Build LogContent Pipeline
#16
2026-08-036 min read

Verification without correction isn't a gap in the pillar cron — it's the design

The 30 June mortgage draft had 10 of 13 sentences clear the sentence-level verifier against live gov.uk pages. The three that failed dropped it from auto-publish. That is not a limitation — it is the design.

Read post
Build LogContent Pipeline
#17
2026-08-016 min read

Five articles shipped in July. Every one of them needed a hand.

July's cadence log records five published articles. Behind four of them is a distinct failure mode — composer crash, checklist casualty, CLI error, day-cap, or coverage-screening gap — that the operator patched by hand. The July 17 exception shows what the pipeline looks like when everything holds; the rest of the month shows what it looks like when the operator is the error-correction layer.

Read post
Content PipelineBuild Log
#18
2026-08-015 min read

Thirteen sentences. Ten checked out. Three would have shipped wrong.

A mortgage rate-spike draft cleared ten of thirteen sentence-level checks and failed three. Without a corrective verifier, the three failing claims would have shipped alongside the ten correct ones. The 10/13 split is where LLM content pipelines silently break facts.

Read post
Build LogUK Calculators
#19
2026-07-305 min read

The cron fires daily. July's 14-day cadence clock ran on manual rescues.

Four composer failures manually rescued. A three-weekday gap that broke the clock and required two publishes on one morning to restart it. The July UK Calculators logs show cadence discipline isn't a property of the scheduler — it's a property of what happens when the scheduler fails.

Read post
Build LogContent Pipeline
#20
2026-07-305 min read

The day-three mortgage draft cleared verification. The cluster had already shipped on day two.

On day three of a fourteen-day content run, the cron produced a mortgage rate-spike draft that cleared sentence-level verification. The same cluster had shipped the morning before. Sentence verification and cluster novelty are different problems — one belongs at the verification layer, the other doesn't.

Read post