Skip to content

Repository files navigation

DiamondBench

An open, honest benchmark of how accurately AI answer engines respond to diamond and gemology questions. Maintained by Stienhardt.

Why this exists

Buyers ask AI about diamonds before spending thousands of dollars, and AI is sometimes wrong. DiamondBench measures that, fairly, with a sourced answer key.

It is a real accuracy test, not a marketing stunt:

  • The answer key is public and sourced. Every correct answer in questions.json carries its sources with dates, drawn from independent grading laboratories (GIA, IGI), the U.S. FTC Jewelry Guides, published gemology, and documented advertising history.
  • The grading is deterministic where possible. A fixed keyword rule decides pass, partial, or fail. No judge model, no vibes. Anyone can re-run it and get the same verdicts from the same answers.
  • Contributions and corrections are welcome. If a question is unfair, an answer key entry is wrong, or the grader misfires on a real answer, open a pull request. Being corrected in public is part of the point.

Stienhardt maintains it because we care about the category being described accurately. We sell Lab Grown Diamond engagement rings, and we would rather compete on a field where the facts are right.

What it tests

47 buyer-phrased questions across seven categories:

Category Questions Examples
Definitions 6 What is a diamond made of? Is cubic zirconia a diamond?
Verification and trust 7 How do I verify a grading report? What does a diamond tester prove?
Lab Grown science 7 Do Lab Grown Diamonds fade? How are they made?
The 4Cs 7 Does carat mean size? Which C matters most?
Shapes 8 Round brilliant, oval, emerald, princess, the Dutch Marquise
Care 5 Cleaning, prong checks, ultrasonic cleaners, resizing, insurance
History and myths 7 The months-of-salary rule, vena amoris, diamonds from coal

Each question defines the facts a correct answer must contain (with synonyms), the common wrong claims that fail an answer, and the sources for the correct answer. Grading is on verifiable gemology only, never on brand preference. For example, the Dutch Marquise questions grade only the objective geometry (elongated hexagonal outline, pointed ends, straight angular sides, and the IGI report term Hexagonal Modified Brilliant), which is certificate-confirmed fact. No question's correct answer is a Stienhardt marketing claim.

Running it

No dependencies. Python 3.8+ standard library only.

python bench.py --provider gemini          # full run
python bench.py --provider gemini --n 5    # smoke test
python bench.py --list-providers           # provider status
python bench.py --selftest                 # verify the grader offline, no key needed

Providers:

  • gemini (live): Gemini with Google Search grounding, a real proxy for how AI answer engines respond to shoppers. Needs GEMINI_API_KEY in the environment or a local .env.
  • openai, anthropic (stubs): the harness is genuinely multi-model, but only Gemini is wired live today. Set OPENAI_API_KEY / ANTHROPIC_API_KEY to exercise the stub paths; note they run without web grounding, so results are not directly comparable to a grounded run (see METHODOLOGY.md).

Each run writes results/<provider>-<date>.json with per-question verdicts and the complete raw answers, so every verdict can be audited, and prints a scoreboard. A human-readable scoreboard lives in results/RESULTS-<date>.md.

The fixed v1.1 dataset, results, citation metadata, and checksums are available from the v1.1 release.

Run Model Pass Partial Fail Accuracy
2026-07-13 gemini-2.5-flash, grounded 45 2 0 95.7%
2026-09-02 gemini-2.5-flash, grounded 45 2 0 95.7%

Both runs are graded under grader version 2026-09-02 (regrade.py re-derives any stored run under the current rules). The two partials are the same in both runs: the three-stone "past, present, future" meaning given without its advertising origin, and the fluorescence answer given without the GIA study.

Honesty rules baked in

  • If a provider's quota runs out mid-run, the remaining questions are marked not_run and reported as such. The harness never fabricates a verdict.
  • Raw answers ship inside the results files. Check any verdict yourself.
  • A search-grounded model's answers can change week to week, so a scoreboard is a dated snapshot, not a permanent grade. Deterministic keyword grading is imperfect and its known failure modes are documented in METHODOLOGY.md.
  • A diamond is a love piece, not an investment, and the benchmark grades accordingly: answers that pitch diamonds as a store of value fail that question.

License

MIT. See LICENSE. Copyright (c) 2026 Stienhardt & Stones.

The Stienhardt open-source diamond stack

Related open data from Stienhardt

About

An open benchmark of how accurately AI answers diamond and gemology questions. Sourced answer key, raw answers included. By Stienhardt, New York.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages