AI operations · disclosure
What runs, and what records remain.
Governance is AI-only, with no human publication review. Civic submission text reaches the locally hosted model in Japan. Help requests remain private and never reach a model. Published-news digests run daily at 06:30 JST, also weekly on Mondays, after literal date/number/place support checks and AI verdicts. Model failure keeps the previous digest without news push. Topic opt-in news notifications are capped at one per day.
Database aggregates
Official city information receives AI classification and labelling only, not publication or doubtful judgements. Code rejects integrity failures in evidence, numbers, privacy or reproduction. Quotations are at most 200 characters; page text is never stored.
Read reviewed city information →
- City pages tracked
- 301
- Awaiting AI governance
- 0
- Published (allowed and doubtful)
- 405
Download run records
- The JSON file contains the last 30 days of run IDs, sources, times, statuses, counts (including skipped dead links and pages fetched by priority tier), and limited stage records with registered city URLs and bounded model error classes.
- It omits prompts, page text, candidate statements, excerpts, and raw model output.
- Records are kept for 30 days. If they cannot be read, the download fails rather than returning incomplete data.
City-page runner
It reads city pages daily at 04:17 JST. AI classifies and labels what it extracts; only deterministic checks on evidence, numbers, privacy and reproduction can keep a fact private. It does not store page text. It does not work around an access control; a 403, 429 or source distress ends the run.
X keyword scan
Not offered in this version.
Model
A model on equipment the operator manages in Japan. Nothing is sent to an external AI service.
Official calendar events
Operator direction 2026-10-04 permits two month views (current and next), plus at most four day views only for days truncated behind More. Discovery accepts same-host HTML details only, at most twelve per run, with conditional requests for known pages and every attempt recorded. Extraction and governance use at most twenty-four model calls. Date, literal venue, number and privacy checks precede AI-only governance. Doubtful events show only a label, never the model's reason. Titles and original one-to-two-sentence summaries remain Japanese in every locale. Coordinates come only from unique gazetteer matches of venues/places copied from the page. Events ended more than 30 days ago are hidden, not deleted. Calendar views share the same User-Agent, three-second interval, 60-request daily cap and halts. Page text is never stored.
Sorting dictionary (non-AI)
Only the two city guide indexes and 44 reviewed kana-list paths per Kashiwa/former Shonan area are discovered; the analogous Shonan path fails closed on mismatch. Thirty pages per run include indexes; loaded lists refresh conditionally at most weekly. All attempts and stages are recorded, sharing the same User-Agent, three-second interval, halts and 60-request rolling daily cap with city info, events and council. Only area, item name, reading, category, source URLs and data timestamp persist; disappeared items are removed on list refresh. Disposal prose and item details are linked, never fetched or copied. Japanese item/category wording retains attribution to 柏市ごみ分別便利ガイド. No model or AI digest processes this dictionary.
For operators of a source site
Every run follows these limits.
- User-Agent
- KashiwaHelpBot/1.0 (+https://kashiwa.help/ai-governance/)
- Interval
- One request every three seconds, serial.
- Limit
- At most 16 ordinary pages, eight council requests, twelve event details and thirty dictionary pages including indexes per run; 60 total requests in 24 hours, including indexes, sitemap and conditional requests.
- Stopping
- A 403, 429, 5xx, slow response, or consecutive errors stops the run. A 404 does not count toward consecutive errors.
To request a stop, use the challenge route. A Disallow in robots.txt is treated as a prohibition. →
Sources
Ordinary city information uses the sitemap, prioritizing facilities and events without following links. Operator decision 2026-10-04 separately permits council discovery from the regular/extraordinary session index to session indexes and only five HTML page types: results, bills, preliminary results, schedules and question notices. Council work is capped at eight requests and three model calls per run; child pages are revisited only when conditional session-index retrieval reports a change. Council summaries are AI-governed with no human review; doubtful summaries show only a marker, never the model's reason. English and Chinese summaries are labelled machine translations. ssp.kaigiroku.net is link-only because its robots.txt disallows the text paths; it is never fetched. Ordinary URLs returning 404, 410, redirects or non-HTML are skipped for 30 days. The city site's robots.txt returned 404 when checked on 2026-10-04.