Math + Econ Reasoning Portfolio (NDA-safe excerpts)

This repository contains carefully selected, NDA-safe excerpts from a larger body of math- and economics-based analytical work used to design, evaluate, and verify LLM reasoning and numerical reliability.

The materials here focus on the final verification layer of much broader analyses:

• reduced-form problem statements
• distilled numerical cores
• deterministic validation logic

The original tasks were typically more complex, data-driven, and multi-stage, but are presented here in simplified, synthetic form to remain fully public and NDA-compliant.

This is best viewed as a portfolio of evaluation artefacts rather than a full reproduction of the underlying research pipelines.

---

What this repo demonstrates

• Non-trivial numerical methods
  (bisection / root-finding, verification inequalities, Monte Carlo sanity checks)

• Reproducible reference solutions
  with explicit tolerances and deterministic outputs

• Answer validation & scoring logic
  similar to LLM evaluation / grading pipelines

• Failure-mode awareness
  (bounds, monotonicity assumptions, bracketing errors, model misspecification)

• Clean Python engineering
  (tests, CI, no side effects on import, CLI + JSON outputs)

---

Important context (NDA-safe clarification)

The problems in this repository are not full research problems and not client deliverables.

They are:

• condensed representations of larger analytical tasks
• using synthetic or normalized parameters
• stripped of proprietary data, domain specifics, and contextual complexity

In practice, the original tasks:

• involved richer stochastic structure or real datasets
• required additional constraints, diagnostics, and robustness checks
• were embedded in broader modeling or evaluation workflows

What you see here corresponds to the final reasoning and verification step — the part most relevant for assessing LLM numerical reasoning, correctness, and failure behavior.

---

Repository structure

• problems/ — problem statements + failure modes
• src/econ_math_portfolio/models/ — model implementations (no code runs on import)
• validators/ — validators calling model code
• originals/ — original standalone scripts kept for transparency (not imported)
• rubrics/ — scoring rules inspired by LLM evaluation setups
• tests/ — pytest
• .github/workflows/ci.yml — CI (Python 3.10–3.12)

---

Quickstart

python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
pytest
python -m econ_math_portfolio list

Development

Run linting and tests:

ruff check .
ruff format --check .
pytest

---

CLI usage

python -m econ_math_portfolio reference credit_var_quantile
python -m econ_math_portfolio validate cpi_target_discount 0.26191

---

Notebook demo

jupyter notebook notebooks/demo.ipynb

---

JSON output (tool-calling friendly)

python -m econ_math_portfolio list --json
python -m econ_math_portfolio reference cpi_target_discount --json
python -m econ_math_portfolio validate cpi_target_discount 0.26191 --json

---

Scoring rubric (LLM evaluation style)

python -m econ_math_portfolio score submissions/contract_good.json --json

Submission format:

{
  "task_id": "cpi_target_discount",
  "answer": 0.26191,
  "explanation": "optional short explanation"
}

---

How to interpret this portfolio

This repository is a curated slice of real analytical work, intentionally focused on:

• reasoning clarity
• numerical correctness
• verification and evaluation

The goal is to show how problems are checked, not just how they are solved.