Purpose | Inputs and configurations | Preview before spending | Define a structured forecast | Smoke, phase the spend, then resume | Simulations or repeated expert forecasts? | Aggregate without accidental row-count weights | Diagnose and compare forecasts without project-specific wrangling | What remains downstream

Purpose | 1. Set package options | 2. Define the treatment text and prompt | 3. Preview the prompt before running | 4. Run the post-only simulation | 5. Inspect the raw output | 6. Summarize by identity | 7. Compare another group to Americans | When to use which function

Purpose | 1. Set package options | 2. Create a small text dataset | 3. Build the prompt | 4. Run the analysis | 5. Inspect the output | 6. Evaluate GPT against human labels | 7. Likert-style sentiment or emotion | 8. Repeated runs for reliability | 9. When to use this workflow | Reference

Step 1: Verify your connection with ellmer | Step 2: Run a minimal nalanda workflow

Purpose | Current package status | What the papers add | Recommended implementation steps | Why these steps matter | Step 1: Prompt layer | Step 2: Condition-based simulation wrapper | Step 3: Descriptive summaries and calibration | Step 4: Demographic profile infrastructure | Step 5: Ensemble controls | Step 6: Cumulative chapter designs | What can be applied directly to the existing chapter workflow? | Transferable immediately | Transferable with design adaptation | Less transferable without stronger validation | Where should calibration happen? | Why not pre-adjust inside simulation functions? | Why not leave calibration entirely to user scripts? | Recommended compromise | Proposed object designs | prompt_bank | ensemble_size | demographic_profiles | Proposed next steps | References

Overview | What the papers did, in simple language | What the supplement adds | 1. Prompting strategy matters | 2. Ensemble prompting matters | 3. Demographic conditioning is part of the method | 4. Absolute effect sizes need caution | Recommended take-aways for users | How this relates to nalanda | A concrete way to use nalanda today | Scenario 1: Pre/post chapter simulations | Scenario 2: Post-only simulations | Scenario 3: Control-versus-treatment chapter comparisons | What users should currently do themselves | 1. Build or manage a prompt bank manually | 2. Run ensembles manually | 3. Manage demographic profiles manually | 4. Estimate between-condition contrasts downstream | 5. Apply effect calibration explicitly | How to think about the 0.56 factor | When it is reasonable to use it | When to be cautious | Best current practice | A simple recommended workflow for users | Minimal workflow | Better workflow | What nalanda may support later | Final practical advice | References