Audited thesis methodology and scaffolded reproducible pipeline

• Day: 2026-05-17
• Time: 11:10 to 11:10
• Project: Dev
• Workspace: WP 2: Operational
• Status: In Progress
• Priority: HIGH
• Assignee: Matías Nehuen Iglesias
• Tags: Thesis, Methodology, Reproducibility, Leakage, Spec-Driven-Development, EPH

Description

Session Goal

Review and harden an income-prediction thesis and its associated EPH modeling scripts, with emphasis on methodological correctness, reproducibility, and a migration path toward a spec-driven research pipeline.

Key Activities

• Performed a methodological audit of the thesis draft, focusing on interpretation of log10 differences, consistency of income variable naming, leakage risks, validation design, and reproducibility gaps.
• Reframed the thesis review as a document triage exercise: treat it as nearly deliverable, then prioritize coherence, critical risks, and completion order.
• Mapped the full experimental workflow for the EPH income modeling pipeline, including ingestion, feature engineering, model families, preprocessing differences, split strategy, and diagnostics.
• Identified technical risks in the current workflow, especially household-level leakage, inconsistent comparability for MLP experiments, and uncertainty about whether income is nominal or deflated.
• Proposed a higher-level architecture: convert the thesis scripts into a lab-grade experimental system with explicit data contracts, reproducible splits, leakage checks, standardized model comparison, and an experiment registry.
• Defined a practical development approach for migration: use legacy scripts as temporary source of truth, create a minimal spec-driven scaffold first, and avoid overengineering before the baseline is stable.

Achievements

• Clarified the main methodological corrections needed in the thesis draft and the order in which they should be addressed.
• Established a coherent interpretation of the current research pipeline and its weak points.
• Produced a concrete implementation direction for turning the thesis code into a reproducible, spec-driven repository.
• Set the foundation for a staged takeover: first stabilize the thesis and baseline experiment, then generalize the codebase into a reusable survey microdata modeling framework.

Pending Tasks

• Rewrite the thesis sections in LaTeX to fix methodological wording, variable labeling, and validation descriptions.
• Resolve the income definition issue: confirm whether the target is nominal or deflated income.
• Audit and eliminate any leakage paths, especially at the household level.
• Standardize preprocessing and model-comparison logic so all experiments are comparable.
• Initialize the migration repository scaffold and implement the minimal reproducible baseline experiment without changing the scientific design.

Evidence

• source_file=2026-05-17.sessions.jsonl, line_number=2, event_count=0, session_id=3ca792f53d9336a5201a6d1b795981ae3dc35212aa2d3e99d59d0ac49826fbc2
• event_ids: []