Privacy

On this page 10 sections

This is a personal engineering blog. Here's what's collected and why.

Page analytics
Request information and a daily pseudonymous identifier help explain access to the publication. The routine counter uses no cookies. Its observation rows currently have no automatic expiration. Full collection and retention details.
Private diagnostics
Optional request diagnostics expire from private reads after 60 days; scheduled cleanup can lag. Global Privacy Control or Do Not Track omits that optional layer, while the routine counter remains separate. Diagnostic policy.
Newsletter
Your email supports signup and delivery. Confirmation reservations last seven days; unsubscribed and bounced addresses are purged after 90 days. Every newsletter has an unsubscribe link. Delivery, retention and opt-out details.
Search & services
Interactive search runs locally. A plain search form sends its query in this site's URL. Cloudflare provides infrastructure, analytics and abuse protection; Google Fonts supplies fonts. Search privacy · Other services.

First-party analytics

From classifier version 2026-10-27.2, routine observations also retain closed derived classification evidence: all recognized client names and reviewed product aliases, their category agreement, a coarse browser-family claim, missing/listed/not-listed network evidence, navigation shape, the activation marker's presence/form, and a bounded referrer-context inconsistency flag. Unrecognized free-text products and raw User-Agent values are not copied into this record. These facts remain linked to the routine pseudonymous observation under its existing retention; they do not authenticate a device, person, purpose or vendor.

From October 27, 2026 UTC, routine observations also retain the local signature-verification outcome/profile and derived coverage facts: which request components were authenticated, the signature's age and validity duration, its wire format, and the expiry-only replay policy. From October 28, 2026 UTC, successful verification also records the first-valid selection policy and counts of matched, tagged, attempted and unassessed signature candidates; these counts do not establish how many signers are valid. A separate assessment compares a verified key publisher with reviewed first-party directory origins; its review date, missing mappings and conflicts remain explicit. It is distinct from a documented promise to sign each client's requests. Any available Cloudflare verified-bot or signed-agent booleans and reviewed provider-category values are recorded separately; unavailable fields remain unknown. Category evidence distinguishes empty, unavailable and unrecognized inputs without retaining arbitrary source text. Raw signature bytes, key material, nonce values and the signed header values are not stored in this evidence record. Public reports group events by reviewed verifier-version identifiers; missing or inconsistent profiles remain explicit. Report generation does not reverify historical signatures. These facts remain linked to the routine pseudonymous observation and share its retention; they do not establish a person's identity, authorization or reading.

For a successful, non-prefetch GET that serves an HTML page or negotiated Markdown, this site records the page path, referring hostname, country, a daily pseudonymous client ID, a heuristic traffic class (browser, bot, or AI User-Agent), the matched User-Agent rule name when available, device type, the request's network (autonomous system number and organization name, such as an internet provider or cloud host), the values of the Fetch Metadata headers Sec-Fetch-Mode, Sec-Fetch-Dest, Sec-Fetch-Site, and Sec-Fetch-User when a browser sends them, whether the Accept header included HTML, and whether an Accept-Language header was present (the Accept and Accept-Language values themselves are not stored), whether the request is mine, and the UTC observation time. It does not record query strings, full referrer URLs, city, continent, raw IP addresses, or raw User-Agent strings in the main observation table. This request evidence has been collected since September 3, 2026.

From collection version 2026-10-27.5, routine client-form evidence records exact mobile/tablet User-Agent markers, a coarse browser-family claim, the already-sent Sec-CH-UA-Mobile preference as mobile/non-mobile/absent/unsupported, and differing or conflicting indicators. No extra hints, device model or browser hardware probes are requested. The old three-value device field remains a deprecated compatibility heuristic with an explicit default flag in new evidence; current reports use the new claims and label earlier heuristics separately. Neither non-mobile preference nor missing markers authenticates desktop hardware. These derived facts remain linked to the observation and share its retention.

For recognized client declarations with a configured public IP feed, the collector also compares the edge-supplied address with the vendor’s published ranges. The address stays in memory and is not sent to that vendor or added to this record. Stored facts include the method/profile, address availability/source/family, applicable-source identifiers, listed/nonlisted/unavailable/not-assessed outcomes, and the public feed’s digest, version, prefix count and fetch/validation/freshness times. New records also distinguish fetching, revalidation, memory reuse, shared fetches, cooldown and failure, with closed reasons for blocked reuse; publisher headers are not copied into observations. These facts remain linked to the observation and are not anonymous. We do not perform reverse-DNS lookups or treat range membership as proof of a person, intent or readership. Unrecognized declarations do not trigger a feed lookup.

The edge daily client ID is the first 128 bits of an HMAC-SHA-256 value derived from the site host, UTC date, IP address, and User-Agent using a secret key. The IP address and User-Agent are processed to create it. Raw IP addresses are not retained by this analytics model; User-Agent evidence, when retained, follows the separate diagnostic policy below. The ID can link requests within one UTC day; it is not anonymous, does not identify a person, and cannot establish the same client across dates, devices, or networks. Public stats exclude rows marked as mine; owner marking depends on server-side configuration and may not identify every request I make. Recording is best effort and uses no cookies.

New page observations also distinguish external, internal, absent, and unusable referrers. For an internal referrer, only a recognized public blog page path is retained, without query strings or fragments. This supports aggregate reports of movement between pages. Arbitrary internal URLs, API paths, and confirmation or unsubscribe tokens are not retained as referring paths. Older records did not preserve this distinction and remain unknown where the original value was discarded.

The browser-beacon dataset collected before August 26, 2026 remains intact. Its source table stores page path, cleaned referrer host/path, country, city, continent, a daily date-derived IP + User-Agent hash, heuristic traffic and device classes, owner status, and UTC time; it never stored raw IP addresses or raw User-Agent strings. A minimized, source-marked copy is included in public stats so the historical trends are not lost.

Routine first-party observation rows are retained until I delete them; there is currently no automatic expiration for those rows. The source marker distinguishes historical beacon events from newer edge observations.

Private operational logs record observation attempts and outcomes. Successful storage logs repeat a limited subset of the recorded classification, reviewed client declarations, verification outcome/profile, signature-coverage flags and any recorded candidate-selection counts, reviewed signer name and provider bot facts. The operation reference connects the receipt to its request's Cloudflare log context. From October 27, 2026 UTC, an available database row reference also links it to the exact stored event. Neither reference is included in public stats. These linked records follow separate provider-log retention and sampling, and restricted investigation-copy lifecycles. A missing receipt does not prove that a database write is missing.

Private request diagnostics

To investigate collection errors and incorrect traffic classifications, an optional private diagnostic layer can retain bounded, filtered User-Agent and Accept header evidence, claimed browser and operating-system family and major version, client form, coarse preferred language, response status, HTTP and TLS version/cipher, Cloudflare data-center code, request correlation references, and collection/verifier versions and outcomes. Capture is limited to recognized public pages. It does not request additional device information through browser probes.

Names, email addresses, cookies, authorization, form bodies, unrestricted URL queries, full external referrers and precise location are not requested for this layer. Familiar identifying and credential patterns are filtered from header evidence; arbitrary client-supplied text is still treated as potentially personal. Redaction and truncation are recorded so incomplete evidence is not presented as an exact original. These records are private and are not included in public stats or joined to newsletter/account records.

The optional layer is omitted when the request carries Global Privacy Control or Do Not Track. That does not disable the separate routine page counter described above. Diagnostic source records expire from private reads after 60 days and are removed by scheduled cleanup; cleanup failure can delay physical removal. Cloudflare's database backup history can retain earlier states for up to thirty additional days. Restored copies must be cleaned before diagnostic access resumes. Routine operational logs and Cloudflare Web Analytics have separate retention and processing.

For access, correction, deletion or objections concerning this processing, use the contact links below. A daily pseudonymous client ID does not establish a person's identity; I do not create a cross-day identity solely to answer such requests.

Cloudflare Web Analytics

Explicitly activated owner-excluded measurement trials can retain closed stage counters, a random trial reference, planned condition/operator labels and an expiration time. These references join to routine excluded observations and are not anonymous. Activation lasts at most fifteen minutes and receipt access requires the current-day owner mark. Counters become deletion-eligible 24 hours after activation ends; the nightly job performs bounded cleanup, so eligibility is not a maximum physical retention guarantee. The routine excluded observations retain the lifecycle described above. Planned labels do not establish actual human operation. These trials request no additional device probes and do not enable the optional raw diagnostic layer.

Cloudflare Web Analytics is separate from the first-party dataset above. Its JavaScript beacon supplies Cloudflare's performance dashboard and does not feed this site's public stats. Cloudflare may process request and browser data for that service; see Cloudflare's Privacy Policy.

AI site check and request check

The robots playground and saved-report comparison process pasted policy text and report files in your browser. These tools do not upload those inputs or spend the hosted check allowance. Ordinary page requests still use the site services described here. A saved example is a dated public report; opening it performs no fresh target check and uses no daily allowance.

The AI site check fetches the public HTTPS page you submit, a fixed set of discovery files on that origin, and a few synthetic client-declaration probes. It does not execute the page's scripts or send your cookies, authorization or other incoming headers to the target. Public DNS is checked through Cloudflare DNS-over-HTTPS. The target site receives the checker's requests and may log them. Use public URLs; credentials, queries and fragments are rejected. Reports are returned to you without being saved in the application's database or shared cache.

The request check assesses the request that reaches this site. It returns derived declaration, request-shape, network and signature evidence without echoing raw addresses or User-Agent text. Signature checks may retrieve a public key directory; recognized client declarations may cause a fetch of a vendor's fixed public IP feed. Your address is compared locally and is not sent to the vendor feed.

Both checks share an allowance of ten admitted requests per public edge address per UTC day. An atomic quota row retains a purpose-specific, daily HMAC of the address, the UTC date, a count and reset time. It stores no raw address, target, headers or report and is separate from analytics identifiers. Shared networks share the allowance; changing networks can change it. Rows become cleanup-eligible one day after reset and nightly deletion may be delayed by cleanup failure. Database backups and provider request logs have their separate retention. Browser checks are same-origin. Short burst protection also applies.

Newsletter

If you request a subscription, your email address is stored in a Cloudflare D1 database and sent to Resend to deliver confirmation and newsletter emails. You become a subscriber after opening the confirmation link and pressing the confirmation button. The signup record may also include the page where you subscribed, an allowlisted campaign source and campaign name, the referring site's hostname, a truncated request IP address, a browser user agent, and request and confirmation times.

To limit unwanted confirmation emails and diagnose delivery failures, the site retains confirmation reservations for seven days. These records contain the email address, a random request identifier, a hashed confirmation token, timestamps, send outcome, and any returned provider message identifier and HTTP status. They are private operational records. An accepted send does not establish that an email reached the inbox. Routine confirmation diagnostics use the request identifier and outcome, without the address, form contents, or tokens.

You can unsubscribe via the link in every newsletter. Confirmation emails also offer a way to block further confirmation requests for up to 90 days. Unsubscribed and bounced addresses, including their suppression state, are purged after 90 days; expired pending signup records are removed by nightly cleanup. An already admitted email may still arrive after an opt-out.

Bot protection

The newsletter form loads Cloudflare Turnstile and initializes its widget when the script is available. Verification is requested on submission. Turnstile analyses browser and network signals to assess automated abuse. Verification may require interaction and can fail; the form reports a problem and lets you retry. Cloudflare may process data as part of this assessment — see Cloudflare's Privacy Policy and Turnstile Privacy Addendum.

Error diagnostics

If something breaks in the browser, a small first-party error report may be sent to help debug the problem. Reports can include bounded error messages and stacks, page and source paths, a cleaned referrer, component name, browser User-Agent, build information, and coarse Cloudflare request metadata and correlation references. The built-in reporter does not read form values, cookies, localStorage contents or Turnstile tokens to populate these reports.

URL filtering removes queries, fragments and URL credentials from structured URL fields and recognized HTTP URLs in messages and stacks, and redacts known confirmation and unsubscribe token paths. It cannot detect every personal detail or secret in error text, other path segments, User-Agent or build information. Reports are private and treated as potentially personal data. Database records become eligible for nightly cleanup after 30 days; cleanup failure can delay physical removal. Private operational logs can contain error text and paths even when database storage is unconfirmed; provider logs and database backups have separate retention.

From October 29, 2026 UTC, new error records can also preserve the client-reported occurrence time separately from server ingestion time. Subscription diagnostics can record whether an attempt was idle, verifying or sending, and a six-digit Turnstile callback code when available. Missing, malformed and unsupported evidence stays explicit; arbitrary callback text is discarded. These are unverified browser reports, not proof of a human, a distinct attempt or successful report delivery. Earlier records are not backfilled.

New subscription error reports can include a random reference for one form attempt, elapsed time, the browser's online/offline and page visibility states, widget and script readiness, and whether recovery is automatic or requires another submission. The attempt reference can connect a browser error to the corresponding server operation, including when its response is lost; it is not a persistent reader identifier. Early reporting also records fixed failure stages when the main script or Turnstile cannot load, or the page's security policy blocks Turnstile. Duplicate and per-page budgets limit these reports. Server verification logs retain bounded provider error codes, HTTP status, retry count and timing, without tokens, addresses or arbitrary provider response text.

Interactive word and related-idea searches run locally in a browser worker. Related search downloads self-hosted E5 model files and document vectors; browser caching can retain these public files. Search does not save a query history or send interactive queries to a third-party search or embedding API. The plain search form works without JavaScript and sends its query in a GET URL to this site. Those URLs can appear in infrastructure request logs. Opening a result is an ordinary page request.

Other services

No advertising networks, social tracking pixels, or session recording are used. Cloudflare provides site infrastructure, Cloudflare Web Analytics, and Turnstile; Google Fonts supplies fonts.

Contact

Questions? Reach out on X or LinkedIn.

The license for everything published here is on the license page.