Changelog — 2.25–2.37

Releases 2.25.0 through 2.37.0. Current releases are on the main changelog, and every band is listed in the release archive.

Loading audio...

Version 2.37.0

July 30, 2026

Our voice-over tool was quietly leaving the Arabic out

  • If you gave the tool a passage containing both Arabic and English, it read only the English. The Arabic was dropped without a word — no error, no warning, nothing marked in red. You got back what looked like a finished recording, and the only way to discover the omission was to sit and listen to the whole thing against the original.
  • It now reads the Arabic, written out phonetically, and says so at the start of the recording. This is the same treatment introduced last week elsewhere on the site. The tool had simply never been connected to it.
  • The uncomfortable part is why it went unnoticed for so long. The part of the system that removed the Arabic was not being secretive about it — it marked every affected response to say what it had done. Nothing at our end ever read that mark. The information had been sitting there the whole time with no one listening, so a recording with a passage missing looked exactly like a recording with nothing missing.
  • So the fix is in two halves, and the second matters more. Connecting the tool up properly fixes today’s problem. Reading that mark, and refusing to call a recording finished when it appears, is what means the next one is found by us rather than by you. We would rather fail loudly and make you try again than hand you something quietly incomplete.
  • The spoken note can now be turned off for a single recording, which the team asked for within hours of it appearing. On a twelve-second clip the note was taking nearly eight seconds — most of the recording — and these are short pieces cut into longer videos where the passage is captioned on screen anyway. It is on unless you turn it off, it is not remembered between recordings, and turning it off changes only whether the recording announces the phonetic reading. The passage itself is still read out in full either way. That distinction is the entire reason we were willing to offer the option at all.
  • We also made this particular mistake harder to repeat. There were two ways to ask for audio — a careful one and a direct one — and the direct one skips the phonetic step by design, because some things genuinely do not need it. Four times now, someone has reached for the direct one without meaning to. The build now refuses to accept that, and we added the rule only once the last remaining case was fixed, so it went in working rather than as something to be waved past.
  • Nothing to do at your end. Recordings you made before today are unchanged; if any of them were meant to contain Arabic, they will need making again.

Version 2.36.0

July 30, 2026

We removed the old voice machinery that nothing had used since last week’s change

  • Nothing you can see or hear on the site changed today. The voices are the same voices, and every author still reads in the voice they were assigned.
  • Until last week the site had a second, much older way of speaking non-English text. It was poor at Arabic in particular, so it was avoided in practice. Last week’s change — writing those passages out phonetically before reading them — replaced it properly, which meant the old machinery had nothing left to do. Today it is gone.
  • We proved it was unreachable before deleting it, rather than assuming. There is a check that takes real passages from the site, runs them through the whole pipeline, and confirms no non-Latin text can reach the point where audio is made. We also ran that check against input it was obliged to reject first, so that a pass from it means something.
  • The one thing this could have broken silently, we wrote a test for. Every voice has a name that predates the change, and those names are stored against each author. Had a clean-up renamed them, nothing would have failed — the audio would simply have been generated in the wrong voice. That is now checked automatically on every build.
  • And a small confession about that test. Our own notes said there were 28 voices. There are 29. The first version of the test asserted 28, and failed — which is exactly why it was written to check the voices themselves rather than how many of them there are. Had the number happened to be right, we would have locked in an unverified figure and trusted it forever.
  • Nothing to do at your end.

Version 2.35.2

July 30, 2026

We had written down how many safety checks we run. Every copy of that number was wrong.

  • Nothing you can see on the site changed today. Housekeeping, posted for the same reason as yesterday’s: the work that keeps the site steady is work you are entitled to see, not only the work you can look at.
  • We run a set of small programs whose only job is to prove that things really work — that a page is serving the version we think it is, that a private file stays private, that a change to one feature has not quietly broken another. Our own notes recorded how many of these exist. That figure had been copied into five different documents, and four of them disagreed with each other and with reality.
  • We deleted the number instead of correcting it. Anyone can count the programs in a second by looking; writing the answer down in five places just created five things to keep in step. This is the second time that exact figure has been wrong, which is what persuaded us the number itself was the mistake.
  • More seriously, one of those lists was incomplete without saying so. It named four of the checks under a heading implying it was all of them. Someone working through it would reasonably have believed they had covered everything, and a list that quietly leaves things out is worse than no list, because it gets trusted. It now names every one, with what each proves and when to run it.
  • The steps for publishing a new version now live on one page, in order. Yesterday we wrote down how our release process works; today there is a plain checklist to work down while doing it, with each step linking to the fuller explanation if something refuses. Written deliberately to hold no facts of its own, so that it cannot start contradicting the documents it points at — which is precisely the failure described above.
  • Nothing to do at your end.

Version 2.35.1

July 30, 2026

We had a careful process for publishing updates, and had never written it down

  • Nothing you can see on the site changed today. This is housekeeping, and it is here because you are entitled to see the work that keeps things steady, not only the work you can look at.
  • One readability fix, on a page only our editors see. Two links in a coloured panel were rendering in a shade that was very hard to read against the background. Small, but the people preparing the site’s audio are users of it too.
  • The steps for publishing a new version are now documented. There is a script that refuses to publish if anything is out of order — unsaved work, the wrong branch, version numbers that disagree with each other. It has been doing that quietly for a while, and until now the only person who knew how it worked was the person who wrote it.
  • We also wrote down a mistake that has not happened yet. There is a plausible-looking change to our publishing setup that would appear to succeed while quietly leaving the site running the previous version. It is now recorded, with the reason, so that whoever makes that change knows to avoid it. Catching something before it bites is cheaper than explaining it afterwards.
  • Nothing to do at your end.

Version 2.35.0

July 29, 2026

Arabic in a narrated page was either mangled or quietly removed. Both are fixed.

  • Arabic, Farsi and Hebrew are now spoken properly in narrated pages and articles. Where a page quotes a passage in one of these, we write it out phonetically in English letters first, and the narration reads that — so you hear the sounds of the original rather than a translation of it. It makes no difference what the passage is: a verse, a narration, a line from a scholar or a quotation from another book are all treated the same way.
  • What it used to do instead was worse than we had described to ourselves. One setting produced sounds that were not words at all, with stray capital letters inserted in the middle of them. The other setting removed the passage — so “He said:” was followed by the translation and the verse itself simply vanished. That version was the harder problem, because it sounded completely fluent. Nothing told you something was missing.
  • The recording now tells you when this is happening, and so does the page. You will hear it in the first few seconds, and a notice stays visible for as long as you are on the page. Both, deliberately: if you are picking up where you left off, playback resumes from your saved position and skips the beginning — so the spoken warning is one you might never hear.
  • Please do not learn a verse from the narration. Vowels are often unwritten in the original, and supplying them is a matter of grammar and meaning rather than a mechanical rule. It can be wrong. The written text on the page is the one to trust, and that is what the notice says.
  • A correction only has to be made once. Every phonetic rendering is kept, so when someone qualified corrects a passage, it is corrected everywhere that passage appears — across every page and article, now and in future. Not all of them have been checked by a person yet, which is exactly why the notice is there.
  • Pasting purely Arabic text into our audio tool used to fail outright. It now works. The step meant to handle Arabic was removing it before anything could be read.
  • Whoever prepares the audio can now check the pronunciation first. Rather than generating a recording and correcting it afterwards, the passage is shown alongside the original before anything is made — the English on the left, the Arabic on the right where it naturally reads from. Anyone editing has the two side by side, in the direction each is meant to be read.
  • Nothing to do at your end. Pages written entirely in English are unchanged, and no existing recording has been altered — the change applies the next time a page’s audio is made.

Version 2.34.0

July 29, 2026

Nothing you can see changed this time — here is what did

  • The settings that keep the site running existed in only one place. If the machine holding them had been lost, rebuilding the site would have meant reconstructing them from memory — and the result would have looked healthy while being quietly wrong. They are now recorded properly, with an automatic check that notices if the record and reality ever drift apart.
  • We found settings that looked like they were configuring things, and were not. Nothing is broken by it: the values they would have applied happen to already be correct. But anyone changing one would have seen no effect and no error, which is exactly how a confusing outage gets made months later.
  • Our build and test process got about a quarter faster. Measuring it first was worth doing, because the slow part was not the part anyone assumed: the tests themselves finish in five seconds, while fetching the things they depend on took two and a half minutes every single time.
  • Every background service now documents how it is set up. Two of them had no documentation at all. That matters for the people maintaining the site, which in the end is what decides how reliably it works for you.
  • We designed how planned maintenance will work. When we next need to take something offline for updates, you will see notice on the site beforehand and a proper “back shortly” page rather than an error.
  • You will see more entries like this one from now on. Until today we only wrote up changes you could see on screen, so the work underneath went unrecorded. That was the wrong way round: most of what keeps a site trustworthy is invisible, and you are entitled to know it is happening.
  • Nothing to do at your end.

Version 2.33.0

July 29, 2026

Buttons turned green when you pointed at them — and we had already “fixed” this once, in one place

  • Coloured buttons became hard to read the moment you hovered over them. A button with white writing on a solid colour would have its label turn dark green as your pointer crossed it — sometimes almost invisible against the button behind it. It affected links styled as buttons, all over the site, in both light and dark mode.
  • We had met this before and fixed only the two buttons in front of us. Earlier this month it was reported for the Zoom and Google Meet buttons on event pages, and those two were corrected individually the same day. The thing actually causing it was left in place, so every other button kept the fault and every button written since inherited it. This time the cause is fixed rather than the symptom, which means buttons added in future are right by default.
  • The cause was a site-wide rule about link colour being applied too forcefully. It was only ever meant as a fallback for links that do not choose their own colour — ordinary links in a paragraph. It was instead overruling buttons that had explicitly chosen one. It now yields to anything that states its own colour, and still colours everything else exactly as before.
  • We checked the obvious over-correction, because it was the real risk. Ordinary links in text still change colour when you point at them — that was measured on the page, not assumed, alongside the buttons that were broken. Of every link on the site, exactly one relied on that colour change as its only response to being hovered, and it was given its own.
  • The “Related topics” boxes at the foot of each help article were close to unreadable in dark mode. Dark green writing on a dark grey background, well below the contrast any text should meet. Now plainly legible, and they respond properly when you point at them.
  • The “Contact Us” button in the Help Centre now opens our contact form. It used to try to open your email application, which does nothing at all on many phones and for anyone using webmail — with no message to say so. The form works for everyone.
  • Nothing to do at your end.

Version 2.32.0

July 28, 2026

Three of our publication logos disappeared from the home page, and we did not spot it first

  • Someone on the team noticed before we did. Three of the eight logos on the home page had quietly stopped appearing. Nothing had broken loudly — the space where each logo belongs was still there, so it looked deliberate rather than wrong. That is the part worth admitting: the failure was designed to be invisible, and it did its job.
  • The cause was a list of our publications written into the code. It repeated things the system already knew, and the two copies had to be kept in step by hand. Change a detail in one place and the other silently disagreed. The list has been deleted rather than corrected — there is now one record of each publication and nothing left to fall out of step with it.
  • This had happened once before, to the same list. A logo was updated and the home page kept showing the old one. That was fixed at the time, but only for the one field that had gone wrong, and the same fault was waiting in the field next to it.
  • Each publication can now be checked on demand. Behind the scenes we can ask each one to confirm its own identity, and see plainly whether the answer was right, wrong, or could not be obtained at all — along with when we last asked. Previously there was nothing to look at.
  • Nothing to do at your end. All eight logos are back, and the home page shows publications in the same order it always did.

Version 2.31.0

July 28, 2026

We publish eight titles. Our own tools could not reliably tell them apart

  • This one is entirely behind the scenes, and it is a mistake worth owning. When we bring a study series into the reading room, it has to be filed under the publication it came from. That link was never actually enforced — it was a piece of text that happened to match, and nothing anywhere would have complained if it stopped matching. It is now a proper connection that the system refuses to let us get wrong.
  • The consequence would have been quiet, which is the bad kind. A series filed under the wrong title would have gone looking for new instalments on the wrong site, and found none — reporting success each time. Nothing would have looked broken; the series would simply have stopped growing.
  • And we could not correct a mistake once made. There was no way to move a series to the right publication after the fact, short of deleting it and everything in it. There is now.
  • Fetching recordings was working out which site to ask by guesswork. It guessed from the name, which was right for the publication it was originally written for. We checked all eight properly for the first time: it was wrong for two of them, and — the part that stung — the one title we had flagged as the likely problem turned out to be fine. We had been about to fix the wrong thing. It now asks each publication directly, which is both correct and quicker.
  • And the form was telling us to do the wrong thing. The box where we enter a publication’s address had an example in it suggesting the wrong sort of address entirely — so the tool was quietly coaching whoever used it into the mistake. It now offers a tick-box that copies the address already entered, which is right for every publication we run, with a note explaining the rare case where it is not.
  • Nothing to do at your end. Nothing you can see has changed. This is the sort of thing that costs nobody anything today and a great deal in six months, so it is being paid for now.

Version 2.30.0

July 28, 2026

Making the spoken version of a page was quietly broken, and had been for a while

  • Most pages on this site have a spoken version you can play. Producing one is our job, not yours, but it is worth saying what was wrong: the tool we use to make them would appear to stall a few seconds in, then often fail outright with a connection error — after which the recording it had just spent ten minutes making was thrown away. This page, being the longest one we have, was the worst affected.
  • It failed for an embarrassing reason. The recording was being made inside the browser window that started it. Close the tab, let the laptop sleep, or lose signal for a moment, and the work died with it — even though the machine doing the actual reading carried on regardless, unaware anyone had left.
  • It now runs on the server, properly. Starting a recording hands it over and lets go. You can close the page and come back later; it will still be going, and it will tell you where it has got to. Nothing is lost to a dropped connection any more.
  • And it says what it is doing. A long page is read in a few hundred short pieces. Previously the progress bar sat at 20% for the entire job with no way to tell a slow recording from a dead one. It now counts through them.
  • Nothing to do at your end. Nothing about playing audio on the site has changed — this was all on our side of the fence. The practical difference for you is that the spoken version of a page is now far more likely to be up to date with the page you are reading.

Version 2.29.0

July 28, 2026

Finished programmes stopped pretending they hadn’t started

  • A finished series was saying “Coming soon”. On the Events page, the Ashura series — which has been and gone — announced itself as forthcoming and listed no sessions at all. It was our mistake and it read as though we had simply never held the programme. It now shows its session count like every other event.
  • The home page kept replaying the last night of a finished series. Long after a programme ended, the home page still led with its final session, laid out exactly like the thing to watch tonight. A finished programme now goes back to showing what it actually is — the trailer, the description and the artwork — with a line saying it has finished broadcasting and that every session is still there to watch.
  • It waits a day before doing that. The final night of a series stays on the home page for twenty-four hours after it airs, so if you missed it you will still find it the next morning.
  • The trailer now sits at the top of the card. On a programme with a long introduction you previously had to scroll past all of it to reach the video. The artwork moved below the text, and both now appear together rather than one replacing the other.
  • Nothing to do at your end. Nothing you had bookmarked or shared has changed address — every past programme is where it was, and still complete.

Version 2.28.0

July 28, 2026

We are on Facebook now

  • You can find us on Facebook. It is a new page, so there is not a great deal on it yet, but it is there and it is genuinely ours. The link sits with the other icons at the bottom of every page and on the Community & Channels page.
  • That is the fourth place we have tidied this week. Between yesterday and today our Instagram, X, Threads, Linktree and now Facebook all point where they should — and search engines have been told which accounts are actually ours, which is what stops the wrong one appearing next to our name.
  • Nothing to do at your end. If you use Facebook and would like to follow along there, the icon is at the bottom of this page.

Version 2.27.0

July 27, 2026

Our social links were pointing at the wrong accounts

  • Our Instagram has moved to @truthpromoters313. The account this site had been sending you to is no longer ours. If you have been following it and wondering why it went quiet, that is why — and the new account is the only Instagram we post to. It would be good to see you there.
  • We are now @truthpromoters on X, and @truthpromoters313 on Threads. Both of those links previously went to a personal account belonging to one of us, which was never quite right for anyone looking for the group rather than the person.
  • Our Linktree address has changed too, and the old one had stopped working. It is now at linktr.ee/truthpromoters. If you tried the link on our Community page recently and got a “page not found”, that was this, and it is fixed. It has also been added to the row of icons at the bottom of every page, where it should have been all along — it is the one place that gathers all of our channels together.
  • Every place we list these has been brought into line — the icons in the footer, the Community & Channels page, and the information search engines read to work out which accounts are genuinely ours. That last one matters more than it sounds: it is how a search result decides which Instagram to show beside our name, and it had been vouching for an account we do not hold.
  • Nothing to do at your end, unless you would like to follow the new Instagram — the old one will not be carrying our posts.

Version 2.26.0

July 25, 2026

A guide to listening, and a tidier Help Centre

  • There is a new guide on listening to the site. Quite a lot of what we publish can be played rather than read, and until now nothing explained how any of it worked. The new page covers the difference between a narrated page and a real recording, how to build up a queue, and what the player remembers — including the bit people ask about most, which is that it brings you back to exactly where you stopped, and always brings you back paused.
  • Why it comes back paused. Browsers do not allow a page to start playing sound on its own, and we would not want it to either — a tab opened in the background suddenly talking is nobody’s idea of helpful. So your place is kept, and you press play.
  • The guides for IlmFlow, FiqhFlow, ScribeFlow and MediaFlow have moved. Those are their own sites now, each with its own documentation, and keeping a second copy here only meant keeping a copy that was always a little out of date. The old addresses take you straight to the right site, so any link you have saved still works.
  • The remaining guides have been rewritten. Prayer times, the Hijri calendar, notifications and getting started all say the same things they did, laid out more clearly — the calculation methods and the two scholarly positions on moon sighting are now proper tables rather than boxes you had to read across.
  • Three help pages could never be listened to, and we had not noticed. The list our tools use to produce the spoken version had fallen out of step with the list of pages that actually exist. The two are now the same list, so this cannot drift again.
  • Nothing to do at your end.

Version 2.25.2

July 25, 2026

Housekeeping: faults fixed before they could reach you

  • Nothing here was a bug you could have run into. We turned a code-checking tool back on after it had quietly stopped working, and it flagged a set of places where the code was built in a way that would misbehave in the right circumstances — circumstances that, as it turned out, the site does not currently produce. We have fixed them anyway, because “it cannot happen today” is a weaker guarantee than “it cannot happen”.
  • The clearest example. The listen button on an article decided which recording to offer once, when the page first appeared, and never reconsidered. Move between two articles that both have audio without a full page reload and it would have offered the wrong one. It needs two such articles next to each other to show up, and there are not two yet — so nobody met it. It now works out the answer fresh every time.
  • Tapping a track in the queue is genuinely a little better. Working out whether you tapped or swiped used the clock at the moment we got round to checking rather than the moment your finger actually landed. On a busy page that could read a tap as a swipe and do nothing at all. It now uses the real timing of your touch, so this one you might actually notice.
  • The rest is tidying with no visible effect — the scripture reader panel, the contents sidebar on long pages, the home page feeds, the transcript player, and the part of the audio player that remembers your place, all restructured so a future change cannot turn them into real faults. The last of those we checked by hand afterwards: stop a talk part-way, reload, press play, and it carries on from where you were.
  • Nothing to do at your end.

Version 2.25.1

July 24, 2026

Old shared links work again

  • Links shared before November 2025 had been leading to a “page not found”. This site used to be published elsewhere, and when it moved here the writing moved to Reflections 313. Anything already shared — in a message, an email, a forum post — kept pointing at an address that no longer had anything behind it. If you ever followed a link to one of our older pieces and found nothing there, that was why, and it was our fault rather than yours.
  • Those links now take you to the piece itself, wherever it now lives. Nothing needs re-sharing and no old link needs replacing — they simply work again.
  • Our short links now always open the main site. A short link could previously leave you browsing the rest of the site under the short web address instead of our proper one. It now hands you straight to the full site, on the page you were headed for.
  • Nothing to do at your end.

Version 2.25.0

July 24, 2026

The player now remembers where you got to

  • Refreshing the page used to wipe out what you were listening to. Moving between pages was fine — the player carried on — but a refresh, or following a link that reloaded the site, cleared the track, the queue and your position, all at once. On a forty-minute talk that meant starting again from the beginning, with no way back to where you were.
  • Your place is now kept. Come back to the site — in an hour or next week — and the player returns with the same recording, the same queue, and the same position you left it at.
  • It comes back paused, on purpose. Nothing starts making noise because you opened a page. Press play and it carries on from exactly where you stopped.
  • It is kept on your own device and nowhere else. We do not store what you listen to on our servers — this is your browser remembering, in the same way it already remembered your volume setting. Clearing your browser data clears it, and emptying the queue clears it straight away.
  • Nothing to do at your end. It applies to everything the player handles: article and page narration, and the recordings in our libraries.
← All releases