Latest

A seven-stage skill pipeline for Claude Code that takes a Jira ticket in a codebase you don't know yet and gets it to a pushed branch with small, reviewable commits: intake, decide, spec, slice, build, review, ship, with four human gates and one untracked notes folder per ticket. It is assembled from what Matt Pocock, Addy Osmani, GitHub Spec Kit, obra/superpowers, Dex Horthy and Anthropic actually ship and teach, then stress-tested by four research agents — controlled productivity studies (METR, Stanford, DORA, Ambig-SWE), critics of spec-driven workflows, ten more YouTube transcripts and a hands-on technical audit that ran every git command in a throwaway repo. The verdict: the core loop holds and being asked clarifying questions has the strongest evidence, but it fails if you run the full process on every ticket, so tickets are right-sized into XS/S/M/L tracks. Slices become one commit each on the branch, the platform squash-merges the PR into one commit per ticket on main, and the full kit — eight SKILL.md files, two bundled curl scripts, two hooks, settings and a setup script — is included, ready to copy into ~/.claude/.

Jev, the first model from TypeSafe AI, does not generate text. You give it a state and a set of typed questions, and it returns a probability distribution over the answers you defined, in a single parallel pass that costs you nothing for output. Your code then branches on the result, so Jev works as a smart if-statement rather than a chatbot. TypeSafe trains it with what it calls RLCD, Reinforcement Learning for Calibrated Decisions: a new training target, in the way RLHF was, whose goal is probabilities that mean what they say rather than text a rater likes or an answer a verifier accepts. This post explains what Jev is and how to use it, then goes through RLCD in pictures. It covers what calibration means, why RLHF's reward destroys it and why RLVR's reward cannot produce it, and why a proper scoring rule is the only objective that fits the stated goal. It also shows why that makes the 'reinforcement' in the name mostly a story about where the data comes from, and why the one number where Jev still leads its open clones is confidence ordering, which temperature scaling cannot fix. Nothing about RLCD has been published, so every claim is labelled by who made it.

A looped transformer applies the same block of layers to a token several times over, so depth stops being a property of the weight file and becomes an argument you pass at inference. A rumour about OpenAI's Astra turned that idea into a fight about redlines, and almost nobody arguing had read the architecture. So this is the architecture, mechanically: the three-part forward pass and the two details every summary drops, why the initial state is random and the core is blind to which iteration it is on, how the test-time depth dial is manufactured during training by sampling r from a heavy-tailed distribution and truncating the backward pass to the last eight iterations, and what the loop actually costs — because it trades storage for time and creates no compute at all. The exchange rate has been measured: one recurrence is worth r^0.46 in equivalent parameters, real and comfortably sublinear. The genuinely strange result underneath is that looping moves reasoning without moving knowledge, confirmed independently by three groups using different methods, and the frontier of the idea is not a bigger loop count but a learned per-token router. Verdict: a real third scaling axis and a component rather than a paradigm — and on the monitorability argument, both camps were right about different questions.

On real market data there is no answer key, so you can never tell whether an agent discovered real structure or overfit noise. superquant-bench inverts the trade-off: a fully synthetic 100-asset price panel with 22 known alpha patterns injected into it, giving the grader the exact true conditional mean of every return. That buys three things reality cannot — a prediction score with a literal zero noise floor (oracle 100, iid noise 0), a directly measured false-discovery rate via a pattern-free twin universe, and statistical look-ahead enforcement. A red-team of seven exploits validates the metrics, then four frontier models run under one identical Claude Code scaffold. The spread is 10×, model identity explains 75% of prediction variance — and the strongest model is the least disciplined, spraying false discoveries on the universe that contains nothing.

Crypto leaderboards rank wallets by raw return, but return over a handful of trades is almost all noise. This post builds a three-level hierarchical Bayesian model — population, trader, trade — that learns each wallet's risk-adjusted skill as a posterior Sharpe ratio with honest uncertainty, shrinking thin-data traders toward 'no edge' so a lucky five-trade wallet is not mistaken for a skilled one. It covers the data (7.6M Hyperliquid fills reconstructed into per-position returns), partial pooling and shrinkage, the conjugate streaming updates, the full PyMC fit by NUTS, and a single continuous walk-forward showing the learned skill predicts the next trade and drives a top-10 copy account to 77% winning copies — before an honest autopsy shows most of that edge was concentration in one trending asset.