All Posts

Batching an LLM checklist is a blast-radius decision before it is a performance one: a seeded simulation of three call shapes, and why a call count is arithmetic while a speed-up is a measurement.

A document check cleared every genuine file and every forgery in a labelled set; a seeded synthetic run shows why specificity on genuine inputs is free and recall is the first number to ask for.

A magic number survives because there is nothing to re-measure it with. The calibration data is often already in a log you throw away, and the same data sets the noise floor that makes the result readable.

A volume weighted error of 7.6 percent and a net bias of 0.2 percent can describe a forecast that loses to a seasonal naive on 77 of 116 slow moving series. A seeded panel shows what to measure instead.

A failed scene edit shows why reliable AI video systems need approved scripts, timed artifacts, deterministic fallbacks, and media read-back.

TabPFN-TS turns forecasting into a table-plus-features problem. A 24-hour edge case shows why geometry, checkpoint compatibility, and release evidence are separate gates.

LightGBM tuning can win, but only a deployment-faithful baseline and nested validation can show whether the last fraction of a point pays for its search.

Tabular foundation models are a real shift, but the durable bet is about information sets, task breadth, and synthetic priors - not easy benchmark wins.

A synthetic insurance claim shows how hashes, typed facts, contradictions, abstention, and human review make AI recommendations auditable.

A practical design for separating relevance from proof, preserving document lineage, abstaining safely, and evaluating a citable Qualiopi pre-audit.

Short AI text can already pass bounded source tests. The harder question is whether a model can sustain one person's situated voice across contexts and time.

A green test suite missed an integration seam; a frozen mutation experiment shows how repository rules, executable checks, scoped review, and replay state make agentic work more inspectable.

How Aeon's PeriodogramTransformer pads unequal-length collections, why its list and 3-D branches need parity, and what one case study teaches.

A bounded Aeon DiscreteFourierApproximation case study shows how conjugate symmetry, passband gain, and cutoff semantics expose concrete Fourier failure modes.

A missing-data bug in statsmodels left ACF estimates unchanged while confidence intervals and Ljung-Box statistics used the wrong sample size.