Blame for README.md

---
license: apache-2.0
tags: [szl-holdings, physics, benchmark, receipts-backed, synthetic]
---

# physics-bench-micro

> 120 synthetic physics problems (kinematics, energy, circuits) with programmatically
> verified answers - the estate's first physics benchmark, rights-clean by construction.

## Why synthetic

Every problem is generated and every answer is *computed*, not copied - zero rights
questions, zero contamination from published benchmarks, infinitely regenerable with
a new seed. Synthetic first, then rights-cleared external microbenches (Matbench,
WeatherBench 2) on top, per the leader audit.

## The verification contract

Each row carries explicit `params`; `verify.py` recomputes every answer from them -
a model answer passes only within tolerance. No regex extraction, no ambiguity.

## Honest labels

- **License:** Apache-2.0 (generated by SZL Holdings, no third-party content)
- **Seed:** 20261110 - a new seed produces a fresh, equally valid battery
- **Contamination-proof:** regenerable on demand, so leakage is detectable

*Generated by the estate's benchmark lane; answers machine-verified before publish.*

## Card completeness (2026-11-12)

This section adds required card sections the card above did not have. It states only what files in this repository and the estate's own receipt datasets show; nothing here is a new result.

### Provenance

Not yet documented. As of 2026-11-12 the card has no provenance statement: who created this data, how it was collected or generated, from which sources, and what filtering was applied. The owner should add this before the dataset is relied on.

### Schema

Read from the data files in this repository on 2026-11-12 (field types are from the first record):

| File | Records | Size | Fields |
|---|---|---|---|
| `data/problems.json` | 120 | 40.5 KB | `id` (str), `topic` (str), `problem` (str), `params` (dict), `answer` (float), `solution` (str) |