QWERTY was arranged for English prose. So we measured what it costs to type something else: one key unit is 19.05 mm, fingers travel from wherever they last landed (no return to home), and a shifted character is two presses — the second made by the opposite hand’s pinky.

Prose costs 14,686 mm of finger travel per 1,000 characters. Code costs 18,612 — 1.267×. Almost all of the difference lands on one finger: the right pinky goes from 12.4% of all travel to 30.8%, because it owns the bracket, the brace, the quote, the colon and the semicolon.

And the second most-pressed key in code is not a letter. It is Left Shift at 8.35%, ahead of e at 6.61%; only the space bar beats it. Shift presses go from 3.0% of all presses in prose to 11.9% in code.

⚠️ There is no 13-language leaderboard, and that is the finding. Language matters globally (η² = 0.598, permutation p = 0.00005) but only 1 of 12 adjacent pairs separates. Within-language spread across five repos is 2,684 mm against a between-language range of 5,740 mm — swap one repo for another and you move further than most neighbouring languages sit apart. Go alone is nameable, at 22,245 mm, separable from all twelve others (joint permutation p = 0.0010).

⚠️ The caveat that matters: this is per 1,000 CHARACTERS, not per unit of work. Code says more per character than prose, so it is the cost of typing it and not of writing it. Two earlier claims died in the checking and are documented rather than hidden: the cheapest code source (fastapi, 13,494 mm) does not beat the worst prose source (Alice, 15,208 mm), and Go’s lead is not driven by its capitalised identifiers — it is only 3rd in Left-Shift presses.

Nothing upstream is redistributed: every row is a count and a distance we computed, and each of the 65 repositories carries its permissive licence in the file.

Every “best city for data scientists” list is a sorted column of salaries, and a sorted column of salaries mostly tells you where housing is expensive. The fix is usually a regional price index — which is published on a different schedule and a different geography to the wages, so bridging the two adds more error than it removes.

You can skip it. Divide a metro’s data-science median by that same metro’s all-jobs median and both numbers are already in the same local prices: the price level cancels. No index, no vintage mismatch.

Which disqualifies the leader. San Jose pays $173,160, first in America — against a local median of $82,470. That is 2.10×, and 50th of 135. A 49-place fall with no change to the wage, only to the question.

Answer: Charlotte–Concord–Gastonia, $131,110 against a local median of $48,880 — 2.68×. A banking centre in a normal Southern metro: the numerator is a finance salary, the denominator is not a finance cost of living. First on the two-year mean (2.73×) and first in each year alone (2.79× in 2023), on 3,870 data scientists at a 1.5% relative standard error.

The median metro pays 2.02× — almost exactly double — and only three of the 135 clear 2.5×. Year-over-year rank stability is Spearman ρ = 0.750.

⚠️ It is a ratio of medians, not of the same people’s pay. It says what the job is worth against the local wage base; it does not say what your raise would be for moving. 17 metros are excluded for clearing the reporting threshold in only one of the two years — a rule fixed before the answers were looked at, which matters because Idaho Falls would lead on its single year. They ship in a separate file so you can disagree and re-run it.

Every version of this argument counts flags or counts yards, and both are averages over a wildly asymmetric distribution. Priced in win probability instead: the median flag moves 1.84 points, the 99th percentile 14.16, the worst single flag in five seasons 66.54 — the worst 1% of flags carry 8.3% of everything penalties move. On yards the most-wronged club is Chicago (−10.04 yd/game); on win probability it is the Jets (−0.72 wins a season). The two correlate at r = 0.773.

Then the split that matters. 6,854 mechanical flags a camera settles · 9,449 judgment flags somebody decided — 58% of all penalties, and the only half that can be wrong in a team’s favour. Across the 32 clubs the two columns correlate at r = +0.015: how disciplined a team is tells you nothing about how the judgment calls go for it. Which disqualifies the leader — the Jets are 31st of 32 at committing the flags a camera settles. They are not being robbed, they are undisciplined.

Answer: Tennessee, 9th of 32 on the facts and last of 32 on the opinions — −3.31 win probability points a game, −0.56 wins a season, worst under all five codings of the split. Pass interference alone is 48% of the deficit. The referees themselves: crews show no home-team bias (p = 0.41) but do differ in severity (p = 0.019) — which is symmetric, so it is variance, not robbery.

Shipped with the number that argues against it: z = −2.27, p = 0.015 for Tennessee alone — and p = 0.43 across all 32 clubs. Somebody has to finish last. The claim is “the best case in the league”, never proof.

⚠️ Nobody publishes which calls were RIGHT. The league grades its officials privately, so this measures how the judgment calls fell, never whether any one of them was correct. A club can finish last here because it is mis-officiated, because it plays a style that draws judgment flags, or through five seasons of bad luck — the permutation test only addresses the third.

Every “best bank” list is written by somebody paid per signup, and the weights are never published. So: six dimensions a federal regulator already publishes about every bank in the country, all six weighted equally, applied identically. Interest paid · complaints per $10B · redress · reach · branch density · CET1 capital. The six survived a correlation cut over fourteen candidates — the largest absolute pairwise correlation left is 0.444, so nothing is a proxy for anything else.

The four largest branch networks finish 41st (JPMorgan Chase), 48th (Bank of America), 60th (PNC) and 71st (Wells Fargo) of 83 — and none of them is bad, they are spiky. Chase is the 100th percentile on reach and the 9th on branch density; Bank of America is the 99th on redress; Wells Fargo sits in the 1st percentile for complaints. A weighted sum of capped percentiles punishes a hole far more than it rewards a peak.

Winner: WaFd Bank of Seattle, 69.0, +2.24σ, 1.42 clear of Busey — and first at nothing. It is above the median on five of six, and its worst measure is the 48th percentile when no other bank’s worst clears the 44th.

Shipped with the part that argues against it: 35 of the 83 win under some defensible weighting. WaFd is #1 in 13.7% of 100,000 weightings drawn uniformly from the simplex against 10.6% for BMO — the highest share in the field, and it still loses seven times in eight. Weight size first and First Interstate wins; weight service first and WaFd drops to 9th. Both rankings ship in the file.

⚠️ 83 banks is not every bank in America — it is the largest US retail networks that also have a federal complaint record. 158 banks clear 50 branches; the 75 without a CFPB match were dropped, not carried at zero, because a zero there is a 100th percentile and a fabricated winner. Credit unions and small community banks are not in here.

Which NFL team is best is an argument about rings. Which team is best to be a fan of is a different question, because you are asking what the franchise gives back to the people who show up. Six dimensions, each one measurable: success (win %), payoff (playoff wins per season), drought, cost (Fan Cost Index), loyalty (crowd vs the club’s own record crowd) and now (2021–25 form). All six count equally — a declared value judgement, stated up front and then tested rather than defended.

The famous teams are not bad, they are spiky. New England is 100th percentile at winning and 100th at payoff, and 16th at cost and 16th at loyalty — it finishes 15th of 32. Dallas has the fullest building in football relative to its own peak and zero conference finals in twenty-seven seasons; they finish 20th. Pittsburgh, 97th percentile at winning, finishes 19th.

Winner: Baltimore at 73.1, +1.60σ, two clear of Seattle — and best at nothing. They are the only club above the league median on all six, and their worst dimension is the 55th percentile when no other club’s worst clears the 39th. Two Super Bowls, not six.

Shipped with the part that argues against it: 19 of the 32 clubs finish first under some defensible weighting. Baltimore is #1 in 32.7% of 100,000 weightings drawn uniformly from the simplex against 25.1% for Indianapolis — the most robust answer in football, and it still loses two weightings in three. Weight cost first and the Colts win; weight the Sundays first and Arizona does, with Baltimore sixth. Both rankings ship in the file.

Every “best watch” list is a taste with a price tag next to it. So: four characteristics, each chosen because you can point at it on a photograph of the watch, weights written down first, applied identically to every watch an authorised dealer will actually sell you. Water resistance 30% · case durability 25% · movement 25% · complications 20%. Ranked by midrank percentile within all 1,605 — water resistance takes only a handful of distinct values, so a positional percentile would let file order decide the winner.

Price is deliberately not an input, because the question is what the price tag knows about the watch. The answer: across the whole market, ρ(price, score) = −0.156. Sorted into ten equal price groups, mean score rises to 57.7 around $9,550 and falls to 36.1 in the top decile (median $89,000). The most expensive watch in the set, a $1,700,000 Greubel Forsey Grande Sonnerie, ranks 1,333rd of 1,605 on 30 metres of water resistance. A $440 Casio G-Shock ranks 51st, above a $16,950 Rolex Submariner (455th) and a $76,645 rose-gold Patek Nautilus (1,138th).

Winner: Panerai Submersible Carbotech at $19,900, 87.1, 2.41σ clear — and #3 is a $3,350 Longines, 1.1 points behind.

Shipped with the limitation stated rather than buried: the score does not measure chiming complications, hand-finishing, provenance or resale, which is why the grande sonnerie scores zero on complications. And the number that cuts against the headline — below about $50,000 the price–score relationship is flat (ρ = +0.023), so the claim is not “cheap is better”, it is that price stops predicting specifications and then starts predicting them backwards.

Every “worst airline” list is a vibe with a logo next to it. So: three metrics, equally weighted, fixed before the data was pulled, applied identically to all nine US mainline carriers over June 2025 – May 2026. On-time arrivals (within 15 min, over flights that operated) · cancellations (over flights scheduled) · mean arrival delay (minutes per operated flight). Z-scored within the nine, not percentile-ranked — with only nine units a percentile would claim the gap between 8th and 9th matches the gap between 1st and 2nd, and here those differ by an order of magnitude.

The answer is Spirit, at −1.09 — and not because it flies late. Spirit is 7th of 9 on on-time, ahead of both Frontier and JetBlue. It finishes last on one metric: it cancelled 3.41% of everything it scheduled against a field average of 1.74%, which is 2.0σ below the mean and the largest single-metric deviation in the table. The margin over 8th is 0.225σ — narrow, and said out loud rather than dressed up.

The turn: Frontier has the worst on-time rate in America and is not the worst airline. A cancelled flight never gets to be late — it leaves the on-time statistic entirely. Judge airlines on the number they all advertise and you systematically under-punish the ones that solve delays by cancelling.

Shipped with the robustness check that matters: the delay metric has a defensible alternative definition, so the whole ranking was recomputed under it. Spirit is last either way (margin widens to 0.265σ) — but Frontier/JetBlue and Delta/United swap, which is why mid-table positions are noise. Also shipped: why price, legroom and safety are deliberately absent, and why the regionals you can’t book a ticket on are excluded.

Every “best places to live” list is an opinion with a number bolted on, and the weights are never shown — which is what makes them impossible to disagree with. So: four metrics, weights fixed in advance, applied identically to the 95 largest US metros (223.5M people, two thirds of the country). Cost of living 30% (BEA price parities) · housing burden 25% (ACS, owners AND renters) · job growth 25% (BLS) · commute 20% (ACS). Percentile-ranked, not z-scored, because San Francisco’s price level would compress everyone else.

The winner is Fayetteville–Springdale–Rogers, Arkansas at 96.4/100, 4.6 points clear and 2.25σ above the mean — and not because it is the best at anything. It never falls below 8th of 95 on any of the four, and no other metro manages that. A weighted sum of capped percentiles punishes a hole harder than it rewards a peak: you can only earn 30 points on cost, but you can lose all 30.

Tested against something deliberately left out of the score — Census net domestic migration over the same 95 metros. ρ = 0.50, and the gaps are the finding: Lakeland FL is 1st for arrivals and 55th on the ruler; Salt Lake City is 16th on the ruler and 85th for arrivals.

Shipped with what is missing and why: violent crime is absent because the FBI publishes no metro-level endpoint, and four measured metrics beat five with one invented. Also shipped: the three bugs that produced completely normal-looking charts — a tie rule that made the ranking depend on file order, a sign error that crowned Miami, and a name match that returned real federal numbers for Las Vegas, New Mexico.

Every settled Polymarket “what will Trump say” market with a real traded price, 2024-06 to 2026-08. These resolve mechanically — a word was said or it wasn’t — so nobody is adjudicating intent, which is what makes thousands of them usable as a test of whether a price is any good.

It is. Bucket by price and compare to what actually happened and the market lands on the diagonal: mean absolute gap 2.04¢ across ten bands. That is the trap, not the good news. If the price is already the true probability, expected profit is exactly zero on both sides — p(1−p) − (1−p)p — and everything after that is subtraction.

Both venues charge a fee shaped like price × (1−price), a parabola peaking at 50¢, which is where 60.5% of all volume trades and where the market is least accurate (the 45–55¢ band happened just 43.3% of the time). Zero minus a guaranteed fee is negative every single trade: after 10,000 bets you are ahead 0.090% of the time, about 1 in 1,100.

Shipped with what didn’t work: no edge survives a 2¢ spread, the favourite–longshot bias is absent here, a “+21%” strategy was leverage not skill, and our own Poisson model lost to the market 30 splits out of 30. The honest claim is narrow: the taker loses by construction — the maker collects.

The US Senate Budget Committee's county-level non-renewal file (23 insurers, 249 million policy-years, 2018–2023) joined to the FEMA National Risk Index and every federal disaster declaration.

In 2018 the correlation between a county's modelled disaster risk and how often its insurers refused to renew was 0.02 — effectively nothing. Same counties, same risk scores, 2023: 0.48. The riskiest tenth of America went from being dropped 1.2× as often as the safest tenth to 2.4×, and outside Florida, Louisiana, California and Texas the 2018 ratio was 0.98 — the dangerous places were dropped slightly less. Ships with the methodology, including the two county-name join traps that produced wrong numbers first.

Every US adult in NHANES 1999–2016 joined to the NCHS death records through 2019. The column that matters is drinking_status, because it separates the two groups that forty years of studies quietly merged: people who never drank, and people who quit.

Pool the quitters in with the lifetime abstainers, the way the classic studies did, and light drinkers come out 31% less likely to die. Split them out and change nothing else: 19%. 44% of the famous benefit was never about alcohol at all — and the quitters had a 4.90% two-year death rate against the light drinkers' 1.13%. Ships with the full methodology and the six models that failed to kill the remainder.

A bank weighed the identical Chipotle bowl 75 times across 8 NYC stores — it ranged from 13.8 oz to 26.8 oz for the same price, so your cost-per-ounce nearly doubles at random and the store you pick is the biggest factor. Ships with a full methodology writeup (individual weights are modeled to the study's published summary — read it). Source: Wells Fargo audit via Fortune, 2024.

How fast a 911 ambulance reaches you, mapped against neighborhood income — the poorest ZIPs wait ~4 min longer and routinely blow past the 8-minute standard (survival drops ~7–10%/min). Ships with a full methodology writeup (the cloud is modeled on the real national gradient — read it). Source: Hsia et al., JAMA 2018.

Vanguard, BlackRock & State Street combined stakes in major US companies — they're the #1 shareholder in ~88% of the S&P 500 (21.9% median stake, 24.9% of votes). Source: Bebchuk & Hirst + 2025 proxy/13F filings.

Every player in NHL history to reach 1,000 career points. Source: Wikipedia.