With real testers now running the Android closed test, the first bugs they found were all about the same thing: content squeezed off the edge of a phone. On the roadmap the row of view tabs (This week / Full roadmap / Costs / Timing) plus the PDF and Share buttons was crammed onto one line, so on a narrow screen the Share button was clipped and — worse on Android, where the system font runs wider — the whole right side vanished. And during the interview, the last row of answer chips could hide behind Android's on-screen navigation bar. The fix splits the roadmap controls into two rows: the view tabs now swipe sideways in their own scroller so they never clip no matter the language, and PDF/Share get their own always-visible line beneath. The interview now pads its answers by the exact height of the device's bottom navigation bar, so the last chip always clears it.

The fresh-eyes pass had flagged one thing to come back to: the same one-day timezone slip it caught in the new timeline probably ran through the whole app. It did. Every date in the app is stored as midnight UTC, but the roadmap, the printable report, and the weekly email printed those dates in the reader's own timezone — so anyone west of Britain (most of the Americas) saw every deadline a day early, and a step due today could read as already overdue. It never surfaced in testing because it's invisible anywhere at or east of London — which is exactly where the app is built and where the automated tests run. The fix makes the whole engine read and write dates one consistent way, midnight UTC end to end, and the new regression test deliberately pretends to be in Los Angeles so this exact bug can never hide behind the timezone the machine happens to sit in.

Before real testers arrived, a cold read of the just-shipped Costs and Timing code found a handful of things worth fixing. The tax-residency pivot was off by a day: the rule is MORE than 183 days, so the last arrival that still counts is 1 July, not 2 — corrected in the engine and the copy. The timeline's milestone dates were built at midnight UTC but printed in the reader's own timezone, so a 1 July arrival showed as “Jun 30” for anyone in the Americas while the headline said July — now printed consistently. The tax card stated residency as a certainty; it now carries a one-line caveat that it assumes you actually stay past 183 days. On the Costs side, the “Fixed fees” total quietly mixed government-checked fees with two nationality-dependent estimates — those estimates now stay in the range but out of the number we present as certain — and a buyer who hasn't entered a home price yet gets told why the total looks low instead of being misled.

A second view that answers a question a checklist never can: when to arrive. Pick an arrival month and your real roadmap re-computes — every deadline shifts with it — because the plan engine already dates everything relative to your arrival. The payoff is a rule almost no one sees coming: Spain treats you as a tax resident for a whole calendar year only if you spend more than 183 days in it, so landing a few weeks either side of early July changes which year Spain first taxes you — and moves your first Spanish tax return by a full year. The screen computes that from your arrival date and says it plainly. Two honesty lines held: the milestone dates that are still estimates stay marked, and school-enrollment windows show as a flag (“ordinary window” vs off-cycle “fuera de plazo”), never a made-up date — because those windows change yearly and we won’t ship them secondhand.

The roadmap already knows which steps apply to you and when they're due — this adds what each one costs, answering the question every mover asks. A researched pass priced all 73 catalog steps and sorted them four ways: fixed official fees (a set government tasa — the exact euro), costs that scale with a number only you know (property transfer tax as a percent of your home price), personal costs we won't put a figure on (a scouting trip, a lawyer — your budget, not ours), and steps that cost nothing (kept, but folded away). It reads as three honest layers behind a new "Costs" tab, with the total always a range, never false precision. It ships switched OFF behind a flag: the fixed government fees have since been checked one by one against their official source (the 2026 BOE tasas order, Instituto Cervantes, the DGT and Seguridad Social schedules) and now show "verified"; the percentages and market prices stay honest "estimates" until they're checked too.

Everything Google needs to start the two-week closed test is now in and submitted for review. Phone screenshots were captured on a real Android phone and framed to the brand — with a purpose-built renderer that keeps them inside Google's exact size rules (a raw phone screenshot is actually too tall to accept, and just under the resolution that qualifies for promotion). The store listing was finished (category, contact), and the verified build was promoted to the closed track — reused, not rebuilt, because Google refuses a re-uploaded version number. And the Android-visitor invitation on the site went live: it points to a short form to leave an email, so we collect testers first, then enrol them.

The Android app cleared its hardest gate. Play Integrity — the sibling of the iPhone's hardware install-proof — now verifies a genuine Google Play install and mints the short-lived token the interview needs, confirmed on a physical phone (Lola answered). The first production build went up to Google's internal testing track through a one-command automated pipeline, installs cleanly, and runs. Android will ship phone-only, matching the iPhone. And an invitation for Android testers was built into the site itself — shown only to Android visitors — so the closed test recruits from people actually moving to Spain, not a rented crowd.

Last month the home page promised "no official store badge appears until there's a real listing behind it — swapping in the link is a one-line change when the apps ship." The day after launch, that line was paid: the App Store pill now says "Download on the App Store" and actually downloads the app, while Google Play keeps its honest "coming soon." The band's copy was rewritten in all five languages for the true half-way state — the iPhone app is here, the browser still does everything, Android is next, and the EU storefronts unlock soon.

Twelve days and four review rounds after the first submission, Apple approved the app — and the release button was pressed the same morning. After the rival-store pill, the dialogue continued in writing: Apple asked whether the AI is built in-house and where users consent to it (answered plainly: the conversational layer is a leading AI provider, every step and date decision is our own deterministic engine, only conversation text ever leaves — no names, no emails — and Lola's first bubble discloses all of it). Then a screenshot caption containing the word "free" was ruled a price reference (one caption reworded), alongside a questionnaire about paid subscriptions the app doesn't have (answered: nothing is for sale). Round four passed without a word. Approval waited behind a manual-release switch until this morning; the iPhone app is now downloadable worldwide — except the EU storefronts, which unlock when a pending business-identity verification with Apple completes. The web app serves everyone, Spain included, meanwhile.

Eight days after submission, Apple reviewed the app and rejected it on exactly one point: the home screen's "coming soon" pill for the other major app store. Their rule (Guideline 2.3.10) is that an iOS app may not carry references to third-party platforms, and the app renders the same landing page the website does — store-availability band included. The fix: that band is now web-only; inside an installed app it was redundant anyway ("download our app" is a strange thing to tell someone already inside it). One platform gate, a regression test so no rival-store copy can quietly return to a native build, and the corrected build heads back to Apple.

Error monitoring caught two real visitors — one in Barcelona, one in Slovakia, both arriving through Facebook links — whose interview silently died: every answer waited 20 seconds, then failed. The cause: inside Facebook's built-in browser, the invisible "prove you're human" check decides it needs the person to actually click something — but the widget was rendered off-screen, where nobody can click it. Now, when the invisible check can't pass, a small visible "quick security check" appears once, the person taps it, and the interview continues where it left off. The wait before that fallback dropped from 20 seconds to 8, a browser that has failed once skips straight to the visible check on later turns, and the fix carries five new regression tests.

The call was made to stop hunting for bugs internally and let real users be the test: the App Store release-candidate build was cut and sent to review. Before cutting it, the app gained over-the-air updates — future fixes to the app's logic and copy reach installed phones on next launch, without a new App Store review each time. The Android track also moved: a test device is in hand, and the launch playbook was re-checked, catching that Android would need its own hardware "this is a real install" proof (the sibling of the iPhone's App Attest) before its first test build — otherwise the interview would ship dead on Android.

Three fixes straight from testing the real build on a phone. The security handshake now happens the moment the interview opens — quietly, while you're reading the first question — instead of pausing Lola's first reply while it runs. The contact form's confirmation email is fixed: it was being sent in a way the phone's server quietly discards the instant it replies to you, so senders never got the "we got it" note even though our inbox did; now it's sent properly. And Lola's messages now put the actual question in bold, so you can skim the reasoning and still see what's being asked.

On the web, one Cloudflare human-check protects the AI endpoint. The phone can't show that puzzle without nagging you, so the native app proves it's a genuine, unmodified install a different way: Apple App Attest. The moment you open the app, the iPhone's Secure Enclave signs a hardware-backed certificate that our server verifies — no tap, no puzzle, nothing you ever see — and on success you get exactly the same short-lived token the web earns from the check. We built the full verifier (it walks Apple's certificate chain up to their published root and confirms the signed challenge is fresh and bound to this exact device) and proved it end-to-end against a real attestation captured from the actual build on a physical iPhone.

A batch of polish before the next mobile build. A proper "My account" page pulls email preferences (turn the weekly roundup on or off), language, and account deletion out of the hamburger menu into one clear place. The interview's first question got quick chips — "In a few months," "Later this year," "Next year" — so you can tap a timeframe instead of typing one, with the composer still there for a specific month. The household-scope note ("this maps the main applicant; a working partner runs their own quick interview") now shows in the interview too, not just on the finished roadmap. And a monthly "freshness" robot checks every dated official fact against a staleness deadline and pings every government source URL — so the "last verified" stamps stay honest because a machine makes us keep them.

Every fix the five-seat review raised is now done. Lola stops contradicting the roadmap: one shared validator guards every place the AI can touch your profile, the income question is answered by the engine itself (so a band the plan is about to flag never gets praised), a personal-case question in the final note gets an honest "I can't assess that — that's for a consulate or a lawyer" instead of a bland "got it," and the opening bubble states plainly that Lola is an AI and this is guidance, not legal advice. The AI endpoint got a Cloudflare human-check (solve once, a short-lived token covers the interview) and a budget-drain alarm. Penalty steps like the Modelo 720 foreign-asset declaration are now kept — flagged "may apply" — when you'd rather not say, instead of silently vanishing. Share links carry your answers in the URL fragment, which never reaches a server. The national legal figures (the €28,800 income floor and friends) live in one registry, with a test that the prose can't drift from it. The income threshold counts your real number of dependents. The privacy policy gained lawful-basis, data-transfer and retention detail. And the home page now shows one real, cited, dated step instead of four abstract trust badges.

Five independent reviewers — each starting cold, from tech, legal, PR, marketing, and operations angles — went over the whole product before launch, and their findings became a fix queue. The first one shipped: the server endpoint behind Lola no longer accepts a caller-supplied instruction prompt. The app now sends only a named task — extract this answer, phrase this question, coach this step, distill this note — and the server writes the actual instructions itself, choosing the model and length limit per task. A copied link can no longer be turned into a free, general-purpose AI proxy; it can only ever run Lola's five real jobs. Riding the same deploy: a German homepage headline that had drifted into an awkward idiom ("Damit gehen Sie nach Hause") became the clean line it was meant to be ("Das nehmen Sie mit.").

The morning of the App Store submission, one more fresh-eyes pass — this time over everything around the app. The contact form got volume limits (a human files a few reports; a flood now gets turned away politely), and two small bugs in it were fixed before any user met them: iOS reports mislabeling themselves as web, and a signed-in visitor's email not pre-filling. The store screenshots were regenerated from a retake after the review caught one still showing the retired voice button. And the operational side got its own audit: every third-party service behind the app mapped into one document — who charges money, where the renewal notices go — so nothing expires in an unwatched inbox.

One afternoon, four features that had waited in the backlog. A public changelog now records every catalog correction, dated and in plain language — and every step carries a "last verified" date linking to it. Your roadmap became shareable: a read-only link that encodes your answers (never your notes) and renders through the same engine, no account needed. The sample plan became four sample lives — US retirees, a German family on the EU path, a Canadian digital nomad, a British couple who already own a home here. And the first verified regional figures shipped: real transfer-tax rates for the four biggest expat regions plus the Madrid/Andalucía wealth-tax reliefs, each checked that day on the region's own tax portal.

The interview's closing "anything else I should know?" note is no longer just stored — it's distilled through the same bounded extractor as the roadmap's "something changed" box, so "the dog is coming too" adds the pet steps before you even reach your roadmap. Answers that simplify your plan now say so: a "−N steps — simpler for you" pill joins the "+N new steps" one, instead of steps vanishing silently. And a UK tester's flag fixed a US-centric step title: the consulate-appointment step no longer quotes US wait times to everyone — wait times vary widely by consulate, and the step now says exactly that, in all five languages.

A real household using the product found what no test had: a Spanish passport holder offered the foreigner registration, a mixed household whose American spouse had no residence step at all. So the whole rule engine got audited in one sitting — every condition on every obligation, against a new tool that builds plans for 181 realistic profiles and checks class-level expectations. Fifteen findings; the fixes shipped the same night: the padrón no longer vanishes for movers still choosing housing, job seekers aren't routed to the no-work visa, short-stay tourists don't get the residence-visa roadmap, EU licences aren't sent to the driving exam, EU citizens can finally say they want citizenship. Then the audit's backlog: nine new officially-sourced steps (employer-sponsored work visas, self-employment, students, the Schengen clock, the EHIC — and the one-year citizenship fast-track for spouses of Spanish citizens), the driving-licence exchange list corrected against the DGT's own table, corporate tax gated behind one new question, and verified regional tax figures completed for every region on the map, foral territories included.

A full fresh-eyes pass (Cristina) closed the week: Lola's spoken voice is gone entirely — opt-in wasn't enough; text-to-speech never fit the fast chip interview — while microphone dictation stays for those who need it. The marketing copy caught up ("a short interview", not "a conversation"). The interview's ending got its missing beat: your final note is acknowledged, then "Getting your roadmap ready — 3… 2… 1" hands over to the roadmap instead of an abrupt jump. The mic button now draws the standard dictation glyph instead of an emoji. And "Report a problem" grew into a proper Contact page — general questions, feedback, or problem reports, one form, linked from every footer.

An evening testing pass by the builder and his wife produced six findings, fixed the same night. The biggest: toning down Lola's reactions earlier that day had gone too far — she'd stopped noticing things. Her reactions now see the whole conversation again, so she can connect your answers ("the whole family is coming, dog included"), while the hard rule stands: she may never state a fact, number, or deadline outside your sourced roadmap. Also: arriving from a guide page is acknowledged again ("I see you've been reading about the padrón…"), the first question opens with the move itself instead of a clever framing, voice is now opt-in rather than on-by-default, a stale menu item went away, and the roadmap sheet no longer hides its close button under the iPhone's Dynamic Island.

The homepage and the "how it works" page merged into one scrolling story — but not by guessing. Four different homepage philosophies were built as local prototypes and walked side by side: one where the page IS the first interview question, one that demos the product in the hero, one that leads with the finished roadmap, and one that keeps the postcard warmth of Spain up top with the demo one scroll below. The postcard won. The new page keeps the rotating photos, then shows a roadmap literally building itself as questions get answered, then the proof: a real sample roadmap and the promise that matters — 100% free, no catch.

The first strangers to use Get Camino told us — through the analytics, not in words — exactly where it lost them: the interview. The worst question was the second one asked, finishing took twice the promised time, and most people who started never saw a roadmap at all. So the interview was rebuilt around one idea: show the value while you answer, not after. Your roadmap now builds live beside the conversation — every answer visibly adds steps (on a phone, a step counter grows as you go). Most questions became single taps instead of typed sentences, every question now says why it's being asked, and the opener is the question every mover is already asking themselves: when? Fewer questions, too — an audit traced each one to the steps it unlocks, merged two pairs into single taps, and deleted one that turned out to feed nothing at all. And if you leave halfway, your progress now waits for you when you come back.

The first public exposure: one plainly-disclosed post in a moving-to-Spain community, web only. About 1,200 views and two complete stranger journeys — first visit to finished roadmap — one in English, one in Spanish, zero user-facing errors. Hours later the post was auto-removed by the community's moderation machinery, the standard fate of anything reported as self-promotion. The overnight watch stood down; the always-on error and uptime alerts kept the site paged regardless.

The web had spoken five languages for a day; now the iOS app does too. A new TestFlight build carried the localization work and the mid-interview language fix onto the phone, gated behind the full native release checks first — the simulator user-flows had to pass before a single build credit was spent. Verified on a real device: the interview switches languages mid-conversation, every menu item reads right in all five, and the English-only build story hides itself outside English.

With five languages live and real family testing underway, a round of fixes: the sign-in code email now arrives in your language (it comes from a different system than the other emails and had been missed), the interview responds instantly when you switch languages mid-conversation instead of forcing a reload, and small wording polish across languages. The headline, though, was a silent one caught by an eagle-eyed look at the database: signed-in users' saved roadmaps had quietly stopped persisting.

Before translating the app into Spanish, every layer of testing was audited with fresh eyes: the coverage map verified against what actually runs, ten critical user paths given honest verdicts, and last week's under-pressure scoping decisions re-examined. Three new safety nets landed: live tests proving the interview already understands Spanish answers, pixel-exact snapshots of every email and the printable report, and API checks that now run on every staging deploy.

In one day, the whole product learned Spanish: Lola interviews in warm, informal Spanish; all sixty roadmap steps and their sixty written guides read natively; dates, emails, and the printable roadmap follow your language; the legal pages carry a faithful courtesy translation; and the sample roadmap belongs to Susana y Tomás. English users see a byte-identical product — proven by frozen snapshots that didn't change.

The same afternoon Spanish shipped, the plumbing proved itself: French, German, and Italian followed — each with the full interview, all sixty step titles and sixty guide explainers, emails, and the printable roadmap. The web grew real per-language pages that search engines can find, every page telling crawlers about its four siblings. And a user question exposed a real gap, fixed the same hour: your language now follows you from the browser into your inbox, so the weekly email never speaks the wrong language.

Every button, question, and label in the app's core screens moved out of the code and into English string catalogs — the plumbing that lets the whole experience ship in Spanish next. A visible language choice now lives in the menu, and four automatic checks guard every future translation: no missing strings, no translation may ever change a number, placeholders must survive, and the names "Get Camino" and "Lola" stay exactly themselves in every language.

The home page now says where the app is headed: a quiet band announcing the iPhone and Android apps — honestly marked "coming soon," not a fake download badge for something you can't download yet. And long scrolls got a back-to-top button: on the roadmap, the sample plan, and the sixty-step guide list, a floating arrow appears once you've scrolled a screen down and takes you home in one tap.

A moving-to-Spain product that only speaks English lacks authenticity — so localization moved from "someday" to before launch. The technical design is written and approved: Spanish first (verified by a native-speaking family member), then French, German and Italian as fill-in-the-blanks exercises on the same plumbing. The engine stays language-free; translation is purely presentation.

The product stopped using a personal Gmail: feedback@, privacy@ and legal@getcamino.app now route reports and rights requests (delivery tested end-to-end), and the legal pages name the actual operator — AELaboratories, Inc — with governing law updated to match its state. The path to a formal company is deliberately deferred: sole proprietorship until there's a proven revenue model.

The how-it-works page — untouched since the earliest days — got a layout pass: its sections ran edge-to-edge on desktop while every other page reads in a centered column, so each section's content now sits in the same 760px column, and the bureaucracy card says what the plan just learned (rules that differ by comunidad are flagged). This page was also renamed "the product roadmap".

All 55 official source links were fetched and checked against their steps — status, where they redirect, and whether the page is actually about the step. Two failed honestly: the empadronamiento link died when a government portal was retired (now points at the law itself, art. 15 of the Ley de Bases del Régimen Local), and the NIE link landed visitors on a police-site cookie wall instead of NIE content (now the Ministerio del Interior's canonical NIE page).

"Camino" is a lovely word — which is exactly the problem: plenty of businesses already want a piece of it. Until a proper trademark search happens, the app and site go by Get Camino (matching the domain we already own), extended brand "Get Camino: Your Road to Spain" — which happens to fit the App Store's 30-character name limit exactly. Every user-facing surface renamed: the name under the icon, the nav, the emails, the printed report, the legal pages, and how Lola introduces the product.

Family testing round four (build 27) nailed the last voice bug with a precise diagnosis: earlier builds only DUCKED Lola's voice while the mic was open — and never released the duck, so everything after went quiet. Build 27 fixed the volume but left her talking over you. Now it's explicit turn-taking: opening the mic cuts her line outright, and every later line plays at full volume. Also fixed: your spoken answer no longer reappears in the next question's box (the recognizer flushes one last result after it's told to stop — that flush was refilling the input the send had just cleared).

The two endpoints that cost real money per call — Lola's brain and Lola's voice — now carry real volume limits: a per-visitor per-minute cap and a global daily budget each, counted durably in the database. The in-memory counters they replace reset whenever the platform recycled a worker (a 70-request burst produced zero throttles); the new ones trip at exactly their limit, verified live. Cross-site browsers are locked out too: the platform's default wide-open CORS answer is replaced with a strict one pinned to our own origins.

Tab through any page and an amber ring — the brand's waypoint color — shows exactly where you are. Keyboard users get the ring; mouse and touch clicks don't.

Family testing round five, all fixed: the app could cold-start into an eternal "Loading your roadmap…" (it now knows when there's nothing to load and takes you home); an interview answer could spin forever if the network hung (35-second limit, then Lola owns it and your answer is back in the box to resend); the microphone cue now appears when the mic is actually hot — so fast talkers stop losing their first words — and tapping it off keeps your last words instead of discarding them; and the answer box grows again as you dictate.

End-to-end tests grew up: the web suite now signs in as a seeded test user (through the real magic-link machinery, no inbox involved) and walks the journeys that matter — saved roadmap, the week view, mark-done and undo, the no-op honesty gate, sign out. Twelve tests, twelve seconds, against the live staging site. Native flows are written for the iOS simulator and run on a free CI runner with zero build credits. First catch, before it even finished: a React hydration error firing on every page load (the nav wordmark rendered differently on server and client).

Spain isn't one rulebook: transfer tax, wealth tax, school admission windows and property tax all vary by comunidad autónoma. The interview now asks where you're settling (name a city and Lola places it in its region), and every region-varying step says so — on your roadmap, in the guides, and on the printed report — naming your region when it's known.

Search engines now get structured data (article + breadcrumbs on every guide, the full catalog as a list, the site itself); arriving at the interview from a guide gets acknowledged ("so glad you're looking into getting your NIE sorted") before the first question; and an accessibility pass landed — the rotating home photo rests when your system prefers reduced motion, controls carry proper roles and labels, and inputs are named for screen readers.

Every public page now unfurls properly when shared: a branded card (the compass-star tile, the promise, the domain) plus per-page social titles and descriptions. One Seo component owns the whole tag set, so no page can ship half of it.

Privacy policy, terms of use, and aviso legal — plain language, linked from every footer. Delete my account, in the menu: one warning, one tap, immediate and permanent (tested live, start to finish). And web analytics went cookieless, so there's no consent banner because there's nothing to consent to.

A former QA engineer joined the testing and it shows: Lola's voice dropped to phone-call volume after the first line (the microphone session was rerouting audio to the earpiece — playback now reclaims the loud speaker every time); a clarifying question mid-interview got a canned "sorry, didn't catch that" (Lola now answers what you asked, then re-asks — still forbidden from stating legal facts there); "on a date" became "did it earlier? add the date" with a forgiving date field that reads "25 April 2026", Spanish months, and "yesterday" — and shows you what it understood before saving; and the printable report got print-safe contrast and real page margins.

All sixty guide pages gained a written explainer — what the step actually is, how the process feels, who's involved — in Get Camino's voice. Each page's search snippet now uses its own opening sentence instead of a shared template.

A second phone found four more: the feedback spinner could run minutes after the report had already arrived (the response stalls on some networks; a grace period now thanks you once the send is safely away); the step drawer's inputs hid under the keyboard (it now slides above it by the exact overlap the OS reports); Lola claimed to have "remodelled your plan" when nothing had moved (a no-op re-plan now says so honestly, and appointment dates are steered to "mark done on a date"); and the nav bar wrapped into two lines at larger iOS text sizes (bar labels now cap their scaling — menu items still scale fully).

The home and how-it-works pages predated almost everything shipped this week. Now: every public page has a real title, description and canonical URL (only the guides had them); how-it-works finally ends with a call to action, mentions the weekly email and the sourced guides, and uses the shared footer; the home topic strip links into the sixty guides it was always describing.

The roadmap became a report: one tap renders a clean, print-ready PDF — hero next step, estimated vs firm dates, dateless steps that say what they're waiting for, official source URLs printed for the gestor. And a "Report a problem" line joined the menu: one text box, straight to the team inbox with platform, version and route attached. iPad support switched off for the first release.

This whole section — the essay, this log, the roadmap — left the unlisted shadows: it's in the nav menu and the sitemap now, and every entry carries its ship date. And the links in Get Camino's emails gained universal links, so tapping "Open your roadmap" on an iPhone opens the app (signed in) instead of a logged-out browser tab.

A five-build mystery: Apple's own sign-in sheet rejected only this app, while every layer we could check was provably correct — the entitlement inside the signed binary, the App ID capability, fresh provisioning profiles, the auth provider config, active developer agreements. The failure lived in none of them: Apple's server-side provisioning for the App ID had gone stale. Toggling the capability off and back on forced a re-provision, and sign-in started working — even on the older builds.

Real devices found what simulators don't: the emailed sign-in code is 8 digits but the app said — and silently truncated to — 6, so codes could never verify; and the step drawer wouldn't scroll on iOS, hiding Lola's answers.

The roadmap gained a second lens: a toggle that filters to just what needs attention now — what's slipped past, and what's due in the next seven days. A clear week gets an honest "nothing needs you this week", with the next dated step so you know what's coming.

Every catalog obligation got a public page (/guide/<id>): when it's due, why it matters, what comes first, the official source, and a category tip — plus a grouped index, prerendered HTML titles and descriptions for search engines, and a sitemap. The nav also unified: desktop now gets the same ☰ menu as mobile, with Sign out inside.

The welcome email and weekly roundup went live end-to-end: server keys landed, both environments redeployed, a real weekly run fired from the scheduler, and real emails read in a real inbox. The live test caught two bugs no unit test saw: email links leaked the hosting platform's per-deploy URL instead of getcamino.app, and one sign-in sent the welcome email three times.

This section gained a third page: the product roadmap, updated with every release — shipped, in progress, next, and the things we've said no to.

Passwordless sign-in (magic link + one-time code) joined the dialog; a signed-out roadmap can be emailed to yourself — which quietly creates your account, with the roadmap riding along; a welcome email greets new users; and a weekly roundup engine (overdue + upcoming, capped at five, deterministic per-item tips) waits on its cron.

Signing in became one link that opens a dialog (Apple + Google, email next); narrow screens got a hamburger menu holding the browse links, leaving the bar to the two actions that matter.

The interview composer clipped behind the iOS keyboard through two attempted fixes. The framework's keyboard-avoidance component was removed entirely, replaced by a hook that pads by the exact overlap the OS reports.

Steps past their due date now say so in red — on the card, in the stats, and in the step sheet, which nudges: mark it done with the real date, or tell Lola what changed and the plan re-flows.

The official Apple button on iOS, exchanging Apple's identity token directly for a session — no password, no web redirect. Three failed builds taught us Apple's provisioning model the hard way.

A public, read-only roadmap for a fictional couple — built by the real engine on a canned profile — with tap-to-expand deterministic detail and interview CTAs.

A deterministic engine suite (15 tests, ~140 ms) runs in CI on every push and inside the deploy gate; API contract tests and a Playwright smoke suite run against the deployed staging site.

All 28 uncited official obligations re-verified one by one against their issuing authorities (tax agency procedure pages, ministry information sheets, the BOE, the DGT, the Instituto Cervantes) and given canonical links; a missing EU-registration obligation was discovered and added.

A full fresh-eyes review of architecture, catalog, and interview found real bugs: a language-exam rule mishandling Filipino applicants, two interview answers collected but never read, the drift-check invariant living in dead code, and 28 "official" items with no citation attached.

A new interview question (do you hope to become a citizen, or keep renewing residence?) now gates the entire citizenship track; a second new question gates the scouting trip to people still choosing a region.

Three TestFlight builds brought iOS level with web: spoken Lola (expo-audio streaming a GET variant of the TTS route), native crash reporting with readable stack traces, native analytics.

Sentry across web, the API routes, and native iOS — one project tagged by platform and environment — plus an uptime monitor that pages on downtime, and daily budget caps on the paid API routes.

Every webinar-derived claim researched against government sources; corrections applied (regional tax ranges, renewal windows, a voluntary-not-mandatory registry); the taxonomy simplified to official | recommendation.

ElevenLabs TTS behind /api/tts (key server-side), a warm Spanish-accented English voice, on by default, with a visible mute pill.

PostHog on web and later native: a funnel from first visit through interview to roadmap, feature-health events, person profiles.

The azulejo compass-star became the app icon set (iOS, adaptive Android, splash, favicon), generated from a single SVG by a script.

Dev test personas and internal cross-check links, first gated by environment + a user-id allowlist, later replaced by a server-owned is_staff flag on the profiles table.

Native builds through EAS with an App Store Connect API key (fully non-interactive), through review rejections to installable TestFlight builds; native Google sign-in, dictation, and Dynamic Island safe-areas verified on device.

EAS Hosting with development / staging / production environments, a custom domain, and a deploy script that pulls each environment's vars and clears bundler caches.

Mark steps done (today or back-dated); completions re-anchor downstream deadlines to what actually happened; a free-text "something changed" box re-plans deterministically.

Fifteen relocation-webinar transcripts mined by five parallel agents; the catalog grew from ~20 to 54 obligations across tax, property, visas, healthcare, and citizenship.

All Anthropic calls moved behind /api/lola (Expo API route on the hosting platform's Workers runtime); the client bundle no longer contains any key.

Expo Router app targeting web, iOS, and Android from one codebase; Supabase Google sign-in; profiles saved per-user under RLS.

A deterministic plan engine (~7 obligations), a catalog-derived interview, Lola's persona prompt, and a report prototype — the whole pipeline end to end before any real app existed.