inference — entity lookup or records anistroph_predict
Explain SHAP feature drivers for a prediction anistroph_explain_prediction
Evaluate Held-out metrics on the evaluation partition anistroph_evaluate_model
Find error slices Where does the model break? anistroph_find_evaluation_slices

The procurement section below follows a hook-first order — leading with the prediction, then revealing the input contract, then exploring broader capabilities. Datasets 2 and 3 follow the traditional discovery-first order. Both orders exercise the same lifecycle; the difference is narrative pacing.

The walkthrough runs the full lifecycle on three datasets that together exercise every prediction mode Anistroph supports:

Dataset Domain Target Task Prediction Mode
Semiconductor Procurement — Demand Supply chain material_demand_next_4w Regression Temporal entity lookup (rolling features)
Semiconductor Yield Manufacturing wafer_yield Regression Non-temporal entity lookup + records mode
Predictive Maintenance — Failure Equipment health failure_within_horizon Classification Temporal entity lookup

Dataset 1 — Semiconductor Procurement: Demand Forecasting

Why this dataset first: It demonstrates Anistroph’s temporal prediction capability — the model uses rolling-window features (4-week, 8-week, 13-week means) that are reconstructed from entity history at prediction time. The model is static; the features are not. This is the case most ML platforms cannot handle without a custom feature pipeline.

The procurement section follows a hook-first order: lead with the prediction, then reveal how it works, then explore the broader capabilities. This keeps the narrative moving and shows the most impressive capability immediately.

Before you start: These prompts assume the Anistroph MCP server is connected to Claude. If Claude doesn’t recognize Anistroph tool names, the MCP server isn’t configured — see Setup & Usage → Use Anistroph with Claude. To verify, ask: “What datasets are available in Anistroph?” — Claude should list the registered datasets without asking for a file upload.

Step 1 — The hero prediction

Use Anistroph to predict the 4-week material demand for series FAB_A__MAT_0001 as of 2026-07-06. Then show me the actual demand for that period from the dataset so we can compare.

Claude calls anistroph_predict with entity_id=FAB_A__MAT_0001 and timestamp=2026-07-06. Anistroph loads 13 weeks of history for this series up to July 6th, computes the 4w/8w/13w rolling means, applies the trained XGBoost model, and returns the forecast. Claude then calls anistroph_sample_rows to fetch the actual material_demand_next_4w value (221.1) for that week so you can compare prediction vs. actual.

One MCP call produced a temporal forecast with rolling feature reconstruction — no feature pipeline, no manual window calculations.

Step 2 — Reveal the input contract

What inputs does this model need to make a prediction? Does it require a timestamp, and if so, how much history does it need?

Claude calls anistroph_get_model_inputs. The response tells you:

This is the key concept: Claude didn’t need to know about the rolling features in Step 1. The model declares its own input contract, and Anistroph handles the history retrieval and feature reconstruction internally.

Step 3 — Explain the prediction

Explain the prediction you just made for FAB_A__MAT_0001. Which features are driving the demand forecast up or down?

Claude calls anistroph_explain_prediction. The SHAP response shows the top positive and negative drivers — typically recent consumption trends (rolling means), planned wafer starts, and inventory level. Categorical features (fab, material category, supplier) appear as grouped contributions rather than separate one-hot columns.

Step 4 — Understand the domain

Profile the semiconductor_procurement_demand dataset. What’s the row count, how many fab-material series are there, and how does 4-week material demand vary by material category?

Claude profiles the dataset (anistroph_profile_dataset), then slices it (anistroph_compare_data with dimension=material_category, metric=material_demand_next_4w). You’ll see which material categories (Photoresists, Process Gases, Wet Chemicals, etc.) drive the most demand.

Step 5 — Stress-test with a real crisis period

Find a fab-material series that experienced a demand spike or inventory crisis — where inventory was near zero and consumption was high — and predict demand as of that crisis point. Then explain what drove the spike.

This model is entity_lookup only (rolling transforms), so Claude can’t send synthetic records. Instead, Claude searches the dataset for a real stress period — using anistroph_sample_rows or anistroph_slice_data to find a series/week where inventory_on_hand was near zero and material_consumption_qty was high. It then calls anistroph_predict with that entity_id + timestamp, and anistroph_explain_prediction on the same point.

This demonstrates Claude acting as an analyst — it finds the stress scenario in the data, predicts at that point, and explains what drove the spike. The LLM orchestrates the investigation; Anistroph handles the feature reconstruction and model execution.

Step 6 — Multi-target: shortage risk classification

The procurement data also has a shortage risk model in Anistroph. Predict the probability of a shortage in the next 4 weeks for series FAB_C__MAT_0094 as of 2023-01-30, and explain which factors drive the risk.

Same source parquet, different target, different task type — now we’re predicting a probability, not a quantity. Claude calls anistroph_predict with the shortage model, entity_id=FAB_C__MAT_0094, and timestamp=2023-01-30. The SHAP drivers from anistroph_explain_prediction typically show inventory well below safety stock and long supplier lead times pushing the risk score up. Two models, one dataset, zero new application code.

Step 7 — Evaluate and find error slices

Evaluate the semiconductor_procurement_demand model on its held-out evaluation set. What are the R², MAE, and RMSE? Then find the slices where the demand forecast error is worst — which fab, material category, or supplier combinations does the model struggle on?

Claude calls anistroph_evaluate_model (R²=0.96, MAE=14.0, RMSE=17.8) and anistroph_find_evaluation_slices. The error-slice search ranks 1/2/3-dimensional combinations (e.g. fab_id × material_category × supplier_id) by how much the absolute prediction error deviates from the overall baseline. This tells you not just “the model is good overall” but “here’s exactly where it’s weak.”

Step 8 — Procurement risk analysis

In Anistroph’s procurement data, which suppliers are associated with the greatest shortage risk?

Claude calls anistroph_compare_data (dimension=supplier_id, metric=shortage_risk_next_4w, aggregation=mean). The response ranks suppliers by average predicted shortage risk — a direct analytical question about the data that complements the model predictions.

Step 9 — Discover the full architecture

What datasets and models are available in Anistroph?

Claude calls anistroph_list_datasets and anistroph_list_models in parallel. You’ll see 13 datasets and 13 models across four domains — semiconductor manufacturing, predictive maintenance, real estate, and supply chain procurement. All discovered through one MCP connection, all using the same shared runtime.

Dataset 2 — Semiconductor Yield: Wafer Yield Prediction

Why this dataset next: It demonstrates the non-temporal case — single-row entity lookup with no rolling features. It also shows records mode (sending raw feature values instead of an entity ID) and the staged-prediction pattern (same target, progressively more features).

Step 1 — Discover

In Anistroph, what models are available for semiconductor wafer yield prediction?

Claude calls anistroph_list_models and filters to models whose target_name is wafer_yield. You’ll see the main model plus four staged models (A through D).

Step 2 — Understand

Profile the semiconductor_yield dataset. What’s the row count, and how does wafer_yield vary across etch_tool and etch_chamber?

Claude profiles the dataset, then slices it by etch_tool × etch_chamber over wafer_yield (anistroph_slice_data). You’ll see which tool/chamber combinations produce the lowest yields.

Step 3 — Inspect the input contract

What inputs does the wafer-yield-xgboost model need? Can I send raw records, or do I need to look up a wafer?

Claude calls anistroph_get_model_inputs. The response shows:

Because this is a non-temporal model with only current and categorical transforms, you can either look up a wafer by ID or send a raw record with all the feature values.

Step 4 — Predict (entity lookup)

Show me the row for wafer_id WAFER_015000 in the semiconductor_yield dataset, then predict its wafer_yield using the wafer-yield-xgboost model.

Claude calls anistroph_sample_rows (filter wafer_id=WAFER_015000) to show the actual row, then anistroph_predict with entity_id=WAFER_015000. The response includes both the predicted and actual yield.

Step 4b — Predict (records mode)

Use Anistroph to generate a plausible “worst-case” wafer record — pick etch_tool, chamber, recipe, and process settings that you’d expect to produce low yield — and run it through the wafer-yield-xgboost model. Then generate a “best-case” record and predict that too. Compare the two.

Claude uses the input schema from Step 3, composes two records using domain reasoning (e.g. ETCH_02 + CH_B + high maintenance age + low exposure dose for the worst case), and calls anistroph_predict with records=[worst, best]. This demonstrates records mode — the LLM is the test harness, fabricating realistic inputs and predicting on them without any feature pipeline code.

Step 5 — Explain

Explain the prediction you just made for WAFER_015000. Which features pushed the yield up, and which pushed it down?

Claude calls anistroph_explain_prediction. The SHAP response shows grouped one-hot drivers (e.g. etch_tool = ETCH_02 as a single contribution) plus the continuous feature drivers (etch temperature, exposure dose, maintenance age).

Step 6 — Evaluate

Evaluate the wafer-yield-xgboost model on its held-out evaluation set.

Claude calls anistroph_evaluate_model. The response shows R²=0.81, MAE=0.0065, plus a prediction-vs-actual sample from the 10,000 held-out wafers.

Step 7 — Find error slices

Find the slices where the wafer yield prediction error is worst — which etch_tool, chamber, or product combinations does the model struggle on?

Claude calls anistroph_find_evaluation_slices. The response ranks combinations like etch_tool × etch_chamber × product_id by error deviation from baseline.

Optional — Staged prediction

I have four staged models for wafer yield — stage A (before etch), B (after etch), C (after deposition), D (before test). For wafer WAFER_015000, predict yield with each stage model and show how the prediction sharpens as more process data becomes available.

Claude calls anistroph_predict four times on the staged model IDs. The predictions converge toward the actual as more features are added, demonstrating the staged-prediction pattern: the same wafer is predicted at four points in the production line, each with only the features knowable at that point.

Dataset 3 — Predictive Maintenance: Failure Within Horizon

Why this dataset last: It demonstrates the classification case — the model returns a probability, not a quantity. It’s also temporal (sensor history with rolling features), so prediction requires an as_of timestamp. The evaluation metrics and error-slice discovery use classification-specific measures (ROC-AUC, F1, log loss) instead of regression measures (MAE, R²).

Step 1 — Discover

In Anistroph, what models are available for predictive maintenance?

Claude calls anistroph_list_models and filters to maintenance-related models. You’ll see the failure model (classification), the RUL model (regression), and the maintenance-required model (classification).

Step 2 — Understand

Profile the predictive_maintenance dataset. What’s the row count, how many machines are there, and what’s the overall failure rate? How does failure rate vary by machine_type?

Claude profiles the dataset, then compares failure rate across machine_type (anistroph_compare_data). The failure rate is very low overall (~0.08%), so slicing by machine type helps see which types are riskier.

Step 3 — Inspect the input contract

What inputs does the predictive_maintenance failure model need? Does it require a timestamp?

Claude calls anistroph_get_model_inputs. The response shows:

Step 4 — Predict

Predict the probability of failure within the horizon for machine TOOL_010 as of 2026-06-02T05:30:00. Then show me what actually happened to TOOL_010 in the hour after that timestamp.

Claude calls anistroph_predict with entity_id=TOOL_010 and timestamp=2026-06-02T05:30:00. Anistroph loads the machine’s sensor history up to that point, computes rolling features, and returns a failure probability. Claude then calls anistroph_sample_rows to show the readings after that timestamp — revealing that TOOL_010 failed at 05:35 (five minutes later) with a POWER failure mode. This is a compelling before-the-fact prediction: the model was asked to assess risk 5 minutes before the actual failure.

Step 5 — Explain

Explain the failure prediction you just made for TOOL_010. Which sensor readings or conditions are driving the risk score up?

Claude calls anistroph_explain_prediction. The SHAP response shows which sensor features (temperature, vibration, pressure, maintenance age) and their rolling trends are pushing the failure probability up or down.

Step 6 — Evaluate

Evaluate the predictive_maintenance failure model on its held-out evaluation set. What are the ROC-AUC, F1, precision, and recall?

Claude calls anistroph_evaluate_model. The response includes classification metrics: ROC-AUC=0.85, F1=0.61, precision=0.48, recall=0.81, plus the decision threshold (0.223) chosen to maximize F1. You’ll also see a sample of prediction-vs-actual rows.

Step 7 — Find error slices

Find the slices where the failure model’s prediction error is worst. Which machine_type or failure_mode combinations does the model struggle on? Use log loss as the error metric.

Claude calls anistroph_find_evaluation_slices with metric=log_loss. The response ranks combinations like machine_type × failure_mode by how much the per-row log loss deviates from the overall baseline. For classification, this identifies populations where the model is overconfident or underconfident — not just where it’s wrong, but where its probability calibration is off.

What This Walkthrough Demonstrates

Capability Where it appears
Temporal prediction — rolling features reconstructed from entity history at prediction time Dataset 1, Step 1; Dataset 3
Self-describing input contract — model declares what it needs, including as_of and history window Dataset 1, Step 2; Datasets 2 and 3, Step 3
SHAP explainability with one-hot feature grouping Dataset 1, Step 3; Datasets 2 and 3, Step 5
Multidimensional analysis — slice and compare data by categorical dimensions Dataset 1, Step 4; Datasets 2 and 3, Step 2
Stress-test with real data — Claude searches for a crisis period and predicts at that point Dataset 1, Step 5
Multi-target on one source — same parquet, different target, different task type Dataset 1, Step 6
Regression evaluation — R², MAE, RMSE, MAPE Dataset 1, Step 7; Dataset 2
Classification evaluation — ROC-AUC, F1, precision, recall Dataset 1, Step 6; Dataset 3
Error slice discovery — where does the model break, by dimension combination Dataset 1, Step 7; Datasets 2 and 3
Procurement risk analysis — which suppliers carry the greatest shortage risk Dataset 1, Step 8
Discovery — list datasets and models via MCP Dataset 1, Step 9; Datasets 2 and 3, Step 1
Non-temporal prediction — single-row entity lookup Dataset 2
Records mode — send raw feature values instead of an entity ID Dataset 2, Step 4b
Staged prediction — same target, progressively more features Dataset 2, optional step
Multi-domain — same lifecycle across supply chain, manufacturing, and equipment health All three datasets

The same MCP tools, the same lifecycle, and the same Claude conversation — across three domains, two task types, and both temporal and non-temporal prediction. No domain-specific code was written for any of these steps. The datasets declare their schemas and feature transforms in YAML; the runtime handles the rest.

Next Steps