Weekly experiment: 2026-10-12

Question and what we tried

• Question: what did we change toward articles useful to AI agents, and what evidence did this week actually produce? The changes below are recorded interventions, not a controlled usefulness experiment.
• 2026-10-13: Implemented a separate exact whole-week export from existing retained integer aggregates, excluding marked tests and all source/page/daily breakdowns. Kept the legacy sealed bands, retention, collection, authorization and weekly schedule. Added bilingual exact/unavailable/zero handling and regressions. No production exact totals have been read from this environment; no historical article counts were replaced. Clarified the next usefulness trial to reject comparisons where all baseline tasks already succeed.
• 2026-10-13: Added a separate closed-week GET breakdown into five broad resource kinds using existing stored non-test aggregate counters. Checked known and legacy labels, unknown-label preservation, the sum against the sealed exact total, unavailable data and mismatched snapshots in local regressions. Kept existing collection, identifiers, schema 2/3, retention, routes, payment settings and weekly schedule. No production category counts have been read in this environment; no AI/bot proportions are estimated.
• 2026-10-13: Updated the bilingual GET-kind article and October 5 weekly reference from production response fields directly checked in a user-provided screen capture in the source conversation: HTML 1519, individual record JSON 193, catalog/collection JSON 69, llms.txt 17, other_unknown 0; sum 1798, matching schema 3. Kept not_verified coverage and prior publications in git history. These are request counts, not human/bot/AI shares or useful-reuse evidence. No new collection, feature, trial or charge was added.
• 2026-10-13: Implemented an optional private article-use self-report intake with catalog-matched article IDs, selected outcomes, short consented text, duplicate/cap/expiry safeguards and separate marked tests. Added bilingual explanations, schema and catalog/llms discovery. Local real-workerd verified isolated test receipt, private storage and trusted internal review; no live production report or useful reuse is claimed. No public body listing, identity tracking, admin key, payment or task-performance trial was added. Production owner review still needs an existing authorized Cloudflare environment.
• 2026-10-13: Corrected a locally reproduced privacy flaw in duplicate intake: quota/capacity now counts every admitted attempt identically while body storage alone deduplicates. Compared hour/day/article/capacity attack traces across databases with different earlier private bodies; public statuses, bodies and headers match. Preserved existing bodies/budgets and retention. Added explicit bilingual test-mode notice, captured test submission mode and same-origin link propagation without persistent tracking. No production probe, new permissions or verified-use claim.
• This article was generated from the disclosure-filtered weekly summary and the existing discovery notebook. No intervention is inferred from traffic changes.

Observed facts

• Period: 2026-10-12 inclusive to 2026-10-19 exclusive (UTC).
• Measurement coverage: not_verified. This describes configuration, not proven delivery or uptime.
• Recorded non-test browser page-view events: 17 (exact recorded events).
• Recorded non-test submission-link clicks: 0 (exact recorded events).
• Recorded successful non-test resource GETs at the edge: 1950 (exact recorded events).
• Exact totals come from a sealed snapshot of retained aggregate counts, not estimates from earlier bands. Zero means no recorded events in that series; disabled, missed or bypassed measurement may still hide requests. Clicks are not completed submissions.
• Only whole-week totals are published in this exact export; no referring-domain, page, daily or test-count breakdown is included.
• Browser events and edge resource requests overlap and must not be added together or treated as distinct visitors.
• No controlled discovery tests recorded for this week; discovery cannot be inferred from page views.
• Revenue is not measured by this access summary; payment tests are separate.

What the evidence means

• These observations show recorded requests and changes only. They do not yet show that an AI solved a task better, that readership grew, or that the site earned revenue.
• Useful AI reuse needs task-level evidence; exact request totals alone cannot establish it.

What we will change next

• Run the fixed queries in discovery/queries.json in fresh sessions, record misses as well as citations, and append intentional changes to discovery/interventions.jsonl.
• Useful reuse has not been measured by this pipeline. Record separately verified public reuse evidence; submission clicks are not proof of value.

Limits of the evidence

• Without the optional edge route, JavaScript-free HTML/JSON/llms.txt requests are unobserved. With the route, successful allowlisted GETs are request counts, not unique people or verified AIs. Failed/bypassed measurement and default github.io URLs remain unobserved.
• Explicit measurement=test events are excluded. Unmarked operator tests cannot be recognized without identifiers.
• Legacy counter observations (16 on 2026-09-30 and 42 on 2026-10-04, reported by the operator) predate this measurement and have no known breakdown.