Change Log

What changed in each release of Ilena.

1.6.220 September 2026

Under the hood

The two dashboards reporting 5xx errors now also count requests the ingress itself refused or timed out, not only the ones a running service answered — a minute in which every pod is down no longer reads as a quiet one.
Reading a batch of rhyme lookups back now returns them in the order they were submitted rather than alphabetically, and reading the same batch a second time no longer records its usage twice; a side-by-side comparison also confirmed a rhyme's stress marker comes back just as reliably through a batch line as through a synchronous one.
The dictionary dashboards now show how many times each entry kind found no answer in the corpus and had to be asked of the model in real time, beside the existing hit ratio.

1.6.119 September 2026

In the product

The suggestions offered as you write now draw on your idea basket: five entries are taken from it at random and offered to the engine as material that is available. Nothing obliges it to use them — a suggestion that takes up none of them is an ordinary suggestion — and you have nothing to select.
The basket's "image" kind becomes "idiom": a turn of phrase or a set expression you want to keep to hand. Ideas already saved under the old kind are kept as they are, as notes, with their content intact.
The four questions about you now open as four cards, each introduced by its own icon, rather than as a form.
An author whose synthesis has not been written yet now finds a sentence inviting them to answer the four questions, instead of an empty panel that said nothing.

Under the hood

`works` can now draw up to five short entries from an author's basket, with a random generator supplied from outside, so the draw is verified exactly rather than merely observed.
The drawn entries travel to the model under a key of their own, beside the author, project and work circles and outside their order of priority — the basket offers, it does not instruct.
Every idea saved as an image becomes a note before the kind disappears, its content and its source address kept.
The basket drawer, its three translations and the shared type follow the new kind, and the source address is no longer written anywhere.

1.6.019 September 2026

In the product

Selecting your own author node — or, for a studio admin, a member's — now shows a short synthesis of how you write, and an edit button opens a dialog to revise the four sign-up questions (what brings you here, the place writing holds for you, the themes you like exploring, your style) at any time.
The details pane now shows and edits a project's description and tags in place, the same way it already does for a work's.
Suggestions offered while writing a line now also draw on your project's description and tags and your work's tags and emotions, on top of the writing style you already declared.

Under the hood

Registration now records the language the sign-up funnel is running in as the new author's language preference, instead of waiting for the settings screen further along.
A project node can now carry a description and tags, read on the tree and written through a new endpoint.
Reading and writing the four self-description answers now targets any author a session may act for — oneself, or in a studio, a member — answers with the composed synthesis, and can erase a single answer without resending the rest.
Signing in, or finishing the four self-description answers, now recomposes an author's synthesis — once per account, since later sign-ins find nothing left to redo, and with no queue anywhere.
`works` now exposes an internal endpoint that turns four free-text answers into one short synthesis, for `profiles` to call.
An author's row can now hold a composed synthesis and the fingerprint it was built from; the four self-description fields it's built from are marked as personal data, included in the data export and deleted with the account.
The shared library now knows the composed-synthesis fingerprint and its own AI query type, so a change to it never silently changes anyone's suggestions.
Signing in no longer waits indefinitely for your synthesis to be composed: it waits five seconds and then lets you through while the composition finishes and is stored, instead of abandoning it and buying it again every time.
The addresses our services use to call one another are now written once for every environment; one of them was missing in production, where no synthesis could therefore be composed.

1.5.418 September 2026

In the product

The four writing tools each have their own screen during sign-up, instead of a grid where each had a quarter of one: a line’s syllables and emotion, a suggestion arriving, the nine dictionaries with their counts, the search through your catalogue. Each shows a real capture of the product rather than describing itself.
Sign-up now alternates a question and what it brings you, instead of asking everything in a row before showing anything.
The Continue button can be clicked where it is drawn, at every screen size. On a 768-pixel-tall laptop it went behind the footer or below the fold on five sign-up screens.
The opening screen shows Ilena at work: a capture that a click enlarges, rather than a description of what the dictionary can do.
An author who starts an annual trial and then disappears before activating now gets one reminder, and only one: it arrives before the trial ends, and does not repeat if you come back in the meantime.
A definition with one blank sense no longer costs you the others. The word keeps the senses the dictionary managed to write, instead of being asked for again whole.
A malformed dictionary answer is no longer simply dropped: it is asked for again in a shape we know how to read, so the word comes back to you rather than staying empty.

Under the hood

The screen called “social proof” now carries the name of what it says, and the one that announced the tools without showing them has left the walk.
A usage queue that stops being consumed now raises an alarm. The ledger of what the writing engine costs can no longer go flat in silence and read like a day without spending.
The spend dashboard now draws every reading its query returns. An hour in which a single service spent used to break the other lines, and real spending read as a service at rest.
The six services start with no configuration warnings, and a database request gives its connection back with its transaction rather than holding it to the end of the call.
A rhyme refusal that no stricter schema could have prevented is counted on its own series, instead of feeding the one that argues for a stricter schema.
The probe that tests the rhyme instructions against the real model now asks about the shapes that were failing in production — plurals and inflected forms — where its list held only simple words, and so measured nothing that was breaking.

1.5.317 September 2026

In the product

"Already verified? Continue" now actually checks: if your address isn't verified yet, it says so on the spot, next to the resend button, instead of waving you through to a payment you can't get back into.
Suggestions offered while writing a line now draw on the writing style you declared in the sign-up quiz — the metre, rhyme and emotion you asked for still come first.
Signing up with an address that already has an account now offers "Sign in" or "Use another address", instead of a red field error that reads like a typo.
Stepping back before Sign Up and walking forward again no longer tries to create a second account: it picks up the one you already made.
A selected answer in the funnel's quiz steps now reads clearly as chosen, instead of looking almost the same as one you're merely hovering over.

1.5.217 September 2026

In the product

An English or Spanish rhyme answer whose stress mark is missing no longer loses every candidate it found: the exact matches it can still support are kept, instead of the whole answer being thrown away.
French rhymes built from a tail that opens on a glide — jɔ̃, wa, ɥi — are no longer discarded: the vowel is found wherever it sits in the sound, not only at its very start.

Under the hood

An alert now fires when a scheduled dictionary review dies partway through, instead of its silence going unnoticed until someone reads a log by hand.
A model answer carrying a character PostgreSQL refuses now costs only its own dictionary entry: the rest of that review pass is written and closed, instead of the whole batch dying and being retried the same way for ever.
blog.ilena.app no longer competes with www.ilena.app/blog for the same articles: it now redirects to the canonical address, so search results stop being split between the two.
A registration that fails partway through no longer leaves an orphaned project behind: the new writer's starting folder is created only once the account itself has been saved.
The duplicate-work quota NodeService's documentation described is now written down as removed, instead of contradicting code that never enforced it.

1.5.116 September 2026

In the product

Dropping a file or a folder on a writer or a project — or picking the same from its menu — now brings it into Ilena as finished works: read, split into verses and stage directions, given their feet and emotions, and placed exactly where you dropped it. Five formats are read: .txt, .md, .rtf, .pdf and .docx.
Import needs an active subscription — a trial does not count, and the row does not even light up under any other standing.
Nothing an import brings in is ever merged into what is already there: drop the same file twice and you get two works, told apart by a number in parentheses.
The tree fills in while an import runs, and you can watch it happen: how much of your import allowance is left is shown as a percentage, never a line count, and a report at the end names anything that could not be read.
A PDF's running page number no longer becomes a line of the song it is imported from — even where one lands inside a stanza, between two lines of the same chorus.
A visitor whose browser asks for a language Ilena doesn't publish can now sign up in English instead of meeting a failure with no message.

Under the hood

The AI usage ledger now records why a call was made, so import's share of the model bill is exact rather than folded into everything else.
A drop is accepted or refused as a whole, with an upload location handed out per file before a single byte moves; a queued import then walks it, creating the projects and works, and resumes cleanly if its message is redelivered.
One uploaded file becomes one analysed work through works' ordinary save path, and one model call sorts a file's lines into verses and stage directions and names its language.
A writer's import allowance is counted in lines by works and resets on their subscription's anniversary.
The import queue, the storage bucket's CORS and the services' variables are in place on stage and preprod.
An alarm fires when imports fail or a reader starts refusing files, and a dashboard covers the rest.

1.5.015 September 2026

In the product

The payment screen now says what leaving the trial costs, which is nothing: cancel before day 30 and you keep Ilena until then, and an email announces the first charge before it arrives. The monthly plan has no trial and promises neither.
English and Spanish rhyme lists stopped coming back empty: the model was asked for the sounds in a way that made the answer unreadable, and twenty-three answers a morning were paid for and thrown away.

Under the hood

Where a sign-up is lost is now a reading anyone can take: sixteen screens, each saying whether it was read, refused or blocked, and naming the reason when it refuses.
That reading is written every night alongside the others, so the daily briefing has it without anyone running a command.
The funnel splits by device, browser and locale, and the annual trial's cohort is followed from its first day to its first charge.
A Checkout session ends in a readable outcome, and reaching the product's value has an event of its own — activation stops meaning only that a card was charged.
The blog's figures describe the blog people actually read: it moved under www.ilena.app/blog and the three exporters were still measuring its old address.
The daily blog evaluation pairs its sources by the day they describe, so a change of collection hour can no longer compare two different days without saying so.
The backlog alert reads a language's backlog rather than one pod's, so a deployment no longer wakes anyone while the backlog is falling.
Words an earlier release left out of the dictionary's review rotation are back at the head of the queue, instead of a month away.

1.4.914 September 2026

In the product

This page now shows recent releases in full and folds everything older than two weeks into a table, one line per release — it had grown to fifty versions, and the thirty oldest held more than half of it.

Under the hood

A dictionary line the review cannot buy stops being bought: the word leaves the due set instead of coming back for ever, and a word handed back keeps its place in the queue rather than going to the end of it.
The daily briefing separates real visits from scanner traffic, so its figures describe people rather than machines.
A milestone's acceptance can be settled on the local stage: a review pass can be asked for by hand, bounded to three words a language, instead of waiting for a nightly one.
The infrastructure files are laid out by environment and then by what the thing is, so a reader can tell where deployment, environments and tools live.

1.4.814 September 2026

Under the hood

One dashboard now answers whether Ilena is up, healthy and affordable, and it is what Grafana opens on: request and error rates, memory per service, and what the model costs each service hour by hour.
The alert on the dictionary's review fires when a word cannot be repaired at all, instead of waking someone for a gap the next pass fills six hours later.
An empty answer the retry recovers is no longer logged as a warning, so the warnings that remain are the ones worth reading.
The zone's security headers are declared in Terraform rather than living only in a dashboard, so a change to them is reviewed like any other.

1.4.713 September 2026

In the product

A word whose rhymes or definitions come back empty is asked once more before the answer is believed: a single unlucky draw no longer tells every reader, for ever, that the word has none.

Under the hood

A prompt change can be proved against the real model before it ships: a probe endpoint and a script ask it a fixed word list and count what comes back unusable.
The zone's alert rules are now evaluated against sample data with promtool, not just checked for their shape, so a rule that stops firing is caught in CI.

1.4.613 September 2026

In the product

A dictionary entry can no longer be filled with a blank definition: an answer that carries none is refused and asked again, instead of being stored empty.

Under the hood

Every line of a review batch is now counted by its kind and its language, so the quarter of each batch that comes back empty can be attributed rather than guessed at.
A batch line that fails to parse is measured rather than assumed: three production passes show it costs a retry and never a lost word.
The connection budget written in the repository counts the dictionary's own pool, and a build guard sums the pools across the services and fails past what the database accepts.

1.4.513 September 2026

In the product

Signing up twice — a reload, a back button, a retry after a slow answer — now reuses your billing record instead of creating a second one, which used to leave the account unreadable to every part of the service.
An account can only ever hold one billing record, so a subscription activates and a plan is read correctly even when the signup was retried.
The dictionary's review asks the model for a strictly valid answer, and retries a line it cannot read once before giving up: far fewer words are paid for twice.
A word whose answer arrives unreadable goes back to the queue on its own, and is bought again for the kind that failed rather than for all twelve.

Under the hood

Every line of a review batch is costed under its own word's language, so the ledger can attribute what was spent to English, French or Spanish.
The dictionary dashboard reads its data again, and its cost panels are labelled in the unit they display.
A service that exhausts its heap now raises an alert instead of surviving unnoticed.
Pages are held ten minutes by shared caches in production and one minute in preproduction, and the rule that sets those values is declared in Terraform rather than living only in a dashboard.
`works` no longer carries its Redis wiring, which had cached nothing since the model calls moved elsewhere.

1.4.412 September 2026

In the product

The dictionary's periodic review buys entries again: a full-size batch no longer dies before it reaches the provider.

Under the hood

A review pass that ends without submitting now says so, in the log and in a counter, with a Grafana panel per outcome.
A thread dying on an uncaught error is logged through the application logger, so its stack trace reaches the log store.
The corpus is bought over four passes a day rather than one: the same words a day, a batch four times smaller, and an answer applied within six hours instead of the next day.

1.4.311 September 2026

In the product

The foot count appears as soon as it is known, without waiting for the emotion, and one half failing no longer takes the other with it.
Editing one line no longer clears or shifts the feet and emotions of the other verses.
Indentation and double spaces typed in a verse survive reopening the work.
French and Spanish count metric syllables — e-mute, synalepha — where they counted spoken ones.
English counts accentual feet.

Under the hood

The line analysis is two model calls — one metric skill per language, one emotion skill — each with its own confidence.
The foot count keeps its deliberation: answering without it is four times faster, but miscounts eleven lines in a hundred in French, where everything turns on the e-mute and the diaeresis.

1.4.210 September 2026

In the product

The dictionary drawer gains a ninth section, figures of speech, each labelled with its device.
The links in billing and subscription emails open the account page on the right section.
French assonances come back where they used to come back empty.
The dictionary's review has a budget per language; Spanish is paused.
The dictionary's corpus is seeded from wordfreq, the ten thousand most common words of each language (CC BY-SA 4.0).
The review no longer creates the words an answer happens to name.
A finished review batch is applied once; a word with a failed line goes back to the queue.
Each review batch reaches OpenAI under its own name, with its languages and word counts.
The About dialog names wordfreq and its licence among the open-source notices.

Under the hood

The dictionary's review now buys words with a budget per language rather than one shared between them, so a large English list can no longer spend the French half; Spanish is paused.

1.4.110 September 2026

In the product

The dictionary drawer's sections are now an accordion: one open at a time, starting on definitions, and each header shows how many results it holds — a section with nothing to show stays visible, greyed out, instead of looking like it never loaded.
The sections read in the order definitions, synonyms, antonyms, lexical field, family, rhymes, homophones, expressions.
The dictionary drawer gained two sections: the lexical field of a word — the words its concept draws around it, closest first — and its family — the feminine, the noun, the agent, the diminutive and the other words the language ties to it, each labelled with what it is.

1.4.09 September 2026

In the product

The dictionary now keeps what it learns: a word looked up once answers instantly the next time, and grows richer as it is looked up again.
Rhymes now come back in five categories — rich, exact, poor, assonance, near — each loading on its own, so a slow category no longer holds back the rest.
A misspelled or ambiguous word now gets a "Did you mean" suggestion instead of an empty result.
The dictionary drawer gained two new panels: homophones and expressions.

Under the hood

The dictionary service has its own schema, entities and data store on the shared database, with typed identifiers for a word, an entry and an alternative spelling shared across services.
Cost tracking now records which language a model call served, and common/ai can submit a batch of prompts to OpenAI's Batch endpoint rather than one at a time.
An operator can seed, refresh, re-acquire and count the dictionary's corpus through a dedicated API, and a scheduled job reviews it in preprod and prod.
Grafana now shows the corpus hit ratio and the dictionary's model spend.

1.3.88 September 2026

In the product

Links to the blog's old address still work. Each one now lands on the matching page under ilena.app/blog — the article that was asked for, not a listing to search it out of.
The Resources menu separates what it offers: the guides, then the social pages, then support and the referral programme. Facebook and YouTube were sitting under a heading that read "Guides".
On a documentation page, the search field no longer covers the menu you have just opened.
Article titles all end the same way. Three conventions coexisted, inherited from two origins, and an article whose title already named Ilena said it twice.

Under the hood

The blog no longer has infrastructure of its own: its store, its distribution and its certificate are gone, and its corpus was archived before they went. What a reader asks for is served by the main site, like everything else.
The nightly process that used to write the articles is retired: it was filling a second copy of the blog that nobody read any more.

1.3.77 September 2026

In the product

The blog moved to ilena.app/blog, and now wears the same header, footer and language menu as the rest of the site. Choosing a language from an article takes you to that article in the language you chose, not back to a home page — and a language an article does not have yet leads to the English version rather than disappearing from the menu.
Articles are lighter. Each page used to fetch 407 kB of JavaScript and build its own stylesheet before showing anything; the layout is now plain CSS, so a page arrives already laid out and reads correctly even when scripts fail.
The blog's index no longer lists every article twice. English and French have their own listing, each with its own pages and its own feed, so you scroll through the language you came for.
Every article ends with questions and answers — 444 of them across the blog, where half were previously written but readable by nobody but a human already on the page.

Under the hood

Publishing an article is now one process rather than a sequence somebody had to remember: a subject becomes an article in both languages, illustrated, checked for broken links and missing answers, and refused rather than published when a check fails.
The blog keeps its own audience measurement after the move, so its readers are still counted as its own rather than folded into the main site's.

1.3.66 September 2026

In the product

An article on the blog can no longer appear in one language alone. A translation that failed used to leave the English version behind, and the next run counted the article as done — so it stayed half-published, and nothing said so. The batch now keeps an article only once every language of it is written, and an article left incomplete is repaired by the next ordinary run.
The blog keeps its own record of what it has published, so an article is taken exactly once and an article withheld on purpose stays off the site — including one that was already there.
On a phone, an open submenu no longer covers the menu entries beneath it, and the buttons on the home page keep their words instead of shrinking to unlabelled circles.

Under the hood

The build no longer fails over things that are not defects. Checks that asserted what was true on the day they were written now assert the rule they were meant to protect.
The tooling that refreshes the test environment no longer prints the key it restores. It moves the value without ever reading it, and reports which key is in place by fingerprint rather than by value.

Earlier releases

Everything before 5 September 2026, one line per heading. The full entries are in the repository’s history.

Version	Date	In the product	Under the hood
1.3.5	4 September 2026	Seven writing guides went live in English, French and Spanish, the dictionary stopped offering words that don't actually rhyme, public pages were rewritten to read like they were written by a person, and the 27 solution pages got clean, language-written addresses.	Search-engine signals were fixed at the site's root, and every page a crawler called too thin was documented rather than padded.
1.3.4	4 September 2026	A public FAQ page launched with eighteen answers, every page gained its own heading, description and search-friendly title, addresses were cleaned up across the site, and each page now declares its language alternates.	Structured data, sitemap freshness, `llms.txt` and a registry of pages kept out of search were put in place for search engines and AI assistants.
1.3.3	3 September 2026	The home page says what Ilena is for before it says what runs underneath: the mention of the kind of model has left the top of the page, and the section explaining why Ilena exists opens on Ilena rather than on what other tools do badly.	The search-position measurements moved to a new source before the previous subscription ended, under a single project covering the whole domain; and each service identifies itself to the writing engine with its own credential, one per service and per environment.
1.3.2	2 September 2026	Nine fixes around signing out and returning — a shared eight-hour idle timeout, a fixed account-window sidebar, Settings failing with a message instead of silently, the comparison tables growing to seven competitors, the user guide brought back in step with the product, and shorter French blog titles.	Nighttime alerts stopped firing on empty queues recovering, deliberate probe failures and starting containers; our own visits stopped being counted as traffic; and a timing-flaky test stopped deciding releases.
1.3.1	1 September 2026	The blog began asking for cookie consent before measuring anything, and moved behind the same shield as the rest of the site, answering on one address instead of two.	Reaching preproduction or production became a sign-in that expires within the hour instead of a long-lived key, on an encrypted workstation.
1.3.0	1 September 2026	The blog began publishing new articles on its own every day, each one appearing in both English and French.	Articles arrived from a new writing partner, were rebuilt onto Ilena's own pages and passed a set of checks before publishing, with nothing changing when a check failed.
1.2.3	31 August 2026	Blog articles stopped loading a tracking script from another company: the same page, one fewer request leaving your browser to somewhere that is not us.	The blog moved to Ilena's own storage and delivery network — every article, image and feed is our copy, and nothing on the page reaches the previous provider.
1.2.2	30 August 2026	The account screens became part of the product rather than a page beside it — the editor's menu bar, footer and tool column, a sessions list that names each device, a questionnaire that keeps its text box in view, and a pricing card whose annual saving is a ribbon.	The two web applications began sharing one copy of what talks to the server, and the build that checks every change lost a minute and forty-eight seconds.
1.2.1	29 August 2026	Nine corrections around leaving and returning: Quit no longer signs you out, the site stopped offering Sign In to someone already signed in, a chosen language followed the account across browsers, synonyms came back fuller, and the nine Solutions pages answered the questions a reader arrives with.	A change to how the monitoring dashboards read the database began reaching the running service on the deploy that carried it.
1.2.0	28 August 2026	The Privacy Policy and the GDPR page began saying plainly what reaches Google when you refuse analytics.	What each author's model calls cost became a ledger — per lookup, in the billed currency, surviving the account's deletion — and a reporting dashboard carried the three figures a web analytics tool structurally cannot produce.
1.1.11	28 August 2026	A lookup became three requests running side by side: definitions, synonyms and antonyms appear as each one finishes, under a single loading animation, and a visitor with JavaScript disabled met the three-language menu instead of being sent into English.	Three quiet defects: documentation anchors a crawler can read, a parse failure that says where it failed, and a log shipper that stopped warning at every start about a setting written under a name it had never had.
1.1.10	27 August 2026	A dictionary lookup stopped changing shape partway through, and clicking a multi-word synonym stopped failing.	profiles began retrying its call to invoicing instead of falling an author back to the free tier on a single hiccup, and the dictionary and rhyme skills each gained their own reasoning settings.
1.1.9	27 August 2026		The backend moved to Spring Boot 4.1 on Spring Framework 7 — Jackson 3 throughout — and twelve hand-pinned dependencies were brought up to their latest release.
1.1.8	26 August 2026	Five corrections a reader meets: a phone-sized screen is told Ilena needs a computer, a suspended account is told so in its own language, the footer stopped pushing French and Spanish pages wider than the screen, and a misspelling now caches the correction too.	Twenty changes underneath — a memory ceiling for the cache, two caches removed with the feature that justified them, a deletion that reaches all four services, alerting on a rebuilt preproduction node, /api/ closed to crawlers, and six investigations recorded rather than acted on.
1.1.7	25 August 2026	The site's images, fonts and textures began being kept for a day by your browser instead of re-checked on every page.	Ilena started measuring itself: metrics and timed model calls from the six services, structured logs carrying a correlation id and queryable by field, email alerts on what the cluster can see, traffic measured at the front door, all on a disk that survives a machine being replaced — and machines that boot from a chosen, recorded image.
1.1.6	24 August 2026	Account management moved into an application of its own at /account/ — password, sessions, subscription and billing — reachable from a single header menu that replaced Sign In once you are signed in, returning you to the page you came from, in your light or dark mode, with a reminder email for an unfinished signup and a monthly-to-yearly switch.	Backend tests began building their schema with Liquibase the way production does, the deleted-role guard moved into the module every service depends on, and seventy orphaned preproduction rows were cleaned up.
1.1.5	23 August 2026	A foot count can be set by hand, verse lines fully support emoji through editing and version history, and a work exports as plain text or Markdown beside PDF.
1.1.4	23 August 2026	The product comparison gained LyricStudio, MasterWriter and Lyrcs, and its grid moved with them.
1.1.3	22 August 2026	The Privacy Policy and Your data pages began naming Google Analytics and what it keeps, and a full refund now ends the subscription and freezes the account the way an unpaid invoice does.	The sixteen signup screens, the trial, the payment and the cancellation each became a measurement taken once, from Stripe's own confirmation, with a trial no longer counted as a sale until a card is charged — and staging began billing against Stripe's real test mode.
1.1.2	21 August 2026	An artist can be given Ilena for free and for good: a demo account writes exactly as a Solo one does, with no subscription behind it.	Where a visitor first came from began being kept at signup — and nothing at all when cookies were declined — measurement learned that a phone and a laptop can be one author, every event gained a written definition, and the account type that decides what an author may reach was written once instead of copied into four services.
1.1.1	20 August 2026	A word already looked up comes back faster and stays that way for longer.	A machine AWS takes back now leaves the cluster on its own in ninety-one seconds, reaching production takes a sign-in of its own, the monitoring dashboards moved behind a single door, and prices are named by a stable label rather than by their identifier.
1.1.0	19 August 2026	Updating Ilena stopped interrupting it: every service runs two copies on separate machines, and one keeps answering while the other is replaced.	Preproduction took production's three-machine shape, what every service writes began being kept for fourteen days on storage of its own, and twelve of eighteen libraries were raised.
1.0.16	18 August 2026	Memory use across every service, in both environments, was recorded — the figures the next sizing decision would be made on.
1.0.15	18 August 2026	Five corrections to titles and links: translated pages stopped sharing a title with their category, Solutions pages gained a tab title of their own, every link said where it leads, and the documentation build stopped dirtying the repository.
1.0.14	17 August 2026	Three new pages for people searching for a French rhyming dictionary, a Spanish syllable counter or an English rhyming dictionary.	Shared images moved into one source directory, changelog dates began being filled from the release's Git tag, milestone output moved under .tmp/, and a reported redirect problem turned out to be already fixed.
1.0.13	17 August 2026	Every email was rebuilt on one shared template, an invitation stopped emailing a temporary password, the "invoice paid" email stopped showing the word null where its link belonged, signing in with an unregistered Google account said so, and the dictionary began returning up to ten synonyms.	Connections began requiring TLS 1.2, the servers behind the site started refusing anything not carrying Cloudflare's certificate, preproduction monitoring went from six services to seventeen, and the end-to-end suite halved its running time.
1.0.12	16 August 2026		The site and the API became reachable only through the content network, the public site stopped answering on the host's own address, the DNS and edge configuration moved into code, and working on preproduction stopped requiring production credentials.
1.0.11	15 August 2026	A work's emotions began describing the whole work read in one pass, emptying a line cleared its analysis, and the dictionary settled at ten synonyms and ten antonyms.	Analysing a line stopped paying for two model calls every time.
1.0.10	13 August 2026		Releases moved to GitHub — a tag deploys preproduction, production is a separate decision — production images began replicating instead of being copied by hand, administrative access moved behind a portal with multi-factor authentication, the dashboards' database credential was proven read-only, and four critical vulnerabilities were patched.
1.0.9	12 August 2026	Invoice status began reading in your own language, the Spanish rhyme page stopped describing something it did not offer, the footer's social links gained readable names, and the site began measuring how visitors move through signup and checkout under the cookie choice you make.	A request carrying no credentials began getting the same answer from every service.
1.0.8	11 August 2026	The syllable-counter and Spanish rhyme pages stopped offering anonymous lookup and led into Ilena instead, the English songwriting page was rewritten, and two French guides were added.	Eleven changes underneath: hreflang made consistent, valid structured data on the three homepages, minified shared chrome, a rotated Stripe test key with the command that printed it now refused, per-service log levels, and two skills recorded.
1.0.7	10 August 2026	Opening the editor began requiring an active subscription.	A malformed request began getting a plain 404 or 400, and creating several items at once stopped being able to make two share an identifier.
1.0.6	9 August 2026	The billing address became enterable and correctable, the project tree's icons appeared, re-inviting someone worked again, and an item still being created could no longer be dragged into a broken state.	Six corrections: a mistyped request refused plainly, no unintended default administrator, a deployment that replaces what it claims to, the client address read from the header Cloudflare sets, notification failures no longer swallowed, and a hand-edited generated page failing the build.
1.0.5	9 August 2026	Billing email began arriving for a failed payment, a paid invoice and a cancellation, a declined card stopped being a dead end, invitations reached their recipient, a frozen account could be unfrozen, and the Solutions pages were split into more specific ones.	Stripe's events were processed again without a queue retry creating a second subscription, configured prices began being checked at startup, the site gained a rehearsal environment before production, and continuous integration started publishing to the registry deployments actually use.
1.0.4	6 August 2026	Sign in, verify, forgot and reset password became standalone pages, one sign-in page began serving all three applications, signing in and being admitted became two steps, and every device where you are signed in became visible and endable.	The forgot-password page stopped revealing whether an address has an account, dead role checks and an inert quota mechanism were removed, and the end-to-end suite passed in full against the deployed build.
1.0.3	5 August 2026	The ideas basket began reporting its failures and keeping your draft, four actions on a project or work began checking that it belongs to you, and seven pages took the site's own dark background, menus and header.	Three corrections to the guards and the shared partials.
1.0.2	5 August 2026	The ideas basket began surviving a reload, login and recovery followed your browser's language, two keyboard-shortcut defects were fixed, the header became dropdown menus, and five pages were added — this Change Log, the product comparison, Solutions, Referral and Partner.	Coverage below 80 % began failing the build, and the end-to-end suite gained the ability to run against a deployed build.
1.0.1	4 August 2026	Duplicating a work began producing a real independent copy, the suggestion setting became two, AI-assisted suggestions gained a badge, the line-variants picker was removed, and the three legal texts came from a single source.	AI Transparency was recorded as informational rather than consent-bearing, and the frontend suite began running on every push.