Practical Data

Figures out which direction is better

If we shaded all the columns in this table with a uniform "big numbers = green, small numbers = red" rule, it would go against the meaning of a few of them (Losses, and ERA is better when it's lower, for example). The model picks the direction for each column, given its title and context.

docs/tests view it live at baseball-reference.com

Standard Pitching Table
Season	Age	Team	W	L	ERA	IP	SO	WHIP	FIP	ERA+	SO9
2020	23	DET	1	4	5.63	32.0	37	1.219	5.75	81	10.4
2021	24	DET	8	12	4.34	149.1	164	1.259	5.09	98	9.9
2022	25	DET	7	8	3.52	117.2	117	1.156	2.96	111	8.9
2023	26	DET	7	3	2.80	80.1	102	0.896	2.00	155	11.4
2024	27	DET	18	4	2.39	192.0	228	0.922	2.49	174	10.7
2025	28	DET	13	6	2.21	195.1	241	0.891	2.45	185	11.1
2026	29	2TM	12	7	2.72	158.2	186	0.964	2.51	155	10.6
Ignoreda point in time (year, season, age)	Ignoreda point in time (year, season, age)	Ignoredlabels with no natural order	Amounthigher is better · log scale	Amountlower is better · linear	Amountlower is better · linear	Amounthigher is better · linear	Amounthigher is better · linear	Amountlower is better · linear	Amountlower is better · linear	Amounthigher is better · linear	Amounthigher is better · linear

Knows which numbers are amounts

A lot of numbers aren't amounts; if you try to shade everything numeric, identifying things like Jersey # or Zip Code would look like they're on a scale. Practical Data asks the decision model whether a column is really a measure first, so identifiers stay are ignored.

docs/tests synthetic test page

Roster
Player	Jersey #	Born	Zip code	Height (in)	Points per game
Ada Park	23	1998	94110	74	21.4
Ben Ortiz	7	2001	10027	79	12.9
Cy Moreau	11	1995	60614	81	8.2
Di Okafor	34	1999	30307	77	17.5
Eli Sato	0	2003	98103	72	5.1
Ignoredlabels with no natural order	Ignoredan identifier, not an amount	Ignoreda point in time (year, season, age)	Ignoredan identifier, not an amount	Amountno better direction · linear	Amounthigher is better · linear

Knows when to use a log or linear scale

On a linear scale, the one $237 card here would flatten everything under $30 into nearly the same red colour. The model notices when ratios matter more than differences and moves that column to a log scale, so the cheap cards still read as different from each other.

docs/tests view it live at mtggoldfish.com

Price History
Card	Set	Rarity	Price	Weekly (%)
Mox Opal #223	MM2	Mythic	237.09	-4.00%
Sheoldred, the Apocalypse #107	DMU	Mythic	105.01	0.00%
Mox Amber #224	DOM	Mythic	92.22	-2.00%
Gemstone Caverns #274	PLIST	Rare	63.15	-1.00%
Minamo, School at Water's Edge #118 Foil	FMB1	Rare	57.87	0.00%
Cavern of Souls #269	LCI	Mythic	51.74	0.00%
Boseiju, Who Endures #266	NEO	Rare	51.40	0.00%
Amulet of Vigor #121	WWK	Rare	50.99	0.00%
Claws of Gix #107	TSB	Special	12.52	+5.00%
Untimely Malfunction #161	DSK	Uncommon	9.74	+6.00%
Sink into Stupor #241	MH3	Uncommon	8.82	+3.00%
Relic of Progenitus #213	MMA	Uncommon	7.48	-3.00%
Faerie Macabre #86	MMA	Common	6.96	0.00%
Pyretic Ritual #153	M11	Common	6.50	+7.00%
Desperate Ritual #163	CHK	Common	6.46	0.00%
Witch Enchanter #239	MH3	Uncommon	5.99	+1.00%
Lorien Revealed #60	LTR	Common	4.59	-7.00%
Ignorednames or free text	Ignorednames or free text	Ordered labelscommon → uncommon → rare → special → mythic	Amountlower is better · log scale	Amounthigher is better · linear

Puts labels in order

Works on words, not just numbers! The model can put labels in order when the presence of that order is implied (like tropical storms, or card rarity, or medal rankings).

docs/tests view it live at en.wikipedia.org

Costliest Atlantic hurricanes
Name	Nominal damage (Billions USD)	Season	Storm classification at peak intensity
Katrina	$125.0	2005	Category 5 hurricane
Harvey	$125.0	2017	Category 4 hurricane
Ian	$112.0	2022	Category 5 hurricane
Maria	$91.6	2017	Category 5 hurricane
Helene	$78.7	2024	Category 4 hurricane
Irma	$77.2	2017	Category 5 hurricane
Ida	$75.3	2021	Category 4 hurricane
Sandy	$68.7	2012	Category 3 hurricane
Ike	$38.0	2008	Category 4 hurricane
Milton	$34.6	2024	Category 5 hurricane
Andrew	$27.3	1992	Category 5 hurricane
Wilma	$26.5	2005	Category 5 hurricane
Ivan	$26.1	2004	Category 5 hurricane
Michael	$25.5	2018	Category 5 hurricane
Florence	$24.2	2018	Category 4 hurricane
Laura	$23.3	2020	Category 4 hurricane
Ignorednames or free text	Amountlower is better · linear	Ignoreda point in time (year, season, age)	Ordered labelscategory 3 hurricane → category 4 hurricane → category 5 hurricane

Ignores Total and Average rows

The first row here is the World total; considered alongside the others, it would dwarf all their numbers and make them almost indistinguishable. The model flags totals and averages as aggregates and scales around them.

docs/tests view it live at en.wikipedia.org

List of countries and territories by total population
Location	Population	% of world	Date
World	8,232,000,000	100%	13 Jun 2025
India	1,429,404,000	17.3%	1 Jul 2026
China	1,404,890,000	17.0%	31 Dec 2025
United States	341,784,857	4.1%	1 Jul 2025
Indonesia	288,315,089	3.5%	31 Dec 2025
Pakistan	241,499,431	2.9%	1 Mar 2023
Nigeria	223,800,000	2.7%	1 Jul 2023
Brazil	214,211,951	2.6%	1 Jul 2026
Bangladesh	169,828,911	2.1%	14 Jun 2022
Russia	146,028,325	1.8%	1 Jan 2025
Mexico	131,231,919	1.6%	30 Jun 2026
Japan	122,650,000	1.5%	1 Sep 2026
Ignorednames	Amountno better direction · log scale	Amountno better direction · log scale	Datesearliest to latest

Rows judged to be totals and kept out of the scale: "World".

Multiple tables can share a scale

Practical data looks for same-named columns in similar contexts and aligns their scales accordingly so they're more easily compared at a glance.

docs/tests view it live at docs.github.com

OpenAI
Model	Category	Input	Cached input	Output
GPT-5 mini	Lightweight	$0.25	$0.025	$2.00
GPT-5.3-Codex	Powerful	$1.75	$0.175	$14.00
GPT-5.4	Versatile	$2.50	$0.25	$15.00
GPT-5.4	Versatile	$5.00	$0.50	$22.50
GPT-5.4 mini	Lightweight	$0.75	$0.075	$4.50
GPT-5.4 nano	Lightweight	$0.20	$0.02	$1.25
GPT-5.5	Powerful	$5.00	$0.50	$30.00
GPT-5.5	Powerful	$10.00	$1.00	$45.00
GPT-5.6 Luna	Lightweight	$0.20	$0.02	$1.20
GPT-5.6 Luna	Lightweight	$0.40	$0.04	$1.80
Ignorednames or versions	Ordered labelslightweight → versatile → powerful	Amountlower is better · log scale · shared scale	Amountlower is better · log scale · shared scale	Amountlower is better · log scale · shared scale

Anthropic
Model	Category	Input	Cached input	Output
Claude Haiku 4.5	Versatile	$1.00	$0.10	$5.00
Claude Sonnet 4	Versatile	$3.00	$0.30	$15.00
Claude Sonnet 4.6	Versatile	$3.00	$0.30	$15.00
Claude Opus 4.8	Powerful	$5.00	$0.50	$25.00
Claude Opus 5	Powerful	$5.00	$0.50	$25.00
Claude Opus 5.5	Powerful	$4.00	$0.20	$20.00
Claude Sonnet 5	Versatile	$2.00	$0.20	$10.00
Claude Sonnet 5.5	Versatile	$2.00	$0.20	$10.00
Ignorednames or versions	Ordered labelslightweight → versatile → powerful	Amountlower is better · log scale · shared scale	Amountlower is better · log scale · shared scale	Amountlower is better · log scale · shared scale

Notices when the table is meant to be read row-wise

Not every table compares down the column. Here each row is one benchmark and the columns are models, so column-wise shading would compare numbers that have nothing to do with each other. The model reads the orientation and gives each row its own scale.

docs/tests view it live at blog.cloudflare.com

Benchmark scores by model
Benchmark	Clef	Clef-flash	Jev	DiffusionGemma Jev	Kev 9B	Laya
BFCL · case exact	98.47	98.76	95.75	96.52	94.51	38.13
ToolRet · nDCG@10	69.19	66.43	65.28	61.21	64.26	12.69
API-Bank · accuracy	91.93	93.11	88.19	83.66	56.30	11.41
Home appliances · case exact	82.95	97.73	52.27	42.05	25.00	0.00
When2Call · accuracy	72.37	65.58	80.97	75.44	49.62	11.94
BANKING77 · macro-F1	94.20	90.93	79.74	74.28	84.83	14.29
CLINC150+OOS · macro-F1	97.43	66.77	89.27	83.49	79.03	3.19
BRIGHT · nDCG@10	45.91	39.26	47.52	42.94	38.53	19.90
Amazon ESCI · macro-F1	57.48	57.39	55.21	53.37	49.22	24.40
PhishNChips · accuracy	79.60	75.05	62.55	85.35	50.75	50.15
Ignorednames or versions	Amounthigher is better · linear	Amounthigher is better · linear	Amounthigher is better · linear	Amounthigher is better · linear	Amounthigher is better · linear	Amounthigher is better · linear

Understands times and durations

It would be easy to treat 2:42 as text, or a ratio. Practical Data parses it as a duration, and understands minutes/hours correctly (3:00 is one more than 2:59).

docs/tests view it live at baseball-reference.com

Time of Game For Season
Year	G	AvgTime	%≥3:30	Avg Tm Btwn PA	Avg Tm Btwn BIP
2026	2218	2:42	0.5	2:10	3:15
2025	2216	2:38	0.1	2:08	3:10
2024	2209	2:36	0.3	2:06	3:09
2023	2226	2:39	0.4	2:08	3:13
2022	2204	3:03	10.5	2:29	3:42
2021	2089	3:10	18.7	2:33	3:52
2020	720	3:07	15.6	2:31	3:51
2019	2211	3:05	13.5	2:28	3:43
2018	2210	3:00	8.9	2:25	3:37
2017	2246	3:05	12.2	2:28	3:39
2016	2236	3:00	8.3	2:25	3:32
2015	2213	2:56	7.1	2:23	3:25
2014	2194	3:02	11.1	2:28	3:33
2013	2180	2:58	8.3	2:25	3:27
2012	2231	2:55	5.8	2:22	3:24
2011	2185	2:51	4.5	2:19	3:16
Ignoreda point in time (year, season, age)	Amountno better direction · linear	Amountlower is better · linear	Amountlower is better · log scale	Amountlower is better · linear	Amountlower is better · linear

How it works (TLDR)

1. Read the page

When you scroll a table into view, the extension figures out if it's got data or not. If it does, the extension standardizes its values (for parsing; your view of it doesn't change): Things like $0.025, .312, 6' 6", 1.5B, $50–60K, 2:59, ≤ 272K, etc.
2. Reason about the data

A System One model (Jev) answers narrow questions with probabilities:
- Is this column an amount? an identifier? a timestamp? a rank?
- Is a higher value better, worse, or neither?
- Does it make sense to show it in linear or log scale?
- Is this row a total, or part of the data?
- Do same-named columns in other tables share a scale?
- Are values compared down each column, or across each row?
3. Paint the table

The extension uses D3 to map each value onto a colour scale and applies it over the page's own background.

If you bring your own Jev or Clef key, the extension sends the data to the provider directly. If not, it sends it to an intermediary Cloudflare worker with a (daily limited, but these things are pretty cheap!) Jev API key. I don't retain anything either way (Privacy policy)
Nothing is sent on sites you turn off. The toolbar popup turns a page or a whole site off, and ⌥⇧H pauses the current page.
A typical page costs about a tenth of a cent on your own key (or mine).

Open Source (MIT)

You can Bring Your Own Key (BYOK) for Jev or Clef, or use the extension without a key (with a usage-limited free endpoint). System one models are pretty cheap :)

$ git clone https://github.com/philrenaud/practical-data
$ cd practical-data && npm install
$ npx playwright install chromium
$ npm run check

View source on GitHub See the core data tests Report a misread table A bit about how the questions are designed

Install

Install Practical Data from the Chrome Web Store.
It works without a key. To use your own, paste a TypeSafe or Clef API key on the options page.

Examples

How it works (TLDR)

1. Read the page

2. Reason about the data

3. Paint the table

Open Source (MIT)

Install