site: reviews — Kawran Bazar, Elevators, Cycloidal Gearbox (graduate from finds)

daily-finds.log.md

@@ -4788,3 +4788,47 @@ Theme honesty check, since "nothing to install" is the kind of claim that quietl
 Not bundled and deliberately so:
 - `onion-minions-epic-poem` and `joke-ground-beef` — captured today, not shortlisted. Jokes and one-line humour finds have never carried a bundle slot and shouldn't start.
 - `fieldbar-mini-bar` and `do-your-whisky-whiskey-making-kit` — both still sitting in the unbundled shortlisted list and both permanently ineligible under the alcohol rule. They are excluded from the pool up front rather than counted as backlog; noting it again because a raw query of "unbundled shortlisted finds" returns 16, not 14, and the difference is entirely these two.
+
+## 2026-08-02 — reviews: kawran-bazar, elevators, cycloidal-gearbox
+
+Three graduations, all three of today's review candidates taken. **Every fact-check moved something, and in two of the three the correction was to the find rather than to the source** — the inverse of 08-01's pattern, and the reason is that today's candidates are primary documents (a repo README, an article with its own simulator) rather than aggregator retellings.
+
+### review: kawran-bazar
+
+Graduated from find captured on 2026-08-02, source: Atlas Obscura.
+Hero: a Wikimedia Commons night photograph (CC BY-SA 2.0) — traders crouched over onions tipped from a mesh sack, a caged fan running beside them, brass scale pans in the foreground. **Deliberately a different asset from the find's hero**, which is the daylight contributor photo from Atlas. The whole argument of the review is that the market's real hours are the dark ones, and the review's hero is the only image on the site that shows them.
+Category: places (existing).
+Fact-check: **the find survived on its texture and lost its headline number.**
+- **The 57 percent figure is not supported and the review says so.** Atlas offers it as coming from "one source," and no reference work reachable from here carries it. Wikipedia's Karwan Bazar article has no such figure. What is documented is better anyway and went into the review instead: as of 2024 the Dhaka North City Corporation counted four markets across roughly 24 bighas, **1,789 permanent shops and 180 temporary ones**, with daily turnover around **50 million taka**.
+- **The founder's name was corrected.** The find (following Atlas) has "Kawran Sing"; Wikipedia and the Bangladeshi press have **Karwan Singh**, a Marwari trader who established the market in the late eighteenth century — consistent with the find's "1780s". Also recorded: the site carried trade from the seventeenth century, so the market is named for Singh rather than founded from nothing by him. Both spellings of the place name (Kawran / Karwan) are in live use; the find keeps Atlas's, the review body uses Karwan for the historical material because that is what the sources use.
+- **The material addition, and it is the one that changes whether a reader should go:** since **March 2024 the city has been relocating the market to Gabtoli.** The DNCC's own regional office moved first, then shops out of a building declared unsafe. This has been announced repeatedly for decades and never finished — which is consistent with Atlas describing it as fully operational in July 2026 — but a review that told someone to take a 2am rickshaw without mentioning it would be negligent. The review ends on it.
+- Left out deliberately: a Kawran Bazar metro station appears in Commons filenames, and the find's railway detail is a surface line running beside the bazar. Two different pieces of infrastructure, only one of them checked, so neither the station nor the line's name is asserted.
+Amazon: n-a (a place).
+
+### review: elevators
+
+Graduated from find captured on 2026-08-02, source: Hacker News — Show HN.
+Hero: the find's own cropped og image reused — the six-shaft schematic with the wordmark removed. **The only case of hero reuse in this run, and it is a genuine dead end rather than laziness**: the piece is entirely JS-rendered, so every other graphic in it exists only at runtime, and there is no second asset anywhere.
+Category: learning (existing, previously used once).
+Fact-check: **the find undersold it badly, which is the correction worth recording.** The find describes an explainer about SCAN and LOOK. Those are the first two sections. The article then does the work: it rejects mean wait time in favour of **p50/p90** distributions ("people fixate on the times the elevator took FOREVER"), models four distinct traffic shapes (morning, lunch, evening, interfloor), builds up **Otis's RSR** dispatcher with its full scoring formula — ETA plus load penalty plus anti-bunching penalty, minus direction-match, idle-nearby and low-load bonuses — and re-optimises every five seconds. Two benchmark results carry the piece and both are counterintuitive: **LOOK overtakes RSR at high flow rates and in small buildings**, and **destination-dispatch kiosks are generally worse than plain up/down buttons**, because committing a passenger to one car forfeits the five-second re-optimisation that makes RSR work.
+- **One claim did not check out and the review flags it rather than repeating it.** The article opens "The simplest elevator algorithm is called SCAN and was patented in 1961." Wikipedia's *Elevator algorithm* article names no patent at all and puts the earliest published treatment in Knuth's *The Art of Computer Programming*. The 1961/Australian thread that search surfaces leads to **Leo Port** (1922–1978), an engineer and later Lord Mayor of Sydney whose Wikipedia entry credits him with conceiving *destination dispatch* — not SCAN — and notes it was never built in his lifetime. That is a better fact than the one it replaces, so the review uses it: the man remembered for the idea the article spends its final section dismantling.
+- **The correction was pushed back into the find**, which had asserted the 1961 patent flatly in our own voice. The find now reads "The piece dates SCAN to a 1961 patent" — attribution, not assertion. Same call as the Johannes Myronas demotion on 08-01.
+- Trimmed after drafting: the first draft ran 476 words against the 450 budget, cut to 431 by tightening two paragraphs rather than dropping a section. Recording it so the copy-edit stage does not re-derive the same cut.
+Amazon: n-a (a free web page).
+
+### review: cycloidal-gearbox
+
+Graduated from find captured on 2026-08-02, source: Hacker News — Show HN.
+Hero: `media/topless_drive.png` from the repository — the open gearbox from above, the yellow nine-lobed rotor seated inside the black housing with its ten pins. **A different asset from the find's hero**, which is `full_assembly.png` (the closed unit); between them the two pages show the object and its mechanism. The GitHub social card was ignored entirely — it is a signed, expiring URL with the repo name set over it, i.e. the poster-not-screenshot pattern.
+Category: making (existing).
+Fact-check: **the find held up in full and the README carried more than it had room for.** Everything verified first-party from the repository: three versions, with **version two documented as a failure** (a micro drive shrunk to the NEMA 17 footprint, beaten by tolerances a 3D printer cannot hold); version three at **1:9 from ten housing pins and nine rotor lobes, 90mm outer diameter, PLA, 4× M3×8, 2× 6704 bearings, +0.15mm profile offset**, on a 42bygh40-A24dh stepper. The numbers the find had no room for and the review leads on: **1.28 N·m ± 0.0061 output against the bare motor's 0.21 N·m ± 0.0061, at 66% ± 0.220% efficiency.**
+- **One arithmetic note, checked and deliberately not made an issue of.** 0.21 N·m through a 9:1 reduction is 1.89 N·m ideal, and 1.28 / 1.89 is 67.7%, against the stated 66%. The gap is within rounding of the quoted error bars and does not change anything; the review says "about two-thirds" alongside the author's own figure rather than picking a fight over 1.7 points.
+- Two limitations surfaced that the find did not carry, both stated in the review: the generator is a **Fusion 360 script**, so without Fusion you get the CAD files and nothing else; and 1.28 N·m from printed PLA is a demonstrator figure, not a number to design around. The author's own "further room for growth" list (MR128 bearings for the housing pins, M2 screws with metal sleeves for the output pins) is the evidence for that reading.
+- The equations are reimplemented from a published SOLIDWORKS walkthrough, credited as such in the README. Recorded because "I followed a tutorial" and "I turned a tutorial into a parametric generator" are different claims, and it is the second.
+Amazon: n-a (an open-source project).
+
+### Considered and not graduated
+
+- **yamanote-3d** was the strongest find of the run on the merits and was held for one reason: **it has no hero and there is no path to one.** Three sources were checked at capture (og image, Product Hunt, the launches mirror) and all three yield a wordmark poster or a 180×180 thumbnail; the project is WebGL, so no static frame of the carriage exists anywhere. A review is a hero-led format, and a heroless review of a *visual* subject sells it worse than the find does. It stays a find, and it is the strongest argument yet for the standing browser-toy hero problem needing an operator decision.
+- **wavelength-light-app** clears the bar and was held on evidence: the claims are physiological (melanopic lux, circadian phase shifting) and checking them properly means more than reading the developer's own page, which is all that was reachable this run.
+- **the-minds-i** would make a good review and it would be a review of a 1981 book, not of a find — the honest version needs the book read, not summarised from a Book Freak entry.