Children began using the new Reading Corner

✅ NEW, 11 Aug — the empty space below the side column is not a bug

Measured on the live page at 1280, and it is exactly what it looks like: the story column is 893px tall, the side column is 441px, and the empty space below the side column is 452px — precisely the difference between them.

⭐ In plain terms: it is two columns holding different amounts of writing. The story runs several paragraphs. The host card and one quotation run out after about half of it. Below that point there is simply nothing left to put beside the story — the same way one newspaper column finishes before the column next to it.

⛔ Checked for the things that would have made it a defect, and none of them is present: the side column is position:static (nothing pinning it), min-height:auto (nothing holding the box open), margin-bottom:0 (no stray spacing), and no align-self stretching it. Nothing is creating the space; the space is what is left when the content stops.

⚠️ The real option, and I recommend against it. The card could be made sticky so it follows the reader down the story. ⭐ Airbnb does exactly this — measured on the live listing: position:sticky; top:80px, and scrolling to y=1400 moved the rail's viewport position from 580 to 112, so it holds near the top instead of scrolling away.

⛔ But their rail is a BOOKING FORM — a control you act on at any point while reading. Ours is a credit: one fact you read once. Following the reader for 452px to keep saying “Olimalar Foundation ran this” is noise, not utility. ⚠️ It would also do nothing on a phone, where the column is stacked under the story anyway — so it is a desktop-only affordance, on a site that is 62% desktop, buying repetition rather than access.

⭐ The honest fix for the feeling, if the gap still bothers him, is more to say — the other two testimonials. The column is short because it holds one quotation. It was built to stack any number.

⭐⭐ NEW, 11 Aug — the identity unit centres, the prose does not

Owner: the avatar and name are centred on the real card and ours were left-aligned. He is right — and centring the whole card would have been the wrong fix. Measured on Airbnb's 395×265 card (padding: 32px 24px), it is two columns:

• Left column 227×201, display:flex, align-items:center, justify-content:center. The 104×104 avatar sits at x=86 with 62px of column to its left and 61px to its right — centred in ITS COLUMN, not in the card, which still has 206px to its right. The “Superhost” badge under it computes text-align:center.

• Right column at x=275 — the stats — and they compute text-align:start. Left, not centred.

⛔⛔ And every piece of prose is left-aligned AND lives outside the card entirely — “Speaks English and Hindi”, “My work: …”, “Response rate: 100%”, “Responds within an hour”, all text-align:start, all outside the 395×265 box.

⭐ So the rule on Airbnb's page is narrow: the identity unit — face, name, badge — centres, and anything you READ stays left. That is what the measurement says, and it stands as a fact about their card.

⛔⛔ OWNER OVERRULED IT THE SAME EVENING, and the card now centres as one column. Five changes marked directly on it: avatar centred · name centred AND stepped down to the link's size · “Ran this session” struck out entirely · description centred · link centred. Built as marked.

⭐ And it is a defensible departure rather than a mistake, for a reason the measurement itself supplies: Airbnb's card is a booking-adjacent block with a stats column, and that second column is what makes their left-aligned identity unit sit off to one side. Ours has no stats column — the count was dropped this morning — so a single centred stack is the honest shape for what is actually left in the box. A two-column rule copied onto a one-column card would be cargo. ⚠️ We also keep the description inside the card where Airbnb puts it outside, because a 344px sidebar has nowhere else for it.

⚠️ The cost, stated once and not relitigated: centred prose is harder to read than left-aligned — both edges rag, so the eye loses the line start on every return sweep, and the description runs three to four lines at 41–49 characters. Raised before the ruling, overruled, built as marked.

⛔ .hostcard__id and .hostcard__role are deleted, not disabled. The wrapper's only job was to centre part of a card that now centres whole; the role line is struck from the design. ⚠️ That also retires the specificity fix made to .hostcard__role an hour earlier — it is recorded rather than quietly dropped, because the same collision will bite the next single-class rule added to this card.

⛔ A specificity bug found in the same pass, in my own component: .hostcard__role was a bare class at (0,1,0) and lost to .hostcard p at (0,1,1) declared below it, so the 8px gap it asks for was rendering as 12px. Fourth single-class override in this stylesheet beaten by a descendant selector — found by reading the file because it had happened three times, not by looking at the page.

⛔ NEW, 11 Aug — the nav gap was pre-existing, and the sidebar only exposed it

Owner: “no padding between the nav menu and the content.” Measured against the nav's bottom edge on the live page: this page's h1 ink began at −1px, and the host card's top border at exactly 0 — a line drawn hard against the nav's own border.

⭐ But it did not start with the side column. Every other page opens with .lede, which carries padding-block: var(--space-8) and puts its first ink at 69px. /clubs/ and /who-we-are/ both measure 69. This page has no .lede and never had the gap — the entry opened straight onto its headline from the day it was built, and four reviews did not catch it.

⭐⭐ What the sidebar changed was visibility, not the defect. Bare text at −1px reads as tight; a bordered box at 0 reads as broken. ⚠️ Worth keeping as a pattern: adding a bordered element is a cheap way to discover the spacing bugs a page already had, because a border makes a boundary that text only implies.

⚠️ Fixed on .entrygrid, not page-wide, and the scope is deliberate: the other four pages get their 40px by accident of carrying a .lede — spacing riding on a semantic class. The durable fix is a rule that says content does not start flush against the nav, which would re-space four already-reviewed pages and is not this brief.

⛔⛔ NEW, 11 Aug — the headline says Reading Corner and the photograph shows the science class

Look at the screen in the picture. It reads “SCIENCE CLUB”. There are no books and no reading corner in the frame. The body says “The science class continued with hands-on learning, while children also began using the new Reading Corner” — the headline takes the second clause and the only photograph documents the first.

⭐⭐ This is the exact failure this kit already predicted, and it has already shipped. Note 1 below says “a generated headline can be accurate and still pick the wrong one of the day's two events.” It did. Every word of the headline is supported by the body, so the guardrail “nothing in a headline that is not in the body” passes on this entry.

⛔ So the guardrail is not enough, and the fix is one more sentence in the brief: where a day carries two events, the headline takes the one the photograph shows — because on this page the photograph is the largest thing on the screen and it is read before the words. ⚠️ Nothing automated can check this; it needs the human spot-check that is already planned, told what to look for.

⚠️ Second-order, and it dates the picture: Science Club was renamed Curiosity Club on 10 Aug. The photograph now shows a screen carrying a retired name. ⛔ Leave it. It is a record of 4 July and it was called Science Club on 4 July — a record that is edited to match today's vocabulary stops being a record.

⛔⛔ 0. AUDITED 10 Aug — this page opened on a blurry cat and no words

Measured on the served page, then looked at. The photograph was aspect-ratio: 4/3 + object-fit: cover at container width, which rendered 992×744. On a 1280×871 laptop the entire first screen was one image — the headline began at y=947, below the fold — and because the source is a 960×1280 portrait, the crop threw away most of the frame and scaled what was left past its own pixel width. It was soft.

⛔ This is the arrival page for 60.8% direct traffic. Someone opening a WhatsApp link saw a cropped, blurry cat and not one word — no headline, no date, no place. ⚠️ Every geometric check passed: zero overflow at 390 and 1280, prose on --measure, no border anywhere. Only the screenshot found it.

⭐⭐ Airbnb, measured live the same hour (listing 23572181, 1710px viewport): the title sits ABOVE the gallery — TITLE at y=96, HERO at y=150 — and the whole hero is 483px tall in a 1120px column, a mosaic of 560×459 plus four ~272×225 tiles. They never let one image be 4:3 at full column width.

⭐ So the defect was never the photo's WIDTH. Ours was 992 against 544 of prose, a 1.8× ratio; Airbnb runs 1120 against 653, which is the same. It was the HEIGHT. ⛔ Fixed by bounding height and removing the forced ratio — max-height: 32rem, width: auto, no crop. A portrait now renders 384×512, a landscape a wide band, and neither is ever scaled past its intrinsic width.

⭐⭐ And the reason it is a crop rather than a fit is a product argument, not a taste one: in a listing the photograph SELLS, in a record it is EVIDENCE. Cropping a phone snapshot to a tidy landscape strip removes the thing it is evidence of. The card in the feed may crop — it is an index. This page must not.

⚠️ Two smaller faults found in the same pass, both fixed: the width/height attributes said 1600×1200 for a 960×1280 image — so the reserved box had the wrong shape and the browser used the wrong ratio — and the date read “Thursday 7 August” when 7 August 2026 is a Friday. ⛔ For the importer: take both dimensions from Ghost's own image metadata per image, never a constant, and derive the weekday from the date.

⚠️ Deliberate trade, stated: a portrait now occupies 384px of a 992px column, left-aligned with the headline. One left edge is the house rule and it reads correctly, but it is quieter than a full-width hero. ⭐ The alternative — Airbnb's multi-photo mosaic — needs several photographs per entry, and the corpus has one.

⭐ RE-MEASURED 11 Aug on the new exemplar, which is a landscape. The 1143×827 source renders 708×512 at 1280 — height capped at 32rem, never scaled past its own 1143px — against 544px of prose, a 1.30× ratio where Airbnb runs 1.71×. Headline, date and credit all sit above it, so the first screen carries words. ⚠️ The cat numbers above are kept as the record of the audit that produced the rule; they describe the previous exemplar, not this one.

⛔⛔ NEW, 11 Aug — the quotation column was 20px text in a 16px column, on three pages

.says capped at 22rem because Airbnb's review column measures 450px. But Airbnb's review text is 16px and ours is 20px — ⭐ the invariant is the character count, not the pixel width. Counted with a Range on the rendered page: 36 characters on this entry and 38–41 on /clubs/, against Airbnb's ~56 and this system's own 45–75 band. Under the ~35 floor where a paragraph reads as confetti.

⛔ Where it bit hardest: /clubs/'s two student quotations — “At first, I didn't know how to stitch. Today, I made this bag on my own.” — the strongest sentences on this site, set in the narrowest column on it.

⭐ Fixed by one token, --measure-small → --measure. The same three quotations now count 53, 61 and 62. Zero overflow at 390 and 1280 on both pages, and /clubs/ was looked at, not only measured.

⚠️ Beyond this brief, and stated because it changes a page nobody asked me to touch: on /clubs/ the quotation is now wider (544px, 20px) than the club's own description (352px, 16px). ⭐ That is the right way round — the description is secondary metadata and the quotation is evidence — but it is a visible change to an already-reviewed page.

⭐ 1. Headlines: Claude backfills, the writer takes over

Owner ruled Claude writes the backfill headlines, with three guardrails: nothing in a headline that is not in the body; a twenty-entry human spot-check read consecutively rather than at random; and ⭐ backfill only — going forward, the person who writes the entry writes the headline.

⭐ That last one is the important one and it is not about quality. August 7 carried two things — a stitching session and the cats. Only somebody who was there knows which of them mattered. A generated headline can be accurate and still pick the wrong one of the day's two events.

⚠️ Consecutively, not at random, is the right check — errors in generated headlines cluster by pattern, and a random sample of twenty across six months would show one of each while a consecutive twenty shows the pattern repeating.

⭐ 2. Three image states, and only a tag can tell two of them apart

Owner ruled: a placeholder image and a needs-photo tag on every entry waiting for a photograph.

state	image	tag
photo pending	placeholder	needs-photo
photo added	the real photograph, which becomes og:image	tag removed
⭐ genuinely no photo, ever	none	none

⭐⭐ The third row is why the tag exists rather than the image. “Waiting for a photograph” and “there will never be a photograph” are different states, and the image cannot distinguish them — only the tag can. ⛔ The 2 August entry above stays image-less and untagged: a placeholder there would promise a photograph that is never coming.

⛔ A placeholder is a PRE-LAUNCH state that cannot survive launch by construction — see the build gate below. Nobody should design this page around the placeholder looking good, because on the day it goes live none exists.

⚠️ I recommended against the placeholder and was overruled, and the override was right. My objection was that a placeholder becomes og:image, so every shared link renders a grey “coming soon” card in a WhatsApp thread — a bad image card in place of a clean text one, on the 60.8% who arrive that way. ⭐ The owner's answer was operational: his team cannot add photographs to entries they cannot find, and the tag is how they find them. He knows how his team works and I do not. ⭐ The build gate is what makes “it will be changed before live” true rather than assumed — so the objection is answered by a mechanism instead of by trust.

⛔ 3. The launch gate — the build fails while any placeholder remains

The placeholder is one known URL, so it is a count. ⭐ The failure is never that people are careless — it is that nothing tells them. If three entries in 150 keep the placeholder, nobody notices by eye and the check notices every time.

⛔ Break it on purpose before trusting it: point one entry at the placeholder, watch the build go red, revert, watch it go green. A gate that has never failed is not known to work — and this project has produced two gates today that could not fail at all.

⚠️ Count the needs-photo tag as well as the image URL. They can disagree — an entry given a real photograph with the tag left on, or tagged with the placeholder already replaced — and a gate that checks one of two fields passes on the state where they diverge.

⛔ 4. The bodies may be rewritten — and one thing inside them may not

Owner, 10 Aug 2026: “the reader is important. not the author.” So the backfill may rewrite bodies for clarity, not only supply headlines. ⭐ The order still stands: headline first, rewrite only where the headline cannot carry it — a headline that supplies a missing referent is less invasive than a rewrite. ⛔ Do not rewrite 150 bodies because you now may.

⛔ Rewriting for clarity is not inventing. “The same group returned” → “The dance group returned” supplies a referent that exists elsewhere in the corpus. Adding a detail that appears nowhere is a different act and is still forbidden. Every fact in a rewritten body must be traceable to its source month.

⛔⛔ And the line the ruling does not reach: a quotation is not the narrator. “The reader over the author” is about the person narrating the day — not about the people quoted inside it. “I never thought I'd be able to dance at this age” is evidence because it is somebody's actual words. Tidied, it becomes a claim the organisation makes about itself. ⭐ A rewrite pass that treats a quoted sentence as prose would destroy the most valuable content on this site while looking like an improvement. Detect quotation before rewriting; leave it exactly as found.

⚠️ One property, recorded so it is not drifted into: with ~150 rewritten bodies the archive becomes a clear record of what happened rather than a record in the words of the people who were there. Both are legitimate and he has chosen the first. ⭐ It applies only to the backfill — going forward the team writes their own — so the two halves of the archive will differ in voice, and someone reading June beside next June should know that is a boundary rather than a drop in quality.

⚠️ 5. What this page cannot answer

Whether the slug is stable. If a headline is edited after publication and the slug follows it, every link already shared in a WhatsApp thread breaks — and those are exactly the 60.8% this page is designed for. ⛔ The slug must be frozen at publish and never track a later headline edit. That is an importer and Ghost behaviour question, not a design one, and it belongs beside the redirect discipline in ASTRO-PLAN.md.