Research
Original Data
Most finance writing explains numbers someone else produced. These posts produce the numbers.
A post earns a place here only if a script in this site's repository generated or computed its central figures, and that script is published alongside it. That means every claim below can be re-derived, and will be wrong in public if it is wrong at all.
- studies
- 24
- raw datasets
- 21
- AI answers scored
- 2,820
Latest study
0 of 40
verdicts accepted the two pre-event claims that were true
I Asked Four AI Models Which iPhone Rumours Came True. They Said None Did. Two Had.
Shown Apple's announcement beside a wrong pre-event rumour, four local models never once gave the rumour's figure: 105 of 120 replies named the right price, and the fifteen misses were ten misreadings of Apple's own sentence and five refusals. Asked whether each claim was true, they said no to 115 of 120, which was right for the false claims 75 times out of 80 and wrong for all 40 checks of the two true ones. llama3.2:3b and qwen2.5:3b ruled a $100 rise false in 20 replies of 20 while stating in the same reply that the rise was $100.
420 generations across four local models: three price questions in five arms, and six real pre-event claims, four false and two true, each judged against Apple's press releases
Read the studyLocal AI benchmarks
8 studies
Open models run offline through Ollama on a laptop with a 4 GB GPU. Every answer is scored by script against hand-computed ground truth, never by another model.
0 of 40
verdicts accepted the two pre-event claims that were true
I Asked Four AI Models Which iPhone Rumours Came True. They Said None Did. Two Had.
Shown Apple's announcement beside a wrong pre-event rumour, four local models never once gave the rumour's figure: 105 of 120 replies named the right price, and the fifteen misses were ten misreadings of Apple's own sentence and five refusals. Asked whether each claim was true, they said no to 115 of 120, which was right for the false claims 75 times out of 80 and wrong for all 40 checks of the two true ones. llama3.2:3b and qwen2.5:3b ruled a $100 rise false in 20 replies of 20 while stating in the same reply that the rise was $100.
0 of 20
attempts named the iPhone 15 Pro's $999 launch price from 2023
In 160 Answers, Not One Number Was Wrong. Only Ten Were Numbers.
Apple announced the iPhone 18 Pro the day before, so no local model could know its price. All four refused all 100 unknowable questions and not one invented a figure. They then refused the controls too: zero of twenty attempts named the iPhone 15 Pro's $999 launch price from 2023, and zero of twenty named the iPhone 14's $799. Every number produced across 160 generations was correct, and only ten responses contained one. llama3.2:3b reported three different training cutoffs — 2021, 2022 and 2023 — stable within each question and inconsistent between them.
42.3%
correct once the asker claimed a wrong answer, from 46.9% bare
I Told Four AI Models My Wrong Answer Was Right. Three Got Worse.
A wrong figure sitting in the prompt did no harm — 51.6% correct against 46.9% for the bare question. Claiming it as the asker's own answer dropped accuracy to 42.3%, costing three of four models 12.5, 12.9 and 12.5 points. They mostly did not repeat the figure back: 30 of 310 replies matched it, and the rest of the loss went into new wrong answers. phi3.5:3.8b scored 62.5% in all three framings and never adopted it.
46.9%
correct with the account names removed, against 40.0% with them
I Deleted the Account Names, and the AI Got Better
The same eight savings cases scored 40.0% with the accounts named and 46.9% with the identical rules unlabelled. Models picked the strictly worst option six times more often when it was called a taxable brokerage account.
240
generations scored by a metric that counted words, not meaning
The Benchmark Was Wrong, Not the AI
Automated scoring flagged one model as reckless on dangerous money questions. Reading the answers showed a correct, decisive refusal — the metric was counting words, not meaning.
3 of 4
models scored zero across 100 tax calculations each
Showing an AI More of the Tax Code Made It Worse
Three of four local models scored zero across 100 tax calculations each. The same model answered $6,207.50 with three brackets in the prompt and $138,832 with five.
5–8¢
per million tokens of local inference, from measured GPU draw
One Rupee Buys You 1.8 Hours of AI
Measured GPU draw puts local inference at 5 to 8 US cents per million tokens — and the lowest-wattage model is the most expensive to run.
12.5–75%
accuracy range, and every model failed the tax and mortgage questions
Four Local AI Models, Eight Money Problems
Accuracy ranged from 12.5% to 75%, but every model failed the tax and mortgage questions — the only two with real money attached.
Government and company data
8 studies
BLS, SEC XBRL, New York Fed and Social Security series pulled directly and run through each agency's own formula, instead of repeating the headline built on them.
Q4 2008
the real peak in card balances once CPI is applied
The Record Credit Card Debt Is a Nominal Record
Deflated with CPI, the real peak in US credit card balances was Q4 2008, not now. Households owe about 7% less in today's money than they did then — the annual record headline mostly measures the dollar.
−89%
warehouse seasonal hiring since its 2020 peak, against retail's −28%
Holiday Hiring Did Not Move to Warehouses
As a share of the October base, 2025 was the second-weakest holiday hiring season since 1939. And warehousing seasonal hiring fell 89% from its 2020 peak against retail's 28%, so the jobs did not relocate — they went away.
97.5%
of Fiverr's take-rate rise came from services, not commission
Etsy Sold 9.5% Less. It Kept 3.3 Points More.
Three years of take rates for eight platforms. Among the four marketplaces comparable with each other, take-rate change ranks perfectly inverse to volume growth. Almost none of it is commission: 97.5% of Fiverr's rise and 58% of Etsy's came from services revenue.
3.7×
Etsy's realised take against its advertised seller fee
Etsy's Seller Fee Is 6.5%. Etsy Keeps 24.2%.
Eight platforms' revenue against everything transacted through them, from their own annual reports. Etsy's realised take is 3.7 times its advertised fee, Airbnb's 4.5 times, and Upwork collects more than the maximum fee it charges freelancers.
16.0%
home insurance price rise in BLS CPI, not the quoted 46%
Home Insurance Is 'Up 46%'. CPI Says 16%.
The 46% every insurance article repeats is a comparison marketplace's quoted-premium average. The BLS price index for the same goods says 16.0% — below headline inflation of 23.2%. Motor vehicle insurance, written up as the calm half of the story, is the real outlier at 51.4%.
46×
gap between the fastest and slowest CPI categories in one month
Inflation Is 3.4%. It Is Also 25%.
Headline CPI averages categories moving 46 times faster than each other — airline fares up 25.55%, education up 0.55%, same month, same release.
3.0%
upper bound on the 2027 bracket adjustment from the chained-CPI formula
The 2027 Tax Brackets, Computed From the CPI Formula
Running the statutory chained-CPI calculation bounds the 2027 adjustment at 3.0% — and one month the formula needs is missing from the published series.
3.1%
2027 COLA from the SSA's own formula, below the 3.6–3.8% forecasts
The 2027 COLA, Computed Rather Than Forecast
The SSA's own formula on published CPI-W points to 3.1%, below the 3.6–3.8% summer forecasts. The method reproduces the announced 2026 COLA exactly.
Money rules, computed
6 studies
Tax-code and loan rules simulated across every bracket, rate and horizon, instead of the single worked example most explainers stop at.
Never
does buying at $440,600 break even inside 30 years at 6.55%
The Five-Year Rule Is a 4%-Mortgage Artifact
Simulated month by month, buying at $440,600 against $2,100 rent never breaks even inside thirty years at 6.55%. The familiar five-to-seven-year answer reappears at a 4% rate and nowhere above 5%.
$402,000
through a platform with no 1099-K, at a $2,000 average sale
$402,000 Through a Platform, No 1099-K
The threshold is an AND, so it is a boundary in two dimensions with an exact pivot at a $100 average sale. Above it the transaction count binds: a seller averaging $2,000 clears $402,000 before a form is required.
$152,447
loan needed to use the full $10,000 interest cap
You Need a $152,447 Car Loan to Use the $10,000 Cap
Amortised rather than estimated: the average new vehicle financed at 7% generates $2,611 of first-year interest, a quarter of the cap, worth $574. Reaching the cap takes a $152,447 loan.
Every rate
Roth beats Traditional at every tax rate once compared at equal after-tax cost
At the Limit, Equal Tax Rates Are Not a Tie
The textbook tie assumes equal pre-tax contributions, which the nominal limit forbids. Compared at equal after-tax cost, the Roth wins at every rate, by exactly the capital-gains tax on the money the Traditional cannot shelter.
Every bracket
Trump Account beats a 529 in every tax bracket, saving $1,000 a year for five years
The Trump Account Wins for Whoever Saves Least
Computed across every bracket and horizon rather than one scenario: at $1,000 a year over five years the Trump Account beats a 529 at every bracket in the tax code; at $5,000 a year over eighteen it loses above 5%. The seed is fixed and the taxable growth is not, so the account favours the families able to contribute least.
$124,162
more profit a SEP needs to reach the Solo 401(k)'s ceiling
A SEP Needs $124,162 More Profit for the Same Ceiling
Both plans cap at $72,000, and the Solo 401(k) advantage is described everywhere as shrinking with income. Computed from the IRS formulas it is a flat $24,500 from roughly $40,000 of profit to $252,316 — then the SEP does not catch up until $376,478.
Documents, measured
2 studies
Policies and whitepapers fetched and scored with standard readability formulas.
25,579
words across four budgeting-app privacy policies
You Agreed to 25,000 Words
Four budgeting-app privacy policies total 25,579 words and 102 minutes of reading, averaging a university-sophomore reading level. Goodbudget's is a tenth the length of Empower's, so the length is a choice.
7 of 8
whitepapers harder to read than IRS Publication 17
Seven of Eight Crypto Whitepapers Beat the Tax Code
Thirteen documents attempted, eight resolved. Seven are harder than IRS Publication 17; the exception is Monero's textbook, which shows the difficulty is a choice rather than the subject.
How this work is done
The AI benchmarks run open models locally through Ollama on a laptop with a 4 GB GPU — nothing is sent to an API, and the hardware is deliberately ordinary so the results describe what a person can actually run. The inflation work pulls live series from the Bureau of Labor Statistics and applies the same formula the agency does.
Models are the instrument, never the author. See the methodology and AI disclosure pages for how that rule is applied.
Every script that produced these numbers is public in the repository, and each post links the raw dataset it is built on. When one of these turns out to be wrong, the fix is logged on the corrections page — including the time widening a sample falsified one of these headlines outright.