Scholar Experiment: Results Analysis Workflow

A systematic workflow for analyzing ML/AI experimental results and generating publication-ready Results sections. Transforms raw experimental data into validated statistical analyses, publication-quality visualizations, and well-structured paper content.

Pre-load (before execution)

1. Codebase docs: If .workflow/codebase/ARCHITECTURE.md exists, read for project context
2. Specs: maestro load --type spec --category coding — load coding conventions
3. Wiki knowledge: maestro search "academic writing research paper" --json — top 5 entries as prior context
4. All optional — proceed without if unavailable

Architecture Overview

                     scholar-experiment
                           |
     ┌─────────────────────┼─────────────────────┐
     |                     |                       |
  [User Input]    [Experiment Context]    [Preferences]
     |                     |                       |
     └─────────┬───────────┘                       |
               v                                   |
┌──────────────────────────┐                       |
│  Phase 1: Data Loading   │ <─ preferences ───────┘
│  Load, validate, inspect │
└────────────┬─────────────┘
             │ cleanedData, dataProfile
             v
┌──────────────────────────┐
│  Phase 2: Statistical    │
│  Analysis & Testing      │
└────────────┬─────────────┘
             │ statisticalResults
             v
┌──────────────────────────┐
│  Phase 3: Visualization  │
│  Plots, charts, tables   │
└────────────┬─────────────┘
             │ figureSpecs, tableSpecs
             v
┌──────────────────────────┐
│  Phase 4: Results Writing│
│  Draft Results section   │
└────────────┬─────────────┘
             │ resultsDraft
             v
┌──────────────────────────┐
│  Phase 5: Quality Check  │
│  Validate & verify       │
└──────────────────────────┘
             │
             v
        [Output Files]
        - analysis-report.md
        - results-draft.md
        - visualization-specs.md

Key Design Principles

1. Statistical rigor first: Every claim must be backed by appropriate statistical tests with complete reporting (mean, SD/SE, p-value, effect size)
2. Pre-test before test: Always check assumptions (normality, variance homogeneity) before selecting parametric vs non-parametric tests
3. Publication-quality output: All visualizations must meet journal standards (vector format, colorblind-friendly, proper error representation)
4. Complete reporting: Never report p-values alone — always include effect sizes, confidence intervals, and descriptive statistics
5. No cherry-picking: Report all planned comparisons, not just significant results
6. Reproducibility: Document all analysis steps, parameters, and random seeds

Statistical Tools and Libraries

This workflow requires statistical computing capabilities. Recommended implementations:

Python Stack (recommended for ML/AI research):

python
import numpy as np              # Numerical computing
import pandas as pd             # Data manipulation
import scipy.stats as stats     # Statistical tests
import matplotlib.pyplot as plt # Visualization
import seaborn as sns           # Statistical visualization
from statsmodels.stats import multitest  # Multiple comparison corrections

R Stack (alternative for advanced statistics):

r
library(tidyverse)  # Data manipulation and visualization
library(stats)      # Statistical tests
library(effsize)    # Effect size calculations
library(multcomp)   # Multiple comparisons

Minimum Requirements:

• Statistical test functions (t-test, ANOVA, Mann-Whitney, Kruskal-Wallis, etc.)
• Effect size calculations (Cohen's d, eta-squared, r)
• Multiple comparison corrections (Bonferroni, Holm, FDR)
• Normality tests (Shapiro-Wilk, Kolmogorov-Smirnov)
• Variance homogeneity tests (Levene's test)

Note: If these libraries are not available, the workflow will guide you to use online statistical calculators or manual computation, but automated analysis is strongly recommended for reproducibility.

1. Pre-test before test: Always verify assumptions (normality, homogeneity of variance) before selecting parametric tests
2. No cherry-picking: Report all experimental runs, not just the best ones
3. Publication-quality output: All visualizations follow colorblind-friendly, vector-format, error-bar standards
4. Reproducibility: Track random seeds, hyperparameter ranges, compute resources, and experimental setup

Interactive Preference Collection

Before dispatching to phases, collect analysis preferences:

javascript
const prefResponse = AskUserQuestion({
  question: "How would you like to configure the analysis?",
  options: [
    {
      label: "Analysis Type",
      description: "Select the type of analysis",
      choices: [
        { value: "full", label: "Full Analysis", description: "Complete pipeline: stats + visualization + writing" },
        { value: "comparison", label: "Model Comparison", description: "Focus on comparing multiple models/methods" },
        { value: "ablation", label: "Ablation Study", description: "Focus on component contribution analysis" },
        { value: "visualization", label: "Visualization Only", description: "Generate visualization specs only" }
      ]
    },
    {
      label: "Statistical Reporting",
      description: "Choose error reporting format",
      choices: [
        { value: "sd", label: "Standard Deviation (SD)", description: "Report mean +/- SD (describes data variability)" },
        { value: "se", label: "Standard Error (SE)", description: "Report mean +/- SE (describes mean uncertainty)" },
        { value: "ci", label: "95% Confidence Interval", description: "Report mean [95% CI: low, high]" }
      ]
    },
    {
      label: "Visualization Style",
      description: "Choose visualization palette",
      choices: [
        { value: "okabe-ito", label: "Okabe-Ito (Recommended)", description: "Most widely used colorblind-friendly palette" },
        { value: "paul-tol", label: "Paul Tol", description: "Alternative colorblind-friendly palette" }
      ]
    }
  ]
});

const workflowPreferences = {
  analysisType: prefResponse.analysisType || "full",
  errorFormat: prefResponse.statisticalReporting || "sd",
  colorPalette: prefResponse.visualizationStyle || "okabe-ito"
};

Auto Mode Defaults

When user provides explicit analysis type (e.g., /scholar-experiment comparison):

• analysisType: from argument
• errorFormat: "sd" (standard deviation)
• colorPalette: "okabe-ito"

Execution Flow

> COMPACT DIRECTIVE: Context compression MUST check TodoWrite phase status. > The phase currently marked in_progress is the active execution phase -- preserve its FULL content. > Only compress phases marked completed or pending.

Phase 1: Data Loading

Load experimental data, validate format, perform initial inspection.

• Ref: phases/01-data-loading.md
• Input: raw data files (CSV, JSON, TensorBoard logs, pickle), experiment context
• Output: cleanedData, dataProfile (format, dimensions, completeness, outlier flags)

Phase 2: Statistical Analysis

Compute descriptive statistics, run pre-tests, perform hypothesis testing, calculate effect sizes.

• Ref: phases/02-statistical-analysis.md
• Input: cleanedData, dataProfile, workflowPreferences
• Output: statisticalResults (descriptive stats, test results, effect sizes, multiple comparison corrections)

Phase 3: Visualization

Generate visualization specifications for publication-quality figures and tables.

• Ref: phases/03-visualization.md
• Input: cleanedData, statisticalResults, workflowPreferences
• Output: figureSpecs, tableSpecs, visualization-specs.md

Phase 4: Results Writing

Draft the Results section with proper statistical reporting and figure/table references.

• Ref: phases/04-results-writing.md
• Input: statisticalResults, figureSpecs, tableSpecs, workflowPreferences
• Output: resultsDraft, results-draft.md

Phase 5: Quality Check

Validate analysis completeness, check reproducibility, verify statistical reporting.

• Ref: phases/05-quality-check.md
• Input: all prior outputs
• Output: analysis-report.md (final validated report), quality checklist

Phase Reference Documents (read on-demand when phase executes):

| Phase | Document | Purpose | Compact | |-------|----------|---------|---------| | 1 | phases/01-data-loading.md | Load and validate data | TodoWrite driven | | 2 | phases/02-statistical-analysis.md | Statistical testing | TodoWrite driven + sentinel | | 3 | phases/03-visualization.md | Figure/table specs | TodoWrite driven | | 4 | phases/04-results-writing.md | Draft Results section | TodoWrite driven + sentinel | | 5 | phases/05-quality-check.md | Validate and verify | TodoWrite driven |

Compact Rules:

1. TodoWrite in_progress -> preserve full content, do not compress
2. TodoWrite completed -> may compress to summary
3. sentinel fallback -> phases marked with sentinel contain compact sentinel; if after compact only sentinel remains without full Step protocol, MUST immediately Read("phases/0N-xxx.md") to recover before continuing

Core Rules

1. Never skip pre-tests: Always run normality (Shapiro-Wilk) and homogeneity of variance (Levene) tests before parametric testing
2. Report completely: Every statistical result must include test statistic, degrees of freedom, p-value, and effect size
3. Correct for multiple comparisons: When running multiple tests, apply Bonferroni or FDR correction
4. Use appropriate tests: Select parametric or non-parametric tests based on pre-test results
5. Vector format only: All figure specs must target PDF/EPS output format
6. Colorblind-friendly: All visualizations must use Okabe-Ito or Paul Tol palettes
7. Honest reporting: Report all runs including negative results; never cherry-pick

Input Processing

User provides data path and optional analysis type:

USER INPUT: [data_path] [analysis_type?]

Structured format:
  DATA_PATH: path/to/results/ or path/to/results.csv
  ANALYSIS_TYPE: full | comparison | ablation | visualization
  EXPERIMENT_CONTEXT: (read from experiment session if available)

Supported data formats:

• CSV files: tabular data with headers
• JSON files: structured results objects
• TensorBoard logs: training curves
• Python pickle: complex objects

Data Flow

Phase 1 output:
  cleanedData    → Phase 2, 3
  dataProfile    → Phase 2

Phase 2 output:
  statisticalResults → Phase 3, 4, 5

Phase 3 output:
  figureSpecs    → Phase 4, 5
  tableSpecs     → Phase 4, 5

Phase 4 output:
  resultsDraft   → Phase 5

Phase 5 output:
  analysis-report.md      (final report)
  results-draft.md        (paper-ready text)
  visualization-specs.md  (figure specifications)

TodoWrite Pattern

Phase starts:
  -> Sub-tasks ATTACHED to TodoWrite (in_progress + pending)
  -> Execute sub-tasks sequentially

Phase ends:
  -> Sub-tasks COLLAPSED back to high-level summary (completed)
  -> Next phase begins

Example:

[x] Phase 1: Data Loading (completed - 3 datasets loaded, validated)
[ ] Phase 2: Statistical Analysis (in_progress)
    [ ] 2.1 Descriptive statistics
    [ ] 2.2 Pre-tests (normality, variance)
    [ ] 2.3 Hypothesis testing
    [ ] 2.4 Effect size calculation
    [ ] 2.5 Multiple comparison correction
[ ] Phase 3: Visualization
[ ] Phase 4: Results Writing
[ ] Phase 5: Quality Check

Error Handling

| Error | Recovery | |-------|----------| | Data format unreadable | Ask user for format clarification, try alternative parsers | | Sample size too small (< 3 runs) | Warn user, proceed with non-parametric tests, note limitation | | Normality assumption violated | Switch to non-parametric tests (Wilcoxon, Mann-Whitney U, Kruskal-Wallis) | | Variance homogeneity violated | Switch to Welch's t-test or Welch's ANOVA | | Missing values detected | Report percentage, suggest imputation or exclusion strategy | | Outliers detected | Report via IQR method, run sensitivity analysis (with/without outliers) |

Coordinator Checklist

Before each phase:

• ] Previous phase output available
• ] TodoWrite updated (current phase in_progress)
• ] Phase document read via Read("phases/0N-xxx.md")

After each phase:

• ] Output files generated
• ] TodoWrite updated (current phase completed, next phase in_progress)
• ] Data flow variables passed to next phase

After all phases:

• ] All output files present (analysis-report.md, results-draft.md, visualization-specs.md)
• ] Quality checklist completed
• ] User notified of results

Related Skills

• scholar-writing: Integrates results-draft.md into full paper structure
• scholar-review: Reviews the Results section for quality and completeness