site: reviews — Isopolis, ssh.place (graduate from finds)

daily-finds.log.md

@@ -4939,3 +4939,38 @@ The theme is the honest through-line and it is a narrow one: **none of these fou
 `sixwhyo` is in the bundle although it was not shortlisted. That is deliberate and worth recording as a small precedent: it was held off the shortlist because its delight is a one-line joke rather than a use, and this is the one theme where that is the qualification rather than the objection. Jokes-proper (`joke-cashew`) still get no bundle slot, per the 08-02 note.
 
 **Hero floor met at the minimum, and the ratio is better than it looks:** 2 of 4 items carry images, against a floor of 2 and an aim of ~4. Two heroes in a four-item grid is half the tiles, where the skill's warning case is two in an eight-item grid. It was also the only ratio available — the entire day produced exactly two hero-bearing finds, so no swap could have improved it.
+
+## 2026-08-03 — reviews: isopolis, ssh-place
+
+Two graduations from three review candidates. **Both fact-checks paid, and one of them corrected the find rather than the source** — the same pattern as 08-02, and for the same reason: today's candidates publish their own technical writeups, so the primary document is available and it disagrees with the summary.
+
+### review: isopolis
+
+Graduated from find captured on 2026-08-03, source: Hacker News — Show HN.
+Hero: `img/pofa.webp` from the project's own behind-the-scenes report — the Palace of Fine Arts rotunda and lagoon with the Marina grid running down to the bay and the Presidio tree canopy filling the lower half, 1536×1536. **A different asset from the find's hero**, which is the cropped postcard og image at 832×300. The dev report publishes 24 full-size stills and this one was picked over the obvious alternative (`img/bridge.webp`, 2048×2048, the Golden Gate) because the bridge frame is two-thirds flat blue water — it is the better *story* and the worse *tile*. `img/viewer.webp` (1600×883, the app in situ) was rejected as a UI screenshot rather than the artwork.
+Category: design (existing).
+Fact-check: **the find got the data source wrong and the review corrects it in both places.**
+- **The city is not OpenStreetMap.** The find asserted "OpenStreetMap footprints, a CARTO basemap, neighbourhood boundaries from DataSF" as the geometry, reading the page footer as a provenance statement. The author's own report (`sf.isopolis.city/dev.html`) is explicit: the base is **Google Photorealistic 3D Tiles**, scraped and rendered in three.js through a single orthographic camera for the entire city — azimuth 22.5°, elevation 30°, 440 blocks of 1024² frames — and repainted by a LoRA fine-tune of Qwen Image Edit. The three footer credits are real but they are the *viewer's*: CARTO supplies the dark basemap warped underneath at 45% alpha, DataSF the neighbourhood polygons, OSM the tour paths. **The correction was pushed back into the find**, same call as the `elevators` 1961-patent demotion on 08-02.
+- Verified and used: ~121 km², roughly a 22-gigapixel image; the published set is 2,630 quadrants (1,606 generated, 1,024 water stamps); the tile pyramid is 3,570 tiles across nine levels at 287 MB.
+- **The two failure stories are the best material in the piece and both are the author's own.** The Golden Gate was so far out of distribution that the model hallucinated extra structure every attempt, and roughly half the span was eventually generated by a different model unconnected to the city and stitched in. Open water failed the other way — shadows and flat green fields kept resolving into lakes — and the fix was to stop generating water entirely and stamp one tile through a hand-drawn mask, so every flat blue quadrant on the map is literally the same tile.
+- **Limitation carried into the review because the author states it:** scraping Google's 3D Tiles is against Google's terms of service. Also carried: seams survive, with two pointed at directly. Principle 5 — a review that omitted the TOS line while quoting the pipeline would be selling the glossy version.
+- Not asserted: the og card's "1.6 M/PX" footer was read as a scale indicator, not a pixel count, and is left out rather than guessed at.
+Amazon: n-a (a free web page).
+
+### review: ssh-place
+
+Graduated from find captured on 2026-08-03, source: Hacker News — Show HN.
+Hero: a **second, later fetch of `ssh.place/canvas.png`** — 1400×780, the live canvas roughly forty minutes after the snapshot the find carries. This is the only hero on the site whose subject changes between two pages by design, and the two frames genuinely differ: the review's has a large "is unsafe" block that did not exist when the find was captured. Both were looked at before committing.
+Category: terminal (existing, previously used twice).
+Fact-check: **the find held up in full and the README carried the better framing.**
+- Verified first-party from the site, the stats page and the repository: 200×60 = 12,000 cells, exactly as the page's own "5000 of 12000" counter implies; fifteen-second cooldown tied to the **SSH key fingerprint** rather than the session, so reconnecting does not reset it; sixteen colours; Go, MIT, `github.com/jeninh/ssh.place`, first committed 2026-07-31. The canvas survives server restarts although the placement counter does not.
+- **The material addition the find had no room for, and it is what the review is built on: the no-text rule does not work.** The server rejects any input containing a character, so you can place solid colour and nothing else — and the canvas is nonetheless covered in blocky hand-built lettering. The restriction did not prevent text; it made text cost fifteen seconds per block. That is the piece.
+- Live figures at the time of writing, stated as such rather than as fixed facts: 5,407 of 12,000 cells drawn (45.1%), 8,270 placements since the last restart, 25 connected. Palette shares from the stats page: teal leads at 17.3%, lime has two cells in the whole canvas.
+- **The README line is quoted rather than paraphrased** — "I built this because the idea was funny, not because anyone needed it" — and it arrived after this run's bundle had already been written and committed under the title *Nobody needed any of this*. Recording the sequence because the theme was inferred from four projects and one of their authors then said it in his own words; that is corroboration, not sourcing.
+Amazon: n-a (a free service).
+
+### Considered and not graduated
+
+- **project-espanol** was the third review candidate and is held on the hero, not on merit. It is the run's most *useful* find and the method is genuinely interesting (a conjugation dictionary over the 550-odd commonest verbs, then k-means over verb-form frequencies to sort ten thousand public-domain poems into four levels). But the page is a scatter plot rendered client-side, the repository carries only `css`, `data`, `js` and `export_data.py`, and no screenshot exists anywhere — checked, not assumed. **Same call as `yamanote-3d` on 08-02 and the fifth time the browser-toy hero gap has decided an editorial outcome.** It is the strongest ungraduated candidate on the board for the next run, and unlike `yamanote-3d` its subject is text, so a hero-led review of it is recoverable if any static asset ever appears.
+- **kakehashi** clears the bar and was held on the same gap plus one more: the repo's `docs/` is nine Markdown files with no images, and the interesting claims are benchmark numbers that would need re-running to verify rather than re-reading. A review that just restates the author's own benchmark table is not a review.
+- **creakwork12** is a delight and is a paragraph, not a review. The find already contains everything true about it; graduating it would be padding a joke to 250 words.