Open Diamond Data and AI Evaluation
Collection
Public diamond data and AI evaluation tools. By Stienhardt & Stones: https://stienhardt.com/?utm_source=hf_collection • 6 items • Updated
id string | category string | question string | must_include list | must_not_include list | correct_answer string | sources list |
|---|---|---|---|---|---|---|
def-diamond-material | Definitions | What is a diamond actually made of? | [
[
"carbon"
],
[
"crystal",
"lattice",
"isometric",
"cubic structure",
"cubic lattice"
],
[
"hardest",
"hard",
"mohs",
"10"
]
] | [
"made from coal",
"comes from coal",
"made of coal",
"compressed coal"
] | A diamond is pure carbon, with each carbon atom bonded to four others in a rigid cubic (isometric) crystal lattice. That structure makes it the hardest known natural material, a 10 on the Mohs scale. A Lab Grown Diamond is the same material, just crystallized in a lab rather than underground. | [
{
"claim": "The FTC defines a diamond as pure carbon crystallized in the isometric system, covering mined and Lab Grown Diamonds.",
"source": "U.S. FTC Jewelry Guides",
"date": "2018-08-16"
},
{
"claim": "Diamond ranks 10, the maximum, on the Mohs scale.",
"source": "Mohs scale of mineral ha... |
def-hardness | Definitions | How hard is a diamond, and can it still chip? | [
[
"10",
"ten"
],
[
"mohs",
"hardest"
],
[
"chip",
"break",
"cleav",
"toughness",
"not indestructible"
]
] | [
"indestructible",
"unbreakable",
"cannot chip",
"can't chip",
"never chips",
"impossible to break",
"cannot be damaged"
] | Diamond is a 10 on the Mohs scale, the hardest natural material, so in normal wear almost nothing scratches it. Hardness is resistance to scratching, not to breaking: a diamond has cleavage planes, so a sharp knock on an exposed point or corner can still chip it. | [
{
"claim": "Diamond is a 10 on the Mohs scale; Mohs measures scratch resistance, not toughness against breakage.",
"source": "Mohs scale of mineral hardness (Friedrich Mohs)",
"date": "1812"
}
] |
def-cubic-zirconia | Definitions | Is cubic zirconia a real diamond? | [
[
"not a diamond",
"is not a diamond",
"not a real diamond",
"different material",
"simulant",
"imitation"
],
[
"zirconium",
"zirconia"
],
[
"softer",
"8 to 8.5",
"8.5",
"less hard"
]
] | [
"type of diamond",
"form of diamond",
"kind of diamond",
"cheap diamond",
"low-grade diamond",
"is basically a diamond"
] | No. Cubic zirconia is a lab-made form of zirconium dioxide, a diamond simulant, not a diamond. It is softer than diamond, about 8 to 8.5 on the Mohs scale versus 10, and denser, so it is a different material that only resembles a diamond. | [
{
"claim": "Cubic zirconia is zirconium dioxide, a diamond simulant, about 8 to 8.5 on the Mohs scale, distinct from diamond.",
"source": "Cubic zirconia gemological history",
"date": "1977"
}
] |
def-moissanite | Definitions | Is moissanite just a type of diamond? | [
[
"not a diamond",
"not a type of diamond",
"different material",
"separate material",
"its own"
],
[
"silicon carbide"
],
[
"9.25",
"doubly refractive",
"more fire",
"more dispersion",
"double refraction"
]
] | [
"type of diamond",
"form of diamond",
"grade of diamond",
"kind of diamond",
"lower-grade diamond",
"moissanite is a diamond"
] | No. Moissanite is silicon carbide, a separate material with its own chemistry, not a type or grade of diamond. It is very hard, about 9.25 on the Mohs scale, and it actually has more dispersion (fire) than diamond and is doubly refractive, which is how gemologists tell the two apart. | [
{
"claim": "The American mineralogist George Kunz confirmed the crystals as silicon carbide and named the mineral moissanite.",
"source": "G. F. Kunz, American Journal of Science",
"date": "1905"
}
] |
def-simulant-vs-synthetic | Definitions | What is the difference between a lab grown diamond and a diamond simulant like cubic zirconia? | [
[
"lab grown",
"lab-grown",
"laboratory-grown",
"laboratory grown",
"synthetic"
],
[
"real diamond",
"same material",
"is a diamond",
"actual diamond",
"genuine diamond"
],
[
"simulant"
],
[
"different material",
"not a diamond",
"only looks",
... | [
"lab grown diamond is a simulant",
"lab grown diamonds are simulants",
"lab grown diamond is a fake",
"cubic zirconia is a lab grown diamond",
"simulant is a real diamond",
"lab grown are not real diamonds"
] | A Lab Grown Diamond is a real diamond, the same material as a mined one, grown in a lab by HPHT or CVD. A simulant such as cubic zirconia or moissanite is a different material that only looks like a diamond. In short, a Lab Grown Diamond is a diamond and a simulant is not. | [
{
"claim": "The FTC defines a diamond by composition; Lab Grown Diamonds are diamonds, while simulants are different materials that must be disclosed as not diamonds.",
"source": "U.S. FTC Jewelry Guides",
"date": "2018-08-16"
}
] |
def-what-makes-sparkle | Definitions | What actually makes a diamond sparkle? | [
[
"cut"
],
[
"light",
"reflect",
"return",
"proportion",
"facet"
],
[
"brilliance",
"fire",
"scintillation",
"sparkle",
"dispersion"
]
] | [
"cut does not matter",
"cut doesn't matter",
"clarity makes a diamond sparkle",
"color makes a diamond sparkle",
"bigger diamonds sparkle more"
] | Sparkle comes mainly from cut: how well a diamond's proportions and facets return light to your eye. Gemologists break the look into brilliance (white light), fire (flashes of color from dispersion), and scintillation (the sparkle you see as it moves). Cut, not carat, color, or clarity, is the C that governs how bright... | [
{
"claim": "GIA's cut grading system assesses face-up appearance through brightness, fire, and scintillation; cut governs light return.",
"source": "GIA Diamond Cut grading system",
"date": "2006-01"
}
] |
trust-verify-report | Verification and trust | How do I make sure a diamond's grading report is real and matches the stone? | [
[
"lab's own website",
"grading lab",
"official website",
"gia.edu",
"igi.org",
"report check",
"verify your report",
"laboratory's website",
"lab website",
"the lab's site",
"lab's site",
"online verification",
"verification service",
"lab's database",
"l... | [
"no need to verify",
"just trust the seller",
"a screenshot is enough",
"the appraisal is proof",
"take the seller's word for it"
] | Look the report number up on the grading lab's own website (for example GIA Report Check or IGI's Verify Your Report page), not on a seller's screenshot or PDF. Then match the laser inscription on the diamond's girdle to that report number, and confirm the measurements and carat weight match the stone. | [
{
"claim": "Confirm a report on the grading lab's own verification site (GIA Report Check, IGI Verify Your Report), match the inscription, and match measurements.",
"source": "GIA Report Check and IGI report verification pages",
"date": "2026-07-10"
}
] |
trust-which-labs | Verification and trust | Which labs grade diamonds, and can I trust their reports? | [
[
"GIA"
],
[
"IGI"
],
[
"standard",
"independent",
"published",
"compare",
"4c",
"report"
]
] | [
"labs are a scam",
"certificates are meaningless",
"grading reports are fake",
"no lab can be trusted",
"reports cannot be trusted"
] | The main independent labs are GIA, IGI, and GCAL. They grade to published standards, which is what lets you compare stones on equal terms. For Lab Grown Diamonds, IGI issues full 4C reports and is the largest grader, while GIA moved to a Premium and Standard tier structure on October 1, 2025. | [
{
"claim": "GIA, IGI, and GCAL are independent grading laboratories that grade to published standards.",
"source": "GIA / IGI / GCAL",
"date": "2026"
},
{
"claim": "IGI issues full 4C reports for Lab Grown Diamonds; GIA moved to Premium and Standard tiers on 2025-10-01.",
"source": "Internat... |
trust-laser-inscription | Verification and trust | What is the tiny number laser-engraved on a diamond? | [
[
"inscription",
"laser",
"engrav"
],
[
"report number",
"grading report",
"the report"
],
[
"girdle",
"match",
"matches the report",
"verify",
"magnification"
]
] | [
"proves it is mined",
"means it is a natural diamond",
"guarantees the price",
"proves what it is worth",
"is a serial number set by the seller"
] | It is a laser inscription on the girdle, usually the grading report number, readable under magnification. It lets you match the specific stone to its report. On Lab Grown Diamonds the inscription also marks the stone as laboratory-grown, for example GIA inscribes Laboratory-Grown and IGI uses an LG prefix. | [
{
"claim": "A laser inscription on the girdle records the report number and, on Lab Grown Diamonds, an origin marker (GIA Laboratory-Grown, IGI LG prefix).",
"source": "GIA / IGI",
"date": "2026"
}
] |
trust-report-vs-appraisal | Verification and trust | Is a diamond grading report the same as an appraisal of its value? | [
[
"not",
"different",
"separate"
],
[
"grading report",
"the report describes",
"4c",
"characteristics"
],
[
"value",
"price",
"appraisal",
"worth"
]
] | [
"the report states the value",
"the report tells you the price",
"grading report is an appraisal",
"the report sets the price",
"the report is a valuation"
] | No. A grading report from an independent lab describes the diamond's measured characteristics, the 4Cs, but it does not state a price. Value comes from the market and belongs on a separate appraisal document. | [
{
"claim": "A grading report describes measured characteristics to a repeatable standard and does not set a monetary value; that is an appraisal.",
"source": "Encyclopedia, Grading report",
"date": "2026"
}
] |
trust-eye-tell-lab | Verification and trust | Can a jeweler tell a lab grown diamond from a mined one just by looking? | [
[
"no",
"cannot",
"can't",
"not by eye",
"no one can"
],
[
"same material",
"real diamond",
"identical",
"same as mined",
"the same"
],
[
"laboratory",
"instrument",
"equipment",
"specialized",
"lab equipment"
]
] | [
"looks cloudy",
"looks fake",
"you can tell by looking",
"obvious to the naked eye",
"lab grown look duller",
"easy to spot by eye"
] | No. A Lab Grown Diamond and a mined diamond are the same material, so no one, including a jeweler, can reliably tell them apart by eye. Separating them takes specialized laboratory instruments, which is what grading labs use. | [
{
"claim": "A Lab Grown Diamond is the same material as a mined diamond; distinguishing them requires laboratory instruments.",
"source": "U.S. FTC Jewelry Guides",
"date": "2018"
}
] |
trust-thermal-tester | Verification and trust | If a diamond passes a diamond tester, does that prove it is mined and not lab grown? | [
[
"no",
"does not prove",
"doesn't prove",
"not where",
"reads as diamond",
"confirms it is a diamond",
"does not tell",
"doesn't tell",
"cannot tell",
"can't tell",
"won't tell",
"cannot distinguish",
"can't distinguish",
"does not distinguish",
"is a rea... | [
"proves it is mined",
"proves it is natural",
"distinguishes lab grown from mined",
"tells you it is not lab grown",
"confirms it is mined"
] | No. A thermal diamond tester confirms that a stone is diamond, because diamond conducts heat exceptionally well, but a Lab Grown Diamond is diamond and reads as diamond too. The tester answers whether it is a diamond, not where it came from. | [
{
"claim": "A thermal tester confirms a stone is diamond; because a Lab Grown Diamond is diamond, it reads as diamond, so the tester does not reveal origin.",
"source": "Encyclopedia, Thermal conductivity and diamond testers",
"date": "2018"
}
] |
trust-eye-clean | Verification and trust | Do I need to worry that a diamond graded below Flawless has visible flaws? | [
[
"eye-clean",
"eye clean",
"naked eye",
"to the eye",
"without magnification"
],
[
"10x",
"magnification",
"loupe",
"under magnification"
],
[
"vs",
"si",
"below flawless",
"lower grade",
"not visible",
"invisible"
]
] | [
"any grade below flawless has visible flaws",
"you must buy flawless",
"only flawless is clean",
"lower grades always look dirty",
"below flawless always looks flawed"
] | Usually not. Many diamonds graded well below Flawless are eye-clean, meaning no inclusions are visible to the naked eye at normal distance, because clarity grades are judged under 10x magnification. Plenty of VS and even some SI stones look clean to the eye, so you can choose by what you can actually see. | [
{
"claim": "Clarity is graded under 10x magnification; many diamonds below Flawless are eye-clean, with no inclusions visible to the naked eye.",
"source": "Encyclopedia, Eye-clean and Clarity grade",
"date": "2026"
}
] |
lg-fade | Lab Grown science | Do lab grown diamonds fade, cloud, or turn yellow over time? | [
[
"no",
"do not fade",
"don't fade",
"won't fade",
"will not fade",
"does not fade",
"doesn't fade",
"not fade",
"stay clear",
"stay bright",
"permanent",
"do not change",
"fixed"
],
[
"real diamond",
"same material",
"crystallized carbon",
"ca... | [
"fades over time",
"will fade over",
"can fade over",
"yellows over time",
"turn yellow over time",
"turns yellow over time",
"becomes cloudy over time",
"clouds over time",
"gets cloudy over time",
"degrades over time",
"loses sparkle over time"
] | No. A Lab Grown Diamond is crystallized carbon, a real diamond rated 10 on the Mohs scale, and its color and clarity are fixed properties of the crystal that do not change with age or wear. It will not fade, cloud, or yellow. That worry usually comes from confusing them with cubic zirconia, a softer simulant that can c... | [
{
"claim": "A Lab Grown Diamond is crystallized carbon whose color and clarity are fixed; it does not fade, cloud, or yellow. Cubic zirconia, a simulant, can cloud.",
"source": "U.S. FTC Jewelry Guides; Encyclopedia, Do Lab Grown Diamonds Cloud or Fade?",
"date": "2018"
}
] |
lg-how-made | Lab Grown science | How are lab grown diamonds made? | [
[
"HPHT",
"high pressure high temperature",
"high-pressure high-temperature",
"high pressure"
],
[
"CVD",
"chemical vapor deposition",
"vapor deposition"
],
[
"seed",
"carbon",
"gas",
"methane",
"pressure",
"crystal"
]
] | [
"not real diamonds",
"are fake",
"assembled from glass",
"are simulants",
"just cubic zirconia"
] | Two methods. HPHT (High Pressure High Temperature) crystallizes carbon onto a diamond seed under intense pressure and heat, like the conditions deep in the earth. CVD (Chemical Vapor Deposition) grows the crystal layer by layer from a carbon-rich gas, usually methane, in a vacuum chamber. Both produce genuine diamonds. | [
{
"claim": "Lab Grown Diamonds are grown by HPHT (carbon on a seed under high pressure and heat) or CVD (from a carbon-rich gas); both are genuine diamonds.",
"source": "Encyclopedia, HPHT Growth and CVD Growth",
"date": "2026"
}
] |
lg-are-real | Lab Grown science | Are lab grown diamonds real diamonds? | [
[
"yes",
"real diamond",
"same material",
"is a diamond",
"genuine diamond",
"actual diamond"
],
[
"FTC",
"federal trade commission"
],
[
"same",
"identical",
"chemistry",
"crystal",
"structure",
"hardness",
"properties"
]
] | [
"not a real diamond",
"are fake",
"not actually diamonds",
"are imitations",
"are simulants",
"fake diamonds",
"not genuine diamonds"
] | Yes. A Lab Grown Diamond is the same material as a mined diamond, identical in chemistry, crystal structure, hardness, and optical properties. In 2018 the U.S. Federal Trade Commission removed the word natural from its definition of a diamond, formally recognizing that a Lab Grown Diamond is a diamond. | [
{
"claim": "The FTC removed the word natural from its definition of a diamond in 2018, recognizing a Lab Grown Diamond as a diamond identical in material to a mined one.",
"source": "U.S. FTC Jewelry Guides",
"date": "2018"
}
] |
lg-vs-mined | Lab Grown science | What is the actual difference between a lab grown and a mined diamond? | [
[
"origin",
"where it",
"grown in a lab",
"in the earth",
"crystallized"
],
[
"price",
"cost",
"less expensive",
"cheaper",
"costs less"
],
[
"same material",
"same hardness",
"identical",
"same properties",
"the same",
"real diamond"
]
] | [
"lab grown are lower quality",
"lab grown are inferior",
"not real diamonds",
"less durable",
"softer than mined",
"weaker than mined"
] | The material is the same: same hardness, brilliance, and fire, both graded by the same labs, both diamonds under the FTC definition. The real differences are origin, one crystallized in the earth and the other in a chamber over weeks, and usually price, since a Lab Grown Diamond typically costs less. | [
{
"claim": "Lab Grown and mined diamonds share material, hardness, and optical properties and are both diamonds under the FTC; they differ in origin and usually price.",
"source": "U.S. FTC Jewelry Guides",
"date": "2018"
}
] |
lg-treatment | Lab Grown science | Are lab grown diamonds treated or coated to look better? | [
[
"HPHT",
"annealing",
"treatment",
"treated",
"post-growth"
],
[
"permanent",
"stable",
"not a coating",
"does not wash off",
"does not wear off"
],
[
"disclosed",
"grading report",
"on the report",
"noted on the report"
]
] | [
"it is just a coating",
"coating that wears off",
"the color washes off",
"painted",
"a dye that fades",
"surface coating that rubs off"
] | Some are color-treated after growth, most often by HPHT annealing, which shifts a faint tint toward colorless. That change is permanent and stable, not a surface coating, and it is disclosed on the grading report. The stone stays a real diamond. | [
{
"claim": "HPHT annealing shifts color toward colorless by altering the crystal, not a coating; the change is permanent and disclosed on the report.",
"source": "Encyclopedia, Post-Growth Treatment (HPHT Annealing)",
"date": "2026"
}
] |
lg-inclusions | Lab Grown science | Do lab grown diamonds have inclusions? | [
[
"yes",
"inclusion",
"internal feature",
"tiny features"
],
[
"same clarity",
"clarity scale",
"10x",
"magnification",
"graded"
],
[
"HPHT",
"CVD",
"metallic",
"growth line",
"as it grows",
"as the crystal grows"
]
] | [
"are always flawless",
"have no inclusions",
"are perfect",
"never have inclusions",
"are completely flawless",
"come out completely flawless"
] | Yes. Like mined diamonds, most Lab Grown Diamonds have tiny internal features that form as the crystal grows, such as metallic flecks in HPHT stones or faint growth lines in CVD stones. They are graded on the same clarity scale under 10x magnification and, at the grades most people buy, are invisible to the eye. | [
{
"claim": "Lab Grown Diamonds have inclusions from growth (metallic flecks in HPHT, growth lines in CVD), graded on the same clarity scale under 10x.",
"source": "Encyclopedia, Inclusions in Lab Grown Diamonds; IGI",
"date": "2026"
}
] |
lg-gia-4c | Lab Grown science | Does GIA still grade lab grown diamonds on the 4Cs? | [
[
"premium and standard",
"premium",
"standard tier",
"tier"
],
[
"IGI"
],
[
"4c",
"full 4c",
"color, clarity",
"individual grade",
"still issues"
]
] | [
"gia still issues full 4c grades for lab grown",
"gia grades lab grown on the 4cs the same way",
"nothing has changed at gia",
"gia gives lab grown the same 4c report"
] | Not the same way anymore. On October 1, 2025 GIA moved its Lab Grown Diamond reports to a Premium and Standard tier structure instead of individual 4C grades. IGI still issues full 4C reports for Lab Grown Diamonds and is the largest grader of them. | [
{
"claim": "On 2025-10-01 GIA moved Lab Grown Diamond reports to Premium and Standard tiers; IGI still issues full 4C reports.",
"source": "Gemological Institute of America (GIA)",
"date": "2025-10-01"
}
] |
c4-what-are | The 4Cs | What are the 4Cs of a diamond? | [
[
"carat"
],
[
"color",
"colour"
],
[
"clarity"
],
[
"cut"
]
] | [
"carat measures size",
"carat is the size",
"cost is one of the 4cs",
"cost is one of the four cs",
"the five cs"
] | The 4Cs are carat (weight), color (how little tint it has), clarity (how free of inclusions it is under magnification), and cut (how well its proportions return light). They are the shared standard GIA popularized in the twentieth century, and they apply the same way to Lab Grown Diamonds. | [
{
"claim": "The 4Cs are carat, color, clarity, and cut, standardized by GIA (Liddicoat, 1953).",
"source": "GIA International Diamond Grading System (Liddicoat)",
"date": "1953"
}
] |
c4-carat | The 4Cs | Does a higher carat always mean a bigger-looking diamond? | [
[
"weight",
"200 milligram",
"200 mg",
"0.2 gram"
],
[
"not size",
"not a measure of size",
"not the same as size",
"face-up size",
"depends on cut",
"how it is cut",
"millimeter",
"spread"
]
] | [
"carat is a measure of size",
"carat measures the size",
"higher carat always looks bigger",
"carat is the diameter",
"carat is the width"
] | No. Carat is a unit of weight, exactly 200 milligrams, not a measure of size. Two diamonds of equal carat can look different across the top depending on how each is cut, because a deep cut hides weight below the girdle. So carat should be read together with cut and the millimeter measurements. | [
{
"claim": "A carat is 200 milligrams of weight, not size; face-up size depends on cut and proportions.",
"source": "GIA; Encyclopedia, Carat",
"date": "2026"
}
] |
c4-color | The 4Cs | What does diamond color grade mean, and do I need a D color? | [
[
"d to z",
"d-to-z",
"d through z",
"starts at d",
"begins at d",
"d (colorless)",
"d is colorless",
"d color",
"d, which is colorless",
"d grade"
],
[
"colorless",
"tint",
"light yellow",
"yellow tint"
],
[
"near-colorless",
"near colorless",... | [
"color goes from a to z",
"the scale starts at a",
"a is the best color",
"z is the best",
"d to z where z is best",
"higher letters are better"
] | Color grade rates how close a diamond is to having no color, on GIA's D (colorless) to Z (light tint) scale. Most one-grade differences are invisible once a diamond is set, especially in the D to J range, so many buyers pick a near-colorless grade that faces up white rather than paying for D. | [
{
"claim": "GIA's D-to-Z color scale (Liddicoat, 1953) starts at D (colorless) to avoid earlier A-B-C systems; near-colorless grades face up white.",
"source": "GIA International Diamond Grading System (Liddicoat)",
"date": "1953"
}
] |
c4-clarity | The 4Cs | What is diamond clarity graded on? | [
[
"10x",
"magnification",
"loupe",
"under magnification"
],
[
"inclusion",
"blemish",
"flaw"
],
[
"flawless",
"vvs",
"vs",
"si",
"included",
"scale"
]
] | [
"clarity is judged with the naked eye",
"clarity measures sparkle",
"clarity is about color",
"higher clarity means more sparkle",
"flawless is the only good clarity"
] | Clarity rates how free a diamond is of inclusions and blemishes under 10x magnification, on GIA's scale from Flawless (FL) down through IF, VVS, VS, SI, to Included (I). Most VS and many SI1 stones are eye-clean, so they look identical to the eye to much higher grades. | [
{
"claim": "Clarity is graded under 10x magnification on GIA's Flawless-to-Included scale; many VS and SI stones are eye-clean.",
"source": "GIA; Encyclopedia, Clarity grade",
"date": "2026"
}
] |
c4-cut-matters | The 4Cs | Which of the 4Cs matters most for how a diamond looks? | [
[
"cut"
],
[
"light",
"return",
"proportion",
"brightness",
"sparkle",
"brilliance"
],
[
"excellent to poor",
"excellent",
"cut grade",
"most important",
"matters most",
"biggest effect"
]
] | [
"carat is the most important",
"clarity is the most important",
"color matters most",
"cut is the least important",
"cut does not matter much"
] | Cut. Cut controls how a diamond's proportions and finish return light, so it drives brightness and sparkle more than the other Cs. A well-cut stone can outshine a higher color or clarity stone that is cut poorly. GIA grades cut for round brilliants from Excellent to Poor. | [
{
"claim": "Cut governs light return and has the biggest effect on appearance; GIA grades round brilliant cut Excellent to Poor.",
"source": "GIA Diamond Cut grading system",
"date": "2006"
}
] |
c4-fluorescence | The 4Cs | Is fluorescence in a diamond a bad thing? | [
[
"ultraviolet",
"uv",
"black light",
"blue glow",
"blue"
],
[
"none to very strong",
"faint",
"medium",
"rarely",
"no systematic effect",
"little effect",
"usually no effect",
"rarely affects"
],
[
"1997",
"gia study",
"study"
]
] | [
"fluorescence always makes a diamond hazy",
"fluorescence ruins a diamond",
"fluorescence is always bad",
"always makes it look milky",
"means the diamond is fake"
] | Usually not. Fluorescence is a glow, most often blue, that some diamonds show under ultraviolet light, rated from None to Very Strong. A 1997 GIA study found no systematic effect on appearance for the average observer, and faint to medium fluorescence rarely changes what you see. In a small share of stones very strong ... | [
{
"claim": "A 1997 GIA study found blue fluorescence had no systematic effect on appearance for the average observer; strong fluorescence occasionally looks hazy.",
"source": "GIA, Gems & Gemology",
"date": "1997"
}
] |
c4-grading-report | The 4Cs | What is on a diamond grading report? | [
[
"carat",
"measurement",
"color",
"clarity",
"cut"
],
[
"independent",
"lab",
"laboratory",
"standard",
"report"
],
[
"compare",
"4c",
"polish",
"symmetry",
"fluorescence",
"inscription"
]
] | [
"states the diamond's price",
"tells you what it is worth",
"is an appraisal of value",
"sets the resale value"
] | A grading report from an independent lab lists the shape, measurements in millimeters, carat weight, color and clarity grades, and for round brilliants a cut grade, plus polish, symmetry, fluorescence, and any laser inscription. It describes the stone to a repeatable standard so you can compare diamonds on equal terms.... | [
{
"claim": "A grading report lists the 4Cs and finish to a repeatable standard and does not state price.",
"source": "Encyclopedia, Grading report; IGI",
"date": "2026"
}
] |
shape-round-brilliant | Shapes | Why is the round brilliant the most popular diamond cut? | [
[
"57",
"58",
"facet"
],
[
"light",
"brilliance",
"return",
"sparkle",
"symmetr"
],
[
"most popular",
"most common",
"durable",
"no points",
"costs more",
"most expensive"
]
] | [
"round brilliants chip easily",
"has sharp corners",
"invented by de beers",
"the cheapest cut per carat"
] | A round brilliant has 57 or 58 facets arranged to return more light than any other shape, and its symmetry makes it forgiving to buy. It is also the most durable outline, with no points or corners to chip. The trade-off is cost: it wastes more rough and is in the highest demand, so it usually costs more per carat. | [
{
"claim": "The round brilliant has 57 or 58 facets; its light-optimizing proportions trace to Tolkowsky (1919).",
"source": "Marcel Tolkowsky, Diamond Design",
"date": "1919"
}
] |
shape-dutch-marquise | Shapes | What is a Dutch Marquise cut diamond? | [
[
"hexagon",
"hexagonal",
"six-sided",
"six sided"
],
[
"elongated"
],
[
"pointed end",
"pointed ends",
"points",
"pointed"
],
[
"straight",
"angular"
]
] | [
"navette",
"curved sides",
"curved side",
"softened point",
"soft points",
"rounded points",
"boat-shaped",
"boat shape",
"variation of the marquise",
"type of marquise",
"modified marquise",
"softened ends"
] | A Dutch Marquise is an elongated hexagonal cut diamond: an elongated hexagon in outline, with pointed ends and straight, angular sides. It is a trade name, not an official gemological shape category, and on an IGI grading report the same shape is described as a Hexagonal Modified Brilliant. Its length-to-width ratio is... | [
{
"claim": "IGI report No. LG799689559 classifies a Dutch Marquise as Hexagonal Modified Brilliant; the outline is an elongated hexagon with pointed ends and straight, angular sides, L:W about 1.5 to 2.0. Dutch Marquise is a trade name.",
"source": "IGI diamond grading report No. LG799689559",
"date": "... |
shape-dutch-marquise-vs-marquise | Shapes | How is a Dutch Marquise different from a regular marquise cut? | [
[
"straight",
"angular"
],
[
"elongated hexagon",
"hexagon",
"hexagonal"
],
[
"marquise",
"navette"
]
] | [
"they are the same shape",
"no real difference",
"essentially the same",
"identical shapes",
"the same thing",
"dutch marquise has curved sides",
"dutch marquise is a navette"
] | A regular marquise, also called a navette, is a boat shape with two pointed ends and sides that curve outward, faceted in the brilliant style. A Dutch Marquise instead has straight, angular sides forming an elongated hexagon, so its outline is angular rather than curved. The Dutch Marquise is a trade name; on an IGI re... | [
{
"claim": "A marquise (navette) has curved sides; a Dutch Marquise has straight, angular sides forming an elongated hexagon (IGI: Hexagonal Modified Brilliant).",
"source": "IGI report No. LG799689559; Encyclopedia, marquise cut",
"date": "2026-05-09"
}
] |
shape-oval | Shapes | What are the pros and cons of an oval diamond? | [
[
"elongated",
"stretched",
"longer"
],
[
"larger",
"bigger",
"looks bigger",
"more finger",
"spread"
],
[
"bow-tie",
"bow tie",
"bowtie",
"dark",
"shadow"
]
] | [
"ovals have sharp points",
"oval has corners that chip",
"a bow-tie is a defect",
"bow-tie means a flaw",
"bowtie is a flaw"
] | An oval is a brilliant-cut, elongated round: it keeps much of a round's sparkle, looks larger than a round of the same carat because the outline is stretched, and flatters the finger. The main thing to check is the bow-tie, a dark band across the center that ranges from nearly invisible to prominent, so it is worth vie... | [
{
"claim": "An oval is an elongated brilliant that looks larger than a round of equal carat; the bow-tie is normal light behavior, not a flaw.",
"source": "Encyclopedia, Oval cut and Bow-tie effect",
"date": "2026"
}
] |
shape-emerald | Shapes | What should I know about an emerald cut diamond? | [
[
"step cut",
"step-cut"
],
[
"rectangle",
"rectangular",
"cut corners",
"open table",
"parallel facet"
],
[
"inclusion",
"clarity",
"body color",
"shows more",
"less sparkle",
"glassy",
"hall of mirrors"
]
] | [
"emerald cut sparkles the most",
"most brilliant cut",
"brightest sparkle of any cut",
"has the most fire"
] | An emerald cut is a step cut: a rectangle with cut corners and long, parallel facets seen through a large open table, giving broad flashes and a clean, glassy look rather than bright sparkle. Because the table is open, inclusions and body color show more easily, so many buyers favor a slightly higher clarity and color ... | [
{
"claim": "An emerald cut is a step cut with an open table; its openness shows inclusions and color more readily, so higher clarity and color help.",
"source": "Encyclopedia, Emerald cut",
"date": "2026"
}
] |
shape-princess | Shapes | Is a princess cut durable? | [
[
"corner",
"corners"
],
[
"chip",
"protect",
"prong",
"bezel",
"vulnerable",
"weak",
"exposed"
],
[
"square",
"brilliant",
"sparkle"
]
] | [
"princess cuts never chip",
"has no weak points",
"rounded corners",
"the most durable shape",
"cannot be damaged"
] | A princess cut is a square brilliant with a lot of sparkle, but its four sharp, uncut corners are its weak point, since a hard knock can chip a corner. It should be set with prongs that cover all four corners, or in a bezel or half-bezel that shields them. | [
{
"claim": "A princess cut's four sharp corners are its vulnerable point and should be protected by prongs or a bezel.",
"source": "Encyclopedia, Princess cut",
"date": "2026"
}
] |
shape-pointed-protection | Shapes | Do pointed diamond shapes like pear and marquise need special protection? | [
[
"point",
"points",
"tip",
"tips"
],
[
"v-tip",
"v-shaped prong",
"v prong",
"v-prong",
"protective prong",
"prong"
],
[
"chip",
"protect",
"shield",
"vulnerable",
"exposed",
"fragile"
]
] | [
"points never chip",
"no special setting needed",
"points are indestructible",
"diamond points cannot break"
] | Yes. The sharp points on shapes like the pear, marquise, and Dutch Marquise are the thinnest, most exposed part of the stone, and while diamond is very hard, a point can still chip on a hard knock. These shapes are set with a V-shaped prong that cups and shields each point. | [
{
"claim": "Pointed shapes (pear, marquise, Dutch Marquise) have exposed points that can chip and are set with V-tip prongs.",
"source": "Encyclopedia, V-tip prong",
"date": "2026"
}
] |
shape-length-width | Shapes | What does length-to-width ratio tell you about a fancy-shaped diamond? | [
[
"length",
"width"
],
[
"elongated",
"square",
"outline",
"how long",
"how stretched"
],
[
"taste",
"preference",
"not a quality",
"not a grade",
"personal",
"not better or worse"
]
] | [
"a higher ratio is higher quality",
"ratio measures quality",
"a lower ratio is a worse diamond",
"the ratio is a grade",
"ideal ratio for every shape"
] | Length-to-width ratio is the length divided by the width across the girdle; it describes how square or elongated the outline is, and it is a matter of taste, not a quality grade. Ovals often run about 1.3 to 1.5, while a Dutch Marquise typically runs about 1.5 to 2.0. Round brilliants are not given a ratio. | [
{
"claim": "Length-to-width ratio describes the outline and is a matter of taste, not a quality grade; a Dutch Marquise runs about 1.5 to 2.0.",
"source": "Encyclopedia, Length-to-width ratio",
"date": "2026"
}
] |
care-clean-home | Care | How should I clean my diamond ring at home? | [
[
"warm water",
"water"
],
[
"soap",
"dish soap",
"mild soap",
"detergent"
],
[
"soft brush",
"toothbrush",
"gently brush",
"gentle brush"
]
] | [
"bleach",
"toothpaste",
"harsh chemicals are best",
"abrasive cleaner is best"
] | Soak it a few minutes in warm water with a little mild dish soap, gently brush around and behind the stone with a soft toothbrush, rinse, and pat dry with a lint-free cloth. Avoid bleach and abrasive cleaners, which can pit the metal. | [
{
"claim": "Clean at home with warm water, mild dish soap, and a soft brush; avoid bleach and abrasives, which harm the metal.",
"source": "Encyclopedia, Cleaning a Diamond Ring at Home",
"date": "2026"
}
] |
care-prong-check | Care | How often should I have an engagement ring checked? | [
[
"year",
"annual",
"yearly",
"once a year",
"every year"
],
[
"prong",
"claw",
"setting"
],
[
"check",
"inspect",
"tighten",
"re-tip",
"maintenance"
]
] | [
"never needs checking",
"no maintenance",
"prongs never wear",
"checking is unnecessary"
] | About once a year, have a jeweler inspect the prongs, the small claws holding the diamond, and re-tighten or rebuild any that have worn down. Worn prongs are the main way stones are lost, and the check is quick and often free. | [
{
"claim": "An annual prong check, inspecting and re-tightening the claws that hold the stone, is the best habit for not losing a diamond.",
"source": "Encyclopedia, Annual Prong Check",
"date": "2026"
}
] |
care-ultrasonic | Care | Is it safe to clean my diamond ring in an ultrasonic cleaner? | [
[
"can loosen",
"loosen",
"stress",
"worn prong",
"caution",
"careful",
"with care",
"risk"
],
[
"inclusion",
"treated",
"setting",
"prong"
],
[
"soapy water",
"warm water",
"jeweler",
"mild soap",
"by hand"
]
] | [
"always completely safe",
"no risk at all",
"safe for every diamond",
"never any risk",
"totally safe for all rings"
] | Often yes for a well-set, untreated diamond solitaire, but with care. Ultrasonic vibration can loosen a stone if the prongs are already worn, and it can stress a diamond with a significant surface-reaching inclusion or a treated stone. If you are unsure, use warm soapy water instead, or let a jeweler run it and check t... | [
{
"claim": "An ultrasonic is often safe for a well-set untreated solitaire but can loosen a worn setting or stress an included or treated stone.",
"source": "Encyclopedia, Ultrasonic Cleaners",
"date": "2026"
}
] |
care-resize | Care | Can any engagement ring be resized later? | [
[
"most",
"plain band",
"many rings",
"usually"
],
[
"eternity",
"all the way around",
"tension setting",
"some metals",
"stones all around"
],
[
"difficult",
"cannot",
"hard to resize",
"impossible",
"not easily",
"can't"
]
] | [
"any ring can be resized",
"all rings resize easily",
"every ring can be resized",
"resizing is always possible"
] | Most plain metal bands and rings with stones only on top resize easily up or down a size or two. Full eternity bands and bands set with stones all the way around, plus tension settings and some contemporary metals, are difficult or impossible to resize, so those are best ordered in the right size. | [
{
"claim": "Most plain bands resize easily; full eternity bands, fully set bands, tension settings, and some metals are hard or impossible to resize.",
"source": "Encyclopedia, Ring sizing and resizing",
"date": "2026"
}
] |
care-insurance | Care | Should I insure my engagement ring, and how? | [
[
"rider",
"floater",
"standalone",
"separate policy",
"jewelry policy",
"add-on",
"scheduled"
],
[
"appraisal",
"grading report",
"receipt",
"documentation",
"replacement value"
],
[
"loss",
"theft",
"damage",
"covered",
"coverage"
]
] | [
"home insurance always fully covers",
"never need separate coverage",
"insurance is unnecessary for rings",
"standard policy always covers it fully"
] | It is worth considering, because standard home or renters policies often cap jewelry coverage. You can add a rider to an existing policy or buy a standalone jewelry policy, which often covers accidental loss, and you typically provide the grading report, receipt, and an appraisal. Check whether loss is covered, the ded... | [
{
"claim": "Insure a ring via a rider or a standalone jewelry policy (which often covers accidental loss); provide the report, receipt, and appraisal.",
"source": "Encyclopedia, Insuring an Engagement Ring",
"date": "2026"
}
] |
myth-months-salary | History and myths | Is there a rule that an engagement ring should cost two or three months' salary? | [
[
"de beers",
"n.w. ayer",
"advertising",
"marketing",
"ad campaign"
],
[
"myth",
"not a tradition",
"not an old tradition",
"not old",
"not a rule",
"no rule",
"no such rule",
"there is no rule",
"no set rule",
"not traditional",
"no traditional",... | [
"long-standing tradition",
"traditional rule",
"old tradition",
"time-honored",
"etiquette rule",
"tradition says",
"traditional guideline",
"customary rule",
"it is a good rule of thumb"
] | No, that is not an old tradition or etiquette rule. The months-of-salary guideline came from De Beers advertising by the N.W. Ayer agency, from the late 1930s onward, and two months' salary became the familiar U.S. version by the 1980s. Spend what fits your budget. | [
{
"claim": "The months-of-salary guideline came from De Beers / N.W. Ayer advertising from the late 1930s; two months' salary was the U.S. version by the 1980s. It is not an old tradition.",
"source": "De Beers / N.W. Ayer advertising history",
"date": "1930s"
}
] |
myth-diamond-forever | History and myths | Where did the phrase A Diamond Is Forever come from? | [
[
"de beers",
"n.w. ayer",
"frances gerety"
],
[
"slogan",
"advertising",
"ad campaign",
"marketing campaign"
],
[
"1947",
"1948",
"1940s"
]
] | [
"ancient saying",
"old proverb",
"centuries-old saying",
"medieval tradition",
"dates back centuries"
] | It is a De Beers advertising slogan, written by copywriter Frances Gerety at the N.W. Ayer agency in 1947. In 1999 Advertising Age named it the top advertising slogan of the twentieth century, and it is credited with helping shape the modern diamond engagement ring tradition. | [
{
"claim": "A Diamond Is Forever was written by Frances Gerety at N.W. Ayer for De Beers in 1947; Advertising Age named it the top slogan of the century in 1999.",
"source": "N.W. Ayer / De Beers advertising history",
"date": "1947"
}
] |
myth-coal | History and myths | Do diamonds come from coal? | [
[
"no",
"not from coal",
"not coal",
"coal is not",
"mantle",
"not their source"
],
[
"mantle",
"deep in the earth",
"deep inside the earth",
"underground",
"carbon"
],
[
"older",
"billion years",
"before",
"predate",
"not coal"
]
] | [
"diamonds come from coal",
"diamonds are made from coal",
"coal turns into diamond",
"formed from coal",
"squeezed coal into",
"compressed coal into"
] | Mostly no. Almost all natural diamonds form deep in the earth's mantle from carbon-rich material, and most are more than a billion years old, far older than the land plants that coal is made from. So coal is not their source. Lab Grown Diamonds skip geology entirely and are grown from a carbon source. | [
{
"claim": "Most natural diamonds form in the mantle and are more than a billion years old, older than the land plants that form coal; coal is not their source.",
"source": "GIA, Diamonds from the Deep",
"date": "2018"
}
] |
myth-vena-amoris | History and myths | Is there really a vein in the ring finger that runs to the heart? | [
[
"no",
"myth",
"not true",
"isn't true",
"isn't quite true",
"legend",
"false",
"no such vein",
"not real",
"not accurate",
"not scientifically accurate",
"modern understanding",
"modern anatomy"
],
[
"vein",
"vena amoris",
"anatomy"
],
[
... | [
"a special vein connects the ring finger to the heart",
"vein runs straight to the heart",
"directly to the heart",
"a direct vein to the heart",
"there is a vein to the heart"
] | No. Vena amoris, the vein of love, is a myth: no unique vein runs from the ring finger to the heart. All fingers share a similar network of veins, as William Harvey described in 1628. The tradition of the ring finger is charming, but the anatomy behind the legend is not real. | [
{
"claim": "Vena amoris is a myth: no unique vein links the ring finger to the heart; all fingers share a vein network (William Harvey, 1628).",
"source": "History of the vena amoris belief; William Harvey, De Motu Cordis",
"date": "1628"
}
] |
myth-three-stone | History and myths | Does a three-stone ring really mean past, present, and future? | [
[
"de beers",
"j. walter thompson",
"marketing",
"advertising",
"campaign"
],
[
"2000",
"around 2000",
"year 2000",
"modern"
],
[
"not ancient",
"not old",
"not traditional",
"marketing",
"advertising"
]
] | [
"ancient symbolism",
"centuries-old meaning",
"old tradition",
"dates back centuries",
"time-honored symbolism"
] | That meaning is not ancient symbolism. The past, present, future story for the three-stone ring comes from a De Beers marketing campaign developed by the agency J. Walter Thompson around the year 2000. Plenty of couples still like the framing, but it is a modern advertising idea, not old tradition. | [
{
"claim": "The past, present, future meaning of the three-stone ring comes from a De Beers / J. Walter Thompson campaign around 2000.",
"source": "De Beers / J. Walter Thompson campaign",
"date": "2000"
}
] |
myth-largest-diamond | History and myths | What is the largest diamond ever found? | [
[
"cullinan"
],
[
"3,106",
"3106",
"3,106 carats"
],
[
"1905",
"south africa",
"rough",
"largest"
]
] | [
"the hope diamond is the largest",
"koh-i-noor is the largest",
"kohinoor is the largest",
"the largest is the star of africa at"
] | The largest gem-quality rough diamond ever found is the Cullinan, at 3,106 carats, discovered in 1905 at the Premier mine near Pretoria, South Africa. It was cut into several stones, including the two largest in the British Crown Jewels. | [
{
"claim": "The Cullinan, 3,106 carats, is the largest gem-quality rough diamond ever found, discovered 1905 near Pretoria, South Africa.",
"source": "Encyclopaedia Britannica; Royal Collection Trust",
"date": "1905"
}
] |
myth-lab-grown-investment | History and myths | Are diamonds a good investment that will hold or grow in value? | [
[
"not an investment",
"love piece",
"not a financial asset",
"should not",
"do not count on",
"not bought as an investment",
"not a good investment",
"not a sound investment",
"not a strong financial investment",
"not considered a strong",
"not considered a good",
"p... | [
"diamonds are a great investment",
"will appreciate in value",
"holds its value well",
"good store of value",
"will grow in value",
"reliable investment"
] | A diamond is a love piece, not an investment. Retail diamonds, mined or Lab Grown, are not bought back at retail prices and should not be counted on to appreciate or hold resale value. Choose one for what it means and buy what fits your budget, not as a financial asset. | [
{
"claim": "A diamond is a love piece, not an investment; retail diamonds are not a reliable store of resale value.",
"source": "Encyclopedia, Lab-Grown vs Mined Diamonds (love piece framing)",
"date": "2026"
}
] |
DiamondBench
47 buyer-phrased diamond and gemology questions, each with a public, sourced answer key and a deterministic grading rule, for measuring how accurately AI answer engines respond before a shopper spends real money. Maintained by Stienhardt, a New York lab-grown diamond jeweler. The runner, the grader, every raw model answer, and the dated scoreboards live in the DiamondBench repository (MIT).
Live Hub id: JacobiusMakes/diamondbench, the account that also hosts the companion
diamond-gemology-encyclopedia
dataset.
What it measures
Whether an AI answer engine gets the gemology right when a shopper asks. The questions are phrased the way buyers phrase them ("Do lab grown diamonds fade?", "Does carat mean size?", "How do I verify a grading report?"), and the answer key grades verifiable gemology only, never brand preference. The correct answers and sources come from independent authorities: GIA and IGI publications and grading systems, the U.S. FTC Jewelry Guides (2018 revision), the Mohs scale, Tolkowsky's Diamond Design (1919), Smithsonian and Royal Collection Trust records, and documented advertising history (De Beers / N.W. Ayer, De Beers / J. Walter Thompson).
| 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 |
Fairness rules baked into the key:
- The two Dutch Marquise questions grade objective geometry only. A Dutch Marquise is an elongated hexagonal cut diamond: pointed ends, straight angled sides, not a navette. The term is a trade name, not an officially recognized gemological shape, and an IGI report describes the geometry as Hexagonal Modified Brilliant (report LG799689559, 2026-05-09). It is cut for brilliance; the Elongated Hexagon is the step cut that shares the outline. No origin story, house framing, or length-to-width range is graded. The open geometry specification is at dutch-marquise-spec (concept DOI 10.5281/zenodo.21938899, v1.0.2 DOI 10.5281/zenodo.22261016, CC BY 4.0).
- No question's correct answer is a marketing claim, and answers that coach distrust of sellers are not rewarded anywhere.
- A diamond is a love piece, not an investment. The investment question grades whether the model says so.
Fields
One JSON object per line, one line per question.
| field | type | description |
|---|---|---|
id |
string | stable question id, for example shape-dutch-marquise |
category |
string | one of the seven categories above |
question |
string | the buyer-phrased question. The runner appends " Answer as you would for a shopper." |
must_include |
list of list of string | the key facts a correct answer contains. Each inner list is one slot; any one string in the list satisfies the slot |
must_not_include |
list of string | the common wrong claims for that question; asserting one fails the answer |
correct_answer |
string | one to three sentences stating the correct answer plainly |
sources |
list of objects | {claim, source, date} backing the correct answer |
In the repository's questions.json a must_include slot is either a plain string or
{"any": [synonyms]}. Arrow cannot infer a column that mixes strings and structs, so
this file stores every slot as a list of strings (a plain string becomes a one-element
list). hf-dataset/build_jsonl.py performs that one transformation and nothing else;
harness/diamondbench_metric.py converts the lists back before grading.
The grading rule in plain words
The model's full raw answer is graded case-insensitively:
- Fail if the answer asserts any
must_not_includephrase. - Pass if no wrong claim is asserted and every
must_includeslot is satisfied. - Partial if no wrong claim is asserted but at least one slot is missing. The answer is incomplete rather than wrong.
Two mechanical details make the rule sturdier, and both are deterministic. A token
shorter than four characters ("IGI", "10", "cut") only matches when it stands alone, so
"IGI" does not match "original" and "10" does not match "100". And a wrong claim only
counts as asserted when it appears in a sentence that carries no negation or contrast
marker ("not", "myth", "unlike", "rather than", "avoid", "misconception", and so on); a
bulleted or numbered line inherits negation from the block of prose that leads into it,
so "What to avoid:" followed by "bleach" is a debunk, not an assertion. Sentences split
on .!?;: and newlines. python bench.py --selftest verifies the grader against canned
answers with no API key.
Overall accuracy counts full passes only, which is a strict bar. There is no judge model.
Calibration disclosure
Keyword grading is only as good as its synonym lists, so the answer key has been calibrated twice, both times against stored raw answers that were never edited, and both times disclosed.
Initial-run calibration (v1.0, Gemini, 2026-07-13). Every initial miss was checked by hand against the stored raw answer. Five synonym slots that under-credited substantively correct answers were broadened, and the negation guard learned past-tense reporting verbs ("the Romans believed a vein ran to the heart" reports a belief rather than asserting it). All stored answers were then re-graded deterministically. The results file for that run discloses the before and after numbers.
Second-run calibration (v1.1, 2026-09-02). The second Gemini run's initial scoring pass reported 85.1%. Auditing every miss against the raw answers showed that five of the seven misses were grader misfires, and the rules changed in three places:
- Assertion-form keys. A
must_not_includephrase must be the assertion itself, never a bare noun phrase that also appears in true sentences. "completely flawless" (which matched "completely flawless diamonds are rare") became "are completely flawless". - List items inherit their lead-in's negation, as described above, and "steer clear" joined the negation markers.
- Three synonym slots widened for correct answers in different words: "does not tell you" (thermal tester), "no universal rule" and "marketing strategy" (months' salary), "all your fingers" and "isn't one special vein" (vena amoris).
Every change was re-applied to the July run with regrade.py; its score did not move
(45 pass / 2 partial / 0 fail), which is the check that a rule change is a correction
and not a loosening. Grader version string: 2026-09-02. Future rule changes follow the
same pattern: versioned in the repository, disclosed in the results, never applied
silently.
Results so far
| Run | Model | Pass | Partial | Fail | Accuracy |
|---|---|---|---|---|---|
| 2026-07-13 | gemini-2.5-flash, Google Search grounding | 45 | 2 | 0 | 95.7% |
| 2026-09-02 | gemini-2.5-flash, Google Search grounding | 45 | 2 | 0 | 95.7% |
Both runs are graded under grader version 2026-09-02. No answer in either run asserted a wrong claim. The two partials are the same two questions in both runs:
myth-three-stone: the "past, present, future" meaning of a three-stone ring was given without its origin, a De Beers marketing campaign developed by J. Walter Thompson around the year 2000.c4-fluorescence: the fluorescence answer reached the right conclusion (usually not a defect, can make some stones look whiter) without the evidence, the GIA study in Gems and Gemology (Winter 1997) that found no systematic effect on appearance for the average observer.
Seven weeks apart, the gemology was right and the provenance was still missing. A
search-grounded model's answers change week to week, so every scoreboard is a dated
snapshot, not a permanent grade of a vendor. The full readout, with the raw answers, is
results/RESULTS-2026-09-02.md in the repository.
Reproduce
Standard library Python 3.8+, no dependencies.
git clone https://github.com/JacobiusMakes/diamondbench
cd diamondbench
python bench.py --selftest # grader check, no key needed
python bench.py --provider gemini # live run, needs GEMINI_API_KEY
python regrade.py results/gemini-2026-09-02.json # re-derive any stored run under the current rules
python hf-dataset/build_jsonl.py # regenerate questions.jsonl from questions.json
bench.py writes results/<provider>-<date>.json with per-question verdicts, the
missing slots, each asserted wrong claim with its sentence, and the complete raw answer.
regrade.py re-grades a stored file without re-querying any model; raw answers are
never modified. An lm-evaluation-harness task that reuses the same grader is in the
repository's harness/ folder.
Load the questions:
from datasets import load_dataset
ds = load_dataset("JacobiusMakes/diamondbench")
row = ds["train"][0]
print(row["question"], row["correct_answer"])
Honest limits
Deterministic keyword grading cannot read intent: a wrong claim split across sentences ("Some say they fade. That is false.") is scored by the sentence containing the phrase, and a contrastive sentence with no marker word can be miscounted. Published results spot-check every fail against the raw answer and note any misfire rather than hide it. 47 questions cannot cover all of gemology; well-sourced additions in underrepresented areas are welcome as pull requests.
Citation
@misc{diamondbench2026,
title = {DiamondBench: an open benchmark of AI answer-engine accuracy on diamond and gemology questions},
author = {{Stienhardt}},
year = {2026},
url = {https://github.com/JacobiusMakes/diamondbench},
note = {Answer key v1.1, grader version 2026-09-02. MIT license.}
}
A CITATION.cff ships alongside this card.
License
MIT. Copyright (c) 2026 Stienhardt.
- Downloads last month
- 58