Audited HGB notebook evidence and experiment design

• Day: 2026-06-10
• Time: 11:45 to 12:30
• Project: Dev
• Workspace: WP 1: Strategic / Growth & Development
• Status: Completed
• Priority: MEDIUM
• Assignee: Matías Nehuen Iglesias
• Tags: Hgb, Notebook-Audit, Experiment-Design, Cross-Validation, Thesis-Methodology, Ols-Governance

Description

Session Goal

Review a set of notebook outputs and YAML experiment artifacts for HistGradientBoosting / regression-model screening, with the goal of separating scientific interpretation from statistical audit and figure-quality checks.

Key Activities

• Defined a three-layer audit framework for HGB notebook outputs:
  1. scientific interpretation,
  2. statistical / ML audit,
  3. figure quality.
• Proposed a first-pass notebook structural review focused on markdown intent, code organization, embedded outputs, and where claims are generated.
• Reframed YAML experiment files as scientific designs, not just configs, by checking sampling/folds, hyperparameter axes, cross-run comparability, and whether each run has a distinct purpose.
• Added evidence-hierarchy guidance for experiment families: exploratory maps, directed sweeps, confirmatory benchmarks, and geo/feature ablations.
• Specified run-selection logic for notebook artifact discovery, including pattern matching, required-artifact checks, audit-table logging, and a manual override mechanism for pinned runs.
• Drafted a reusable thesis-evidence template that separates methodology, results, and discussion to avoid repetition and keep claims aligned with evidence.
• Added governance guidance for OLS feature coding in the EPH notebook, distinguishing numeric, binary, categorical, and fixed-effect variables so coded integers are not treated as continuous measures.

Achievements

• Clarified that several HGB results should be treated as exploratory rather than leaderboard-grade because some runs are not directly comparable.
• Captured the main scientific interpretation that HGB improves predictive performance over linear models, but still shows distributional compression and tail bias.
• Established a thesis-ready structure for evaluating model evidence and for deciding when to revise plots, rerun experiments, or extend backend artifacts.
• Produced a cleaner decision framework for comparing YAML experiment runs without changing the underlying setup unnecessarily.

Pending Tasks

• Apply the audit framework to the actual notebook outputs and figure set.
• Mark non-comparable HGB runs explicitly in the thesis or notebook narrative.
• Decide which plots need revision versus which experiments need reruns.
• Finalize the canonical anchor configuration and confirmatory benchmark run for the experiment family.
• Implement or validate the OLS governance cells and sensitivity specifications in the EPH notebook.

Evidence

• source_file=2026-06-10.sessions.jsonl, line_number=0, event_count=0, session_id=fd0075964b13469b12567fcbff99458fc5416bbd497f53a26c66458f1894a628
• event_ids: []