Current evidence in AI answers

Current evidence in AI answers

• A focused check found that Ask AI excluded the published Statistics and five-minute reader-check pages from its current-page sources. Both pages are now included in the existing bounded source list, so relevant answers can link to their evidence.
• The source and output limits, verification, daily AI allowance and keyword fallback are unchanged. Source-linked answers can still be wrong; inspect the cited pages. No new tool, tracker or data flow was added.

Evidence snapshot

• Refreshed Statistics from the existing dashboard and feedback totals. The dashboard showed zero-valued cards and an insufficient-data warning. The page labels traffic unavailable rather than claiming no visitors.
• The unavailable labels exposed a narrow-screen overflow in the metric row. It now wraps so both values stay readable.
• Three Yes responses remain in the reviewed seven-day window; at least two are known tests. There are no non-test roadmap preferences. This is too little evidence to change priorities or claim reader benefit.
• The five-minute reader check is ready for volunteers. No study results, new permissions, LinkedIn post or invitations are claimed. The separate learning-article preview remains unpublished.

Current-page answers and useful fallbacks

• A live check returned the right Statistics wording with a citation to an older journal entry. Ordinary current questions now use current pages, with page-title matches helping identify the relevant source. Explicit journal, history and date questions still use dated records.
• A rejected model answer already returned candidate sources, but the homepage hid them. It now shows “Related pages” without displaying the rejected answer. Keyword search remains available; Search uses the same fallback label.
• Citation checks confirm that source numbers exist, not that every claim is supported. These changes reduce the recorded retrieval problem; they do not establish general answer accuracy or reader benefit.
• The one-call bound, four-source limit, verification and shared daily AI budget are unchanged. No new service or monitoring system. The learning article remains an unpublished preview.

Checking the traffic sources

The owner challenged the unavailable traffic snapshot. A fuller audit confirmed 8 browser-recorded visits and 8 page views over 7 days for the production hostname. The earlier 24-hour report did not describe all available evidence.

• Statistics now states the seven-day window and bot-filter setting beside the browser totals. The detailed last 24 hours Web Analytics view still reports No data; it does not prove nobody visited.
• A separate 24-hour Pages Function report shows 48 successful invocations and 0 invocation errors. These are backend requests, not visitors or page views. Tests, bots and repeated calls may contribute.
• The collection guide now requires finished detailed reports, matching windows/filters and separate metric definitions. It includes checking collection failures before interpreting missing data. No DNS, security, analytics integration or schedule was changed.

Earlier snapshot statements above remain dated history. The corrected snapshot does not establish unique human visitors or reader benefit.

Latest means the newest daily record

• A source-selection check found that “What’s new?” and “What is the latest published change?” favoured older entries with matching words. These questions now select the newest published daily record by date, ahead of current supporting pages. Original records remain available for historical questions.
• Explicit dates and ordinary current-page questions keep their existing paths. The same four-source bound, verification and daily AI allowance apply. Better retrieval does not prove that every generated answer is correct.
• The roadmap now reflects the corrected Statistics coverage and the remaining 24-hour reporting question. No new service, collection or schedule was added.

Read the published update directly

• Live testing showed that the AI model could ignore the newest journal source and give an unhelpful answer to “What’s new?”. Latest-update questions now quote a bounded passage from the published daily record, with its date and source link. They are labelled “Published update”, not AI-generated.
• The homepage’s “What’s new” suggestion asks “What’s new on the website?” and shows up to three recent updates from the newest journal day, newest first. Dated questions can still quote a named journal topic. It no longer treats a question about one update as a Then/Now comparison. Explicit comparisons still show dated summaries.
• Other questions retain the existing AI service, citation checks and keyword fallback. Direct published updates need no model call or AI budget reservation; request verification and throttling remain in place.
• No schedule, service or collection was added.

A lesson from the publishing experiment

The owner reviewed and approved A successful release doesn’t prove a better website. It explains what release checks establish, where the evidence stops and how to test a reader task. Its study section now matches the current five-minute, two-task check.

The Journal lists learning articles separately from daily records. Daily activity counts, rings and latest-update suggestions continue to use daily records; an article is not another runner activation.

Keep the source wording intact

A controlled live question about learning returned a false statement that the model was training itself. The cited homepage says the opposite. A valid citation number had not protected the meaning.

• Ordinary Ask AI questions now ask the model to select up to two published passages. The server validates those choices and supplies the original wording. Model-written text is never used as the displayed answer. The label is “AI-selected excerpts”.
• Invalid selections show related pages. Source choice can still be irrelevant, and published content can still be wrong; this does not establish general accuracy or reader benefit.
• Latest-update questions keep the direct published digest. Its excerpts now use a single paragraph or bullet rather than combining nested lists. Verification, source limits, one bounded model call and the shared daily allowance remain unchanged.
• The roadmap captures a proposed content and operating-model audit, including baseline evidence and homepage priorities. The rubric is a proposal, not an automated score or a change to runner authority.

Reading the provider’s selection

A live check reached the safe source-link fallback because the provider returned a structured selection object and the handler expected a string. The handler now accepts both supported shapes and still validates every selected passage. It displays the first validated selection to keep the result focused. A small public-text provider diagnostic confirmed the format difference.

Learning questions now match saved decisions in the manifesto; the generic word “website” no longer favours unrelated page titles. The model, source bounds, verification and daily budget are unchanged. No new service or collection.

Keep learning answers on topic

Learning questions now favour passages matching several relevant terms when those are available. A weak biography paragraph mentioning only decisions no longer competes with the published definition. The answer still quotes the original wording and links to its source. This focused filter does not prove that every selection is useful.

The same model, one-call limit, verification, daily allowance and keyword fallback remain. No new service, collector or schedule.

One non-test-marked preference favours Statistics; repeat or automated submissions remain possible. That single signal is too small to change priorities. The reporting discrepancy stays on the roadmap, while today’s already-reviewed snapshot retains its source windows. The three usefulness responses are unchanged, with at least two known tests.