Archives

An extraction confidence is not an adverse-action reason. Document Intelligence returns two scores, and the gate reads the wrong one.

An HCM grants permissions to people holding jobs, and an agent is neither. Least privilege with no native expression, and the primitive that just shipped.

Four vendor agent platforms in one mesh, each sound alone. The seven joins that failed, starting with three catalogs where the loosest one decides.

The judge is a model too, with the same length, position and self-preference biases. Measuring it, versioning it, and knowing when not to use one.

Containment counts the sessions that never needed you. A booking agent's rollback is a refund, and a surge multiple is a scalar over a rotating mixture.

One roofline number sets the speculative-decoding gate, the batch-size knee and the chunked-prefill floor. Why max-num-seqs is a latency control.

Validating a system that never returns the same answer twice: what qualification becomes when the output is distributional, and where sampling breaks.

A conformal coverage guarantee is marginal, not conditional, so it can hold overall and fail on exactly the subgroup your fairness audit is about.

An 8B model on a $250 Android phone hits a bandwidth ceiling, a residency ceiling and a thermal ceiling, in that order, and dies mid-sentence.

Tail sampling fixes head sampling and adds three quieter ways to lose the same traces, none visible in the config and none logged as an error.

The content screens you put inline are themselves services behind the same wall. The endpoint Model Armor needs, and the field that says a screen ran.

Who runs the OAuth flow decides your network topology. Token exchange across agent hops, and why an act claim records who acted but never when.

Leaving a fleet of LLM agents a client's own team can run: discovery by directory, scorecards a stranger can read, runbooks drilled against real failures.

On a versioned corpus, faithfulness metrics pass an answer grounded in a revision retired months ago. Treating the corpus as bitemporal is what fixes it.

The seam between what a model may see and what it may run, enforced by different systems owned by different teams, and the error message as a leak channel.