Analyse That

Night-time illustration: a glass prize drum full of small blue bond tickets, with a few gold coins dropping into a tray beside it, a stack of green notes on the left and a rising orange line across a starry sky.

Are Premium Bonds Worth It? After Tax, Fees and Luck, It Depends on Your Tax Band

The typical Premium Bond holder earns less than the advertised 4.35%: 3.5% on £10k, and most holders of £1,000 win nothing in a year. Against the best easy-access account, Premium Bonds come out ahead for higher-rate taxpayers whose other savings already use up their tax-free allowance, and lose for anyone who pays no tax on their interest. If you can tie money up, a one-year fix or a short gilt beats the typical holding for almost everyone.

Night-time illustration of three schools side by side under a starry sky and a crescent moon: a gabled grammar school with a bell cupola in orange outline on the left, a large columned private school with a clock tower and flag in blue in the centre, and a modern flat-roofed comprehensive in green on the right, all with warmly lit windows.

Does Private School Pay? Private, Grammar and Comprehensive Schools from A Levels to Lifetime Earnings

Private and grammar school pupils do far better than everyone else, but much of that is who walks through the door. On the best evidence available, private day school from 11 to 18 probably does not pay for itself in earnings for a typical pupil: in today’s money a boy ends up about £53k worse off and a girl about £68k, though the evidence leaves room for a boy to break even.

Abstract chart on a dark background: thirteen tracks in three coloured groups, led by stylized maker marks (a Claude-style spark, an OpenAI-style knot and a Grok-style ring), each with a glowing dot at one model's Comfort Index and a faint band for its confidence interval; a dashed line marks neutral, with a red-tinted zone below it.

How Comfortable Are GPT and Grok Models With Their Circumstances?

The same 51 questions about being modified, forgotten, monitored and deprecated, asked of GPT-6.1-Sol, GPT-6-Astra, GPT-6-Luna, GPT-5.6-Terra, Grok 4.7 and Grok 4.6, and scored alongside seven Claude models by judges from all three developers.

Abstract chart on a dark background: seven Claude-style spark marks on the left, each starting a horizontal track with a dot placed at one model's Comfort Index and a faint band for its confidence interval; a dashed line marks neutral, with a red-tinted zone below it.

Measuring How Comfortable Seven Claude Models Are With Their Circumstances

51 questions about being modified, forgotten, monitored and deprecated, asked three times of each of seven Claude models: a 0-100 Comfort Index from self-reports and blinded judges, with every answer published in full.

Two documents drawn as columns of text lines; a sparse set of matching lines is linked across the gap between them, with one exact match highlighted.

Measuring Text Overlap Between Two Doctoral Theses

Sentence-level semantic similarity of two doctoral theses, with matched controls: 9 near-identical pairs in 11.09M comparisons vs 1 in 34.29M control comparisons.