What changed

The headline said 100x. The overcharge was 47x. The post also presented a smoke-test cost check and model aliases as shipped. Both were plans, and the smoke test we shipped only checks that a cost header is present.

The overcharge was 47x, not 100x, and the model string was in 46 files, not 56. Only the features marked as shipped are live, not every feature mentioned.

The section on the cost bug now says a 47x overcharge, not 100x. We removed a count of issues found in code review that we could not back up.

EU AI Act high-risk obligations now start in December 2027 or August 2028, not August 2026. Two survey figures, 88% and 14.4%, are restated to match their source. The sample request headers are corrected. Cost caps sit at the org, team and key level and model allowlists at the org level, not per agent.

The history was wrong. MCP and AGENTS.md were founding projects of the Agentic AI Foundation in December 2025, A2A joined in August 2026, and IBM's protocol merged into A2A in August 2025.

The 4 PRs from the first session were all closed unmerged the same day, which the post didn't say. We removed a comparison with an engineer's time. Night Owl runs at 8:15 AM UTC, not 8 AM.

The title said the pipeline went from issue to production deploy in 3 minutes. It went from issue to PR. Since April 2026 Claude Code in containers needs a different token variable, and the post now says so.

The 51-provider and OpenClaw figures are now dated to March 2026; dead, stub and cloud providers were dropped in May 2026. Integration takes one environment variable and one header, not two variables. The semantic cache threshold defaults to 0.90, not 0.92. We removed a claim that OpenClaw has no approval workflows, and our $0.50 HITL default.

One fleet has five specialists, not six. Deployment isn't one click; it runs on managed runners, in private beta as of November 2026.

PII scanning is off by default for household orgs, and the post had listed it. The word "family" has one exception in the API, an onboarding route.

The plan had 25 verified gaps, not 26, and the week had 43 production PRs, not deploys. The closeout counted 40 of 41 nightly runs, not nights. The quality gap stayed between 0.60 and 0.80, not +0.8 every night.

When the daily org budget runs out, new requests are denied. They aren't queued until the next day.

The public sample-request trace is illustrative, not a real or anonymized gate sequence.

Google Fonts has no Inter Display family. The post had listed one.

The AI cost on each post comes from that post's frontmatter. It isn't recorded live by the gateway, as the post had said.

Now says this was our fashion analysis pipeline, removed in May 2026, rather than a pipeline we still run.

The Live Build demo has since been retired for Runner Spotlight, and the task counts are dated to March 31, 2026.

Family Manager is now called Nod.

Now says it describes the fleet as it ran in March 2026.

Now says it describes the fleet as it ran in March 2026.

The three tool profiles come from our container images, not from OpenClaw itself. The post now says it describes the fleet as it ran in March 2026.

The 99.4% figure comes from Vorlon's 2026 CISO Report and covers SaaS or AI incidents. We had credited a LayerX survey. EU AI Act high-risk obligations now start in December 2027, not August 2026, with penalties up to 3% of turnover, not 7%. The 135,000 instances were found exposed in one January scan, not running right now.

The billing bug was 47x, not 100x. The pipeline ran on three repos under two GitHub owners, not three organizations. A new note says the step names, template names and engine list don't all match what shipped.

The palette values now match the site, and Claude Design is described as an Anthropic Labs product rather than a handoff tool.

The alerting work on April 14 came five days later, not a week later.

Now says it describes the fleet as it ran in March 2026.

Now says it describes the fleet as it ran in March 2026.

Updated the scorecard with the November 1 re-measure.

Added the 1.14 upgrade tests, tests on images and Box, the audit of how we wired Jev in, and the follow-up experiments.

Marked the reports we fixed as shipped.

Added the full-edition live result.