The Bench:
the ten reviews.

Diversions · Sound · on the bench

A turntable modelled properly in a browser tab. One deck, no sync, and an About paragraph that contradicts the page.

Tested 27 September 2026 (UTC) against the local copy of labs.llc build 537 · Comparators checked the same day

16/20 checks met

Scorecard · 16 of 20 checks met

How the checks work Sources 4/4

All four: its wiring file lists no outside source; no relay, no key.

Candour 3/4

Missed C4: the About section says there is no file picker, on a page with twelve load slots.

Proof 3/4

Missed P3: no independent reference for the brake, and playback could not be measured locally.

Footprint 3/4

Missed F4: its three tracks are files fetched from the site.

Reach 3/4

Missed R3: one deck against Mixxx's four. R4 met: a generated track saves as a WAV or a compact WebM.

Open Laband on labs.llc A LABS tool, reviewed by LABS. See the disclosure.

01 What it is

Laband is a single-deck turntable in a browser tab, built for the studio's own three Laband tracks: Neon Horizon Run, Black Glass Tides and Black Glass Undertow. Pitch moves in semitones and tempo as a multiple of the original, and the two are separate controls. A brake spins the record down, and the platter can be dragged either way with the audio following the hand. Twelve slots take your own files or direct links, and a generator will write a new track from a seed.

It is for anyone who wants to scratch, brake or re-pitch a record without installing DJ software, and for the curious who want to hear a platter's physics modelled rather than faked.

Exhibit: Laband, captured from the local copy of build 537.

02 How it works

All of it runs in the browser. The signal path lives in js/dsp.js, which the page loads twice: once as an ordinary script for the ScriptProcessor fallback, and once as an AudioWorklet module, so the read head and the stretcher run on the audio thread wherever the browser allows (dsp.js:1-34; engine.js:709-746).

The read head is signed, so it can run backwards, and interpolates with Hermite curves; pitch and tempo are held apart by a guarded WSOLA time-stretcher (dsp.js:26-31). The source records why: the browser's own playbackRate accepts a negative value and renders silence. The brake is a per-sample one-pole on the rate with a time constant of 0.36 s (deck.js:25), which the source derives as a steady fall of 48.1 semitones a second (dsp.js:553-563). The generator renders in an OfflineAudioContext and fetches nothing (make.js:1-28). A generated track can be saved as a WAV, or encoded in the browser as a compact WebM (make.js:1583-1667); publishing one to the page's library sends only its recipe, a seed and four settings, never audio (make.js:1514-1520). The tool's wiring file lists no outside data source at all.

03 The test

We read the page and its files from the local copy of build 537 on 27 September 2026 and recomputed the brake constant from the source. The three MP3s are not in the local tree, because the studio ships media in a separate archive, so playback itself could not be measured. That is a limit of our test, not a finding about the live page.

Measurements for Laband, 27 September 2026, times in UTC

UTC	What we did	What came back
20:00:12Z	Page on the local copy	HTTP 200, 113,262 bytes of HTML.
20:00:12Z	The three track addresses	Each returned the HTML page rather than audio: the MP3s are absent from the local copy.
20:00:13Z	The DSP worklet, js/dsp.js	HTTP 200, 74,213 bytes, application/javascript.
20:00:13Z	The made-tracks list, made.php	HTTP 200: {"tracks":[],"n":0,"max":1000}.
20:03:00Z	Brake constant, 12 ÷ (0.36 × ln 2)	48.09 semitones a second, the 48.1 the source states.

04 The comparison

Mixxx

fetched 2026-09-27

What it is

Free desktop DJ software for mixing live with digital music files; checked on mixxx.org and its 2.4 manual.

What it does better

Up to four decks with a crossfader and Sync Lock, BPM and musical key detection, timecode vinyl and hardware controllers, chained effects, a library and recording. Laband has no second deck, no sync, no detection and no controller support.

Where Laband goes further

Nothing to install, with the DSP on the AudioWorklet thread and a fallback it announces (engine.js:709-746). Pitch runs −12 to +12 semitones in 0.1 steps and tempo 50–200% as two live dials (index.html:1235, 1249), beside a brake whose physics is written down. The same seed gives the same generated track on any machine (make.js:18-22).

Audacity, Change Tempo effect

fetched 2026-09-27

What it is

A free audio editor; this effect changes the tempo of a selection without changing its pitch.

What it does better

It renders an edited file you can keep, offers an optional high-quality SBSMS stretch for careful work, and sets tempo by percentage, BPM or a target length.

Where Laband goes further

Laband is live: pitch and tempo move while the record plays, and the platter scratches in both directions (dsp.js:26-31; deck.js:734). A modelled brake and spin-up is something an offline editor has no reason to offer (dsp.js:553-563).

05 Where it falls short

• One deck. There is no second deck, crossfader, beat sync, or BPM and key detection; the page's own About section calls it a deck built for one record (index.html:1518-1523).
• The page contradicts itself. That same About section says there is no file picker and no library, while the page offers twelve load slots for files and links (index.html:866-892) and a library of published generated tracks (index.html:707-716).
• Audio linked from other sites loads only if that host allows cross-origin reads; a comment in the page calls streaming links the thing that cannot work (index.html:858-864).
• We could not hear it. With no MP3s in the local copy, stretch quality at 50% or 200% went unmeasured, and we make no claim about it.

06 The ruling

Laband scores where the engineering shows. When the browser's own control could not play a record backwards, the studio wrote a read head that can, and the brake constant we recomputed from the code gives the rate the source promises. It loses marks where a DJ would look (one deck, no sync, no BPM or key detection) and on the stale About paragraph, which is a small fix the studio should make.

As a tab you open to put a hand on a record, it does more than it needs to. As a rig for a set, it is not one; Mixxx is.

On the docket: 16 of 20 checks met. A real turntable model in a tab; one deck, no sync, and an About paragraph the page contradicts.

Try Laband Back to the docket

Diversions · Words · on the bench

A clever shelf over a public catalogue. Left unfiltered, most of what it deals cannot be read.

Tested 27 September 2026 (UTC) against the local copy of labs.llc build 537 · Comparators checked the same day

14/20 checks met

Scorecard · 14 of 20 checks met

How the checks work Sources 4/4

All four: Open Library is the catalogue itself, called from the browser with no key or relay.

Candour 3/4

Missed C2: our evidence shows no stated limit, such as the unfiltered deal being mostly records.

Proof 3/4

Missed P3: we reproduced a deal against the same API, with no independent reference.

Footprint 2/4

Missed F3 and F4: every shelf needs Open Library.

Reach 2/4

Missed R3 (twelve subjects) and R4 (lists and the reading log live on Open Library).

Open Books on labs.llc A LABS tool, reviewed by LABS. See the disclosure.

01 What it is

Books is a reading room laid over the Open Library catalogue. Press deal and it lays out twenty-four books at random from one of twelve subjects. There is also a search box with categories, and an access filter with three settings: Everything, Reads free now, and Borrow with the key. Each book is labelled by whether it can be read without an account.

It is for readers who would rather stumble on a book than search for one, and for anyone who wants to see at a glance which titles are free.

Exhibit: Books, captured from the local copy of build 537.

02 How it works

Everything happens in the browser. books/assets/js/app.js calls Open Library's public search API directly, with a fixed list of fields and twenty-four results a page, and takes covers from covers.openlibrary.org (app.js:25-27, 176-180). There is no relay and no key.

The random deal is engineered. A random subject, a seeded random sort and a random offset are combined so that the API's per-address cache cannot hand back the same shelf twice (app.js:10-13, 199-208). A book with no cover gets a generated typographic one instead of a blank tile (app.js:143-166). The access filter adds ebook_access:public or ebook_access:borrowable to the query (app.js:221-227). The page never handles an account or a password; borrowing sends the reader to Open Library's own forms (app.js:1-9).

03 The test

On 27 September 2026 we fetched the page from the local copy, reproduced a random deal against the same API with the page's own query shape, and ran the free-to-read filter on the same subject.

Measurements for Books, 27 September 2026, times in UTC

UTC	What we did	What came back
20:00:14Z	Page on the local copy	HTTP 200, 40,679 bytes; app.js 16,126 bytes.
20:02:05Z	A poetry deal: seeded sort, offset 1,200, 24 results	HTTP 200 in 1.74 s; 422,892 records found. Of 24 books, 23 had no ebook and one was print-disabled only; 7 had a cover, so 17 would get a generated one.
20:02:08Z	Same subject, ebook_access:public	HTTP 200 in 2.28 s; 21,662 found; all 24 free to read.

04 The comparison

Open Library

fetched 2026-09-27

What it is

The Internet Archive's open catalogue: search, subject browsing, borrowing through Controlled Digital Lending, and reading lists.

What it does better

It is the source. Borrowing, reading, lists, the reading log and yearly goals all happen there; it offers full-text search inside millions of books; and its subject browsing runs far deeper than the twelve subjects Books deals from.

Where Books goes further

One click gives a shelf the cache cannot repeat (app.js:10-13, 199-208), and every card carries an access badge in plain words, with a three-way filter on it (app.js:168-173; index.html:555-558).

Project Gutenberg

fetched 2026-09-27

What it is

A library of free public-domain ebooks with no registration; its home page counted 79,475 on the day.

What it does better

Every book it lists is free to read and download with no borrowing step, and it offers Top 100 and most-downloaded lists and search by language.

Where Books goes further

Books covers borrowable titles as well as free ones, from a far larger catalogue (the poetry subject alone reported 422,892 records in our test), and labels which is which (app.js:168-173). Every book lacking a cover still arrives with a designed one (app.js:143-166).

05 Where it falls short

• Unfiltered, the shelf is mostly records. In our poetry deal 23 of 24 books had no ebook at all. The filter fixes it, but a newcomer has to know to use it.
• It adds no reading tools of its own: no reading log, lists, goals or full-text search. Those live on Open Library.
• It depends on one outside API. Our two calls took 1.7 and 2.3 seconds, and when Open Library does not answer, the page can only say so (app.js:214-216).
• The random deal draws on twelve fixed subjects only (app.js:29-34).

06 The ruling

Books is a thin page and knows it, so its care goes into the two things a thin page can do well. The random deal is built to defeat the catalogue's cache, and the access badge answers the question every stranger's shelf raises, whether you can read the thing, in plain words on every card. Our test showed why the filter matters: with it switched on, all twenty-four books the same subject returned were free to read.

For discovery it is a pleasure. For reading, keeping lists or searching inside texts, go to Open Library, which is where Books sends you anyway.

On the docket: 14 of 20 checks met. A clever shelf over a public catalogue; unfiltered, most of what it deals cannot be read.

Try Books Back to the docket

Live Intelligence · Earth · on the bench

Computes the moon from Meeus and matched the Naval Observatory to the minute. It knows nothing of where you stand.

Tested 27 September 2026 (UTC) against the local copy of labs.llc build 537 · Comparators checked the same day

18/20 checks met

Scorecard · 18 of 20 checks met

How the checks work Sources 4/4

All four: computed in the browser from Meeus; the map is NASA's, and credited.

Candour 4/4

All four: method, accuracy and omissions (libration, a geocentric view) are stated on the page.

Proof 4/4

All four: five phase instants matched USNO's to within a minute.

Footprint 4/4

All four: no data feed; the map is bundled on the same origin.

Reach 2/4

Missed R3 (no rise, set or libration) and R4 (no export in our evidence).

Open The Moon on labs.llc A LABS tool, reviewed by LABS. See the disclosure.

01 What it is

The Moon shows the moon for any moment, computed rather than looked up. It gives the true phase and the illuminated fraction, draws the terminator on the near side as NASA's Lunar Reconnaissance Orbiter mapped it, and reports distance, apparent size and where this month's perigee and apogee fall. Below sit the next four phases, a year of new and full moons with eclipses marked, and a scrubber that plays a whole lunation in thirty seconds.

It is for anyone who wants tonight's moon explained (photographers, walkers, teachers) and for readers who want to know how each figure was reached.

Exhibit: The Moon, captured from the local copy of build 537.

02 How it works

There is no data feed. assets/js/moon-core.js implements Jean Meeus's Astronomical Algorithms, second edition: chapter 47 for the moon, chapter 25 for the sun, chapter 48 for the illuminated fraction (moon-core.js:29-34). The phase comes from the true angle between the computed moon and sun, not from a mean month; the source notes that the real moon runs up to about fourteen hours ahead of or behind the average (moon-core.js:18-27, 190-210).

Each phase instant is found by bisection (moon-core.js:228-248), and perigee and apogee are searched for rather than tabulated, so the closest-this-month line cannot go stale (moon.js:1114-1137). The disc is NASA SVS 4720's LRO albedo map, bundled on the same origin and projected orthographically; only the light on it is computed (moon.js:227-236).

03 The test

On 27 September 2026 we loaded moon-core.js into JavaScriptCore, called the page's own functions for that moment and the phases around it, and set the answers beside the US Naval Observatory's phase service.

Measurements for The Moon, 27 September 2026, times in UTC

UTC	What we did	What came back
20:00:14Z	Page and assets on the local copy	HTTP 200, 57,791 bytes; the LRO map 214,153 bytes; moon-core.js 26,380 bytes.
20:00:45Z	The page's astronomy, run headless	Waning gibbous, 98.32% lit, 16.69 days old, 375,418 km away, 31.82 arcmin across.
20:00:45Z	Five phase instants from the page	Full 26 Sep 16:49:29; last quarter 3 Oct 13:25:52; new 10 Oct 15:50:38; first quarter 18 Oct 16:13:03; full 26 Oct 04:12:16 UTC.
20:04:00Z	The same five from USNO's phase API	16:49, 13:25, 15:50, 16:12, 04:12 UT: all five within one minute of the page.

04 The comparison

US Naval Observatory, Phases of the Moon

fetched 2026-09-27

What it is

Official tables of the primary phases in Universal Time for 1700 to 2100, with illuminated-fraction tables and rise and set data.

What it does better

It is the national authority, and it gives moonrise and moonset for a place, which The Moon does not.

Where The Moon goes further

The Moon draws its answer: an LRO-mapped disc lit by the computed phase, terminator included (moon.js:199-236). It explains distance, apparent size and progress toward perigee, and states its own accuracy (moon.js:1163-1176, 1209-1213), a claim that held to the minute against USNO in our test.

NASA SVS, Moon Phase and Libration (Dial-a-Moon)

fetched 2026-09-27

What it is

8,760 hourly renderings a year of the moon's size, orientation and illumination, with libration and distance data.

What it does better

It shows libration and the true orientation hour by hour, offers north-up and south-up views, and provides high-resolution frames to download.

Where The Moon goes further

The Moon computes any instant live instead of picking the nearest pre-rendered hour, and sets the next four phases, a year of new and full moons with eclipses, and a thirty-second lunation beside it (moon-core.js:190-248; moon.js:1218-1270, 1353).

Moon Giant, Moon Phase Today

fetched 2026-09-27

What it is

A daily phase page with illumination, age, distance and a phase calendar.

What it does better

A month-grid calendar a reader can scan at a glance.

Where The Moon goes further

The Moon names its method and the accuracy of its figures (Meeus chapters, about ten arcseconds, about a minute); Moon Giant's page as fetched did not say how its figures are computed. It also places the moon between this month's perigee and apogee (moon.js:1114-1176).

05 Where it falls short

• No moonrise or moonset, and no location at all. The view is geocentric (moon.js:1211), so it cannot tell you when the moon will be up where you are; USNO can.
• Libration is not modelled, so features near the limb are drawn where they sit on average rather than where they are tonight (moon.js:1212-1213). NASA's Dial-a-Moon shows it hour by hour.
• The disc is drawn north up only; NASA SVS offers both orientations.

06 The ruling

This is one of the most thoroughly evidenced pages on the bench. The Moon carries its astronomy with it, says in the page how accurate that astronomy is, and, the part that matters, the claim survived an outside check: five phase times, five agreements with the US Naval Observatory within a minute. It even takes the trouble to call 'supermoon' what it is, a word coined by an astrologer in 1979 with no astronomical definition (moon.js:1205-1206).

Its gaps are the observer's: no place, no rise and set, no libration. Use it to understand tonight's moon; use USNO to plan when to see it.

On the docket: 18 of 20 checks met. Computes the moon from Meeus and matched the Naval Observatory to the minute; knows nothing of where you stand.

Try The Moon Back to the docket

Studio & Tools · Design · on the bench

Reproducible marks and exports that say what they depend on. The SVG does not outline the type, though its catalogue line says it does.

Tested 27 September 2026 (UTC) against the local copy of labs.llc build 537 · Comparators checked the same day

15/20 checks met

Scorecard · 15 of 20 checks met

How the checks work Sources 4/4

All four: marks are computed locally; the faces come from Google Fonts, named in type.js.

Candour 3/4

Missed C4: the catalogue says the SVG's type is outlined; the code keeps it as live text.

Proof 3/4

Missed P3: determinism was checked against itself; no outside reference exists for a seed.

Footprint 2/4

Missed F3 and F4: the faces load from Google Fonts.

Reach 3/4

Missed R3: twenty-four faces against thousands of fonts and templates.

Open Logo Lab on labs.llc A LABS tool, reviewed by LABS. See the disclosure.

01 What it is

Logo Lab, which lives at labs.llc/dina/, is a mark-making instrument. A seeded geometry engine draws marks in eight modes; a logotype is set in any of twenty-four display faces on a twenty-rung size ladder in four finishes; and the two lock up on an artboard whose ground can be white, black or genuinely transparent. Results export as PNG or SVG.

It is for founders, hobbyists and designers who want to explore marks and wordmarks quickly, and to keep a reproducible record of what they liked: the seed.

Exhibit: Logo Lab, captured from the local copy of build 537.

02 How it works

It runs in the browser. assets/js/engine.js is a pure module: parameters go in, one scene comes out, and two renderers, canvas and SVG, draw that same scene, so the export matches the screen (engine.js:1-25). The eight modes are geometric, script, angular, hybrid, lockup, monogram, tessel and oneline (engine.js:56).

Type comes from Google Fonts in one stylesheet request that asks for exactly the weights each family ships (type.js:104-165). Thirteen faces that ship a single weight are marked single and never requested at 700, so the browser is never left to fake a bold. PNG export uses canvas.toBlob; the SVG is written as a string, and its description records mode, symmetry, carve, grid and ground (engine.js:972-980).

03 The test

On 27 September 2026 we loaded engine.js into JavaScriptCore, built an SVG with the defaults and seed 4242 twice, then with seed 4243, and counted the faces table in type.js.

Measurements for Logo Lab, 27 September 2026, times in UTC

UTC	What we did	What came back
20:00:15Z	Page and scripts on the local copy	HTTP 200, 76,814 bytes; engine.js 55,363 bytes; type.js 37,395 bytes.
20:03:13Z	Seed 4242, built twice	A 3,650-byte SVG, byte-identical on the second run.
20:03:13Z	Seed 4243	A different SVG of 3,250 bytes. The file's description: mark engine v5 · mode=geometric symmetry=cyclic/4 carve=0 grid=off ground=white.
20:03:30Z	Faces in type.js	24 families, 13 flagged single-weight.

04 The comparison

Adobe Express logo maker

fetched 2026-09-27

What it is

A template-based logo maker with Adobe Fonts and Firefly AI tools, on a free plan and a premium one.

What it does better

Thousands of templates and fonts against Logo Lab's twenty-four faces; PDF and JPG downloads with transparent-background options; generation from a brand name, an industry and a style; a wider design suite around the logo.

Where Logo Lab goes further

Logo Lab's marks are generated geometry rather than templates, eight modes with symmetry and carve, reproducible by seed as our run showed (engine.js:56). Every export states what the vector depends on (type.js:510-511; engine.js:959-971), and one spec drives a twenty-rung size ladder (type.js:1-96).

Brandmark

fetched 2026-09-27

What it is

An AI logo generator working from a brand name and keywords, sold with a branding package for a one-off payment.

What it does better

A full package around the logo (brand guide, social profile icons, business-card designs and animated versions) and PDF export alongside SVG and PNG.

Where Logo Lab goes further

Logo Lab is free in the browser with the whole engine exposed: every parameter is a control and is written into the exported SVG's description (engine.js:972-980; app.js:1440). A transparent artboard is a first-class ground, and the PNG export needs no fonts at all (engine.js:38-42).

05 Where it falls short

• The SVG does not outline the type. The logotype stays an SVG text element naming its font, and the export's own note says the file needs that family installed or linked to render as drawn (type.js:510-511). The labs.llc catalogue line promises an SVG with the type outlined; the code does not do it.
• Twenty-four fixed faces from Google Fonts. You cannot bring your own, and a network that blocks Google Fonts loses them.
• No brand kit, business-card or social templates, and no EPS or PDF export, which Brandmark and Adobe Express offer.
• The monogram mask in an SVG uses the reader's system interface font, also a live reference rather than an outline (engine.js:965-969).

06 The ruling

Logo Lab treats a logo as something you should be able to make again, and our headless run bore that out: same seed, same bytes. It is unusually candid about the small dishonesties type tools tend to commit. It will not request a bold that a face does not have, and every export lists what it still relies on.

That candour is also how we found its biggest gap. Because the type is not outlined, a logotype opened on a machine without the font will not look as designed, and the catalogue says otherwise. Until the outlining or the sentence is fixed, export PNG for anything you send.

On the docket: 15 of 20 checks met. Reproducible marks and candid exports; the SVG does not outline the type its catalogue line says it does.

Try Logo Lab Back to the docket

Markets & Capital · Markets · on the bench

A watcher forbidden to write bonuses, and an audit trail to show it. Thirteen cards, a three-week-old check and one known stale headline.

Tested 27 September 2026 (UTC) against the local copy of labs.llc build 537 · Comparators checked the same day

13/20 checks met

Scorecard · 13 of 20 checks met

How the checks work Sources 4/4

All four: issuers' pages read by hand, the wire named, the relay locked to an allow-list.

Candour 3/4

Missed C4: the Active Cash headline contradicts the desk's own note.

Proof 3/4

Missed P3: no independent source was checked against the offers themselves.

Footprint 1/4

Only F2: the watcher and the wire run on the labs server and read Google News.

Reach 2/4

Missed R3 (thirteen cards, no credit-range filter) and R4 (nothing to keep or export).

Open Card on labs.llc A LABS tool, reviewed by LABS. See the disclosure.

01 What it is

Card is the labs.llc page for THE CREDITCARD, a card the studio presents on its own stated terms (2–5% cash back by category, a 0% intro APR for 15 months, no annual fee), set beside thirteen current US credit-card offers (sign-up bonuses, travel, cash back and 0% APR) written up by hand and stamped with the date they were checked (index.html:551-572). A table sets the studio's card against the best of the market and says where the market wins (index.html:640-660). A server-side watcher reads the news for changes, and a live credit-and-finance wire runs alongside.

It is for US consumers comparing headline offers who want to know how recently each figure was checked and where it came from.

Exhibit: Card, captured from the local copy of build 537.

02 How it works

The hand-checked record is card/assets/js/card-data.js: thirteen cards, read on 6 September 2026 from the issuers' own public pages (card-data.js:1-50). A watcher, card-update.php, sweeps Google News for each card and writes only to an overlay file, never to the record. Its rule is in the code: annual fees and the length of a 0% window may be applied automatically, but a sign-up bonus may only ever be proposed for a person to decide (card-update.php:1-45).

The rule came from a measurement. On 13 August a two-publisher agreement rule wrote a wrong bonus to four of seven cards while making no errors on fees or 0% windows (card-update.php:10-28). The page reads the watcher's state to show its flags (app.js:119). The wire passes through a same-origin relay whose allow-list is derived from the feed list; it re-checks every redirect and refuses private addresses (rss-proxy.php:11-22).

03 The test

On 27 September 2026 we fetched the page from the local copy, read the watcher's state with a read-only request, and pulled the credit-card wire through the relay.

Measurements for Card, 27 September 2026, times in UTC

UTC	What we did	What came back
20:00:16Z	Page on the local copy	HTTP 200, 57,758 bytes; card-data.js 7,035 bytes.
20:01:26Z	Watcher state, read-only	Checked 2026-09-06; last sweep 19:57:18 UTC; nothing applied. Review events on three cards, all '[Expired]' headlines from one site, with no proposals.
20:01:27Z	The wire through the relay	HTTP 200 in 0.77 s; 131,616 bytes of RSS, 100 items.

04 The comparison

Bankrate, Best credit cards

fetched 2026-09-27

What it is

A ranked list of recommended cards by category, with its own ratings and a card-matching tool; the page read 'Updated August 19, 2026'.

What it does better

More categories, filtering by credit range, personalised matching through CardMatch, and its own rating method across more cards (fifteen recommendations on the day).

Where Card goes further

Card shows its audit trail: what the watcher heard, what it set aside as expired, and which figures are under review (app.js:136, 252). Its record carries no affiliate links; Bankrate's page carries an advertiser disclosure and issuers' 'Apply now' links.

The Points Guy, Best credit cards

fetched 2026-09-27

What it is

Editorial picks with its own point valuations and a comparison table; the page read 'Updated Sept. 21, 2026'.

What it does better

It puts a dollar value on points and miles, was updated more recently than Card's 6 September check, and covers business and hotel categories across fourteen cards.

Where Card goes further

Card publishes a measured rule for what software may change on the page: fees and 0% windows yes, bonuses never (card-update.php:10-28). Its record carries no affiliate or referral links at all, and each card's fresh-look link is a Google search (app.js:258); The Points Guy states that most cards it features come from partners who compensate it.

05 Where it falls short

• The card at the centre is the one a reader cannot check. THE CREDITCARD's terms are stated with no issuer and no way to apply named anywhere on the page, while every market offer is traced to its issuer's own page.
• Thirteen cards, checked on 6 September: three weeks old on the day of our test.
• A known stale headline. Wells Fargo Active Cash still shows a $200 bonus, while the desk's own note says the issuer's page showed $100 on 6 September. It is flagged Under review but not corrected (card-data.js:59-60).
• Six cards' bonuses and 0% windows could not be re-read on 6 September, because the issuers' pages drew them in by script (card-data.js:29-35).
• No personalisation or pre-qualification, no filter by credit range, and each card's fresh-look link is a Google search rather than the issuer's page.

06 The ruling

Card is built around one question, whether a figure can be trusted today, and answers it with machinery rather than a promise. The watcher that once wrote the wrong bonus to four of seven cards is now forbidden to write bonuses at all, and in our read it had applied nothing and set three expired headlines aside. That is the right instinct.

The same honesty shows the desk's limits: a small set, a hand check three weeks old, and a headline the desk itself knows is wrong. Use it to see how an offer's figures were sourced; confirm the terms with the issuer before applying.

On the docket: 13 of 20 checks met. A watcher forbidden to write bonuses; thirteen cards, a three-week-old check and one known stale headline.

Try Card Back to the docket

Studio & Tools · Design · on the bench

Sixteen presets compose finished multi-photo blends on their own. No layer control, no undo, PNG only, and our copy could not show a blend.

Tested 27 September 2026 (UTC) against the local copy of labs.llc build 537 · Comparators checked the same day

13/20 checks met

Scorecard · 13 of 20 checks met

How the checks work Sources 4/4

All four: bundled photographs and your own files; no relay, no key.

Candour 2/4

Missed C2 and C3: our evidence shows no limit or doubtful figure stated to the reader.

Proof 2/4

Missed P2 and P3: no photographs locally, so no composition could be rendered.

Footprint 2/4

Missed F3 (optional music streams from soundhelix.com) and F4 (photographs load from the site).

Reach 3/4

Missed R3: no layer control, masks or video, which the comparators offer.

Open Multipics on labs.llc A LABS tool, reviewed by LABS. See the disclosure.

01 What it is

Multipics is a generative photo-blend console. It stacks two to eight photographs at a time using CSS blend modes, with per-layer opacity, scale, rotation, blur and hue; cycles through sixteen artistic presets on a timer; takes your own images or a whole folder; and exports the current composition as a high-resolution PNG.

It is for people who want striking collage and double-exposure images without learning a layer editor: artists, social posters, anyone who enjoys a happy accident.

Exhibit: Multipics on labs.llc itself, one page load at 22:11 UTC on 27 September 2026 (build 539). The local copy could not supply the photographs, so this is the one exhibit on this site taken from the public page.

02 How it works

Everything runs in the browser, in one document (multi/multi.php). A bundled list of 117 photographs (multi.php:1143) is stacked in an isolated layer stack using fifteen blend modes, from overlay and multiply to hue and luminosity (multi.php:1226-1230). Each of the sixteen presets, among them CINEMATIC, VHS COLLAGE, GLITCH ART and KALEIDOSCOPE, is written as a character rather than a recipe: a weighted pool of modes plus ranges for layer count, opacity, scale, rotation and per-layer effects (multi.php:1232-1281).

Your own images become local object URLs, a folder opens through the File System Access API, and nothing is uploaded (multi.php:2212-2219, 2532-2546). The PNG export redraws the same layers into an offscreen canvas at up to four times the viewport, scaling blur and shadow lengths to match, and steps down until the buffer fits in about 256 MB (multi.php:2125-2150).

03 The test

On 27 September 2026 we fetched the console from the local copy and counted its parts in the source. The 117 photographs are not in the local tree, because the studio ships images separately, so no composition could be rendered locally; the exhibit above was therefore taken from the public page, where the photographs are served. This test is thinner than the others, and the Proof bar says so.

Measurements for Multipics, 27 September 2026, times in UTC

UTC	What we did	What came back
20:00:17Z	The console on the local copy	HTTP 200, 152,688 bytes: the whole console in one document.
20:00:17Z	The bare /multi/ address	404 locally; the live site redirects it to multi.php with a rule the local server does not run.
20:00:18Z	A bundled photograph	404 locally: images are not in the local tree.
20:02:30Z	Counts from the source	117 photographs, 15 blend modes, 16 presets; layer ranges from 2–4 up to 4–8.

04 The comparison

Photopea

fetched 2026-09-27

What it is

A browser-based raster and vector editor in the Photoshop mould that, by its own account, runs on your device without uploading files.

What it does better

Full manual control (layers, masks, layer styles and adjustments, with undo), and it opens and saves PSD, XCF and Sketch files at the source images' resolution and bit depth.

Where Multipics goes further

No skill required: sixteen presets generate complete multi-layer compositions on a timer, which a manual editor leaves to you (multi.php:1232-1281), from a bundled library of 117 photographs ready from the first second (multi.php:1143).

Pixlr Editor

fetched 2026-09-27

What it is

An online photo editor with layers, the same family of blend modes, filters, AI tools and a subscription tier.

What it does better

Manual layer editing, masks, AI tools such as generative fill, and its own document format for returning to a composition later.

Where Multipics goes further

Automatic, continuously changing compositions from weighted presets instead of one hand-built image, with uploads kept local as object URLs (multi.php:1189-1191, 2212-2219).

Mosh-Pro (formerly PhotoMosh)

fetched 2026-09-27

What it is

A real-time visual effects mixer for images, video and webcam, with a desktop app and a free web version.

What it does better

Video, GIF and webcam in and out, more than sixty effects, and seeded randomisation of effects.

Where Multipics goes further

Multipics blends several photographs together, two to eight layers per preset, rather than distorting one source (multi.php:1236-1255), and there is nothing to install.

05 Where it falls short

• No direct control of layers. You cannot choose which photograph sits where, move or mask it, or undo; the composition is generated.
• PNG only, sized from the browser window (up to four times), not from the photographs' own resolution.
• Uploads over 12 MB are skipped (multi.php:2216, 2221), and the folder picker needs a browser that supports showDirectoryPicker (multi.php:2532).
• No video, GIF or webcam input or output.
• We could not watch it work. With no photographs in our copy, what the presets actually produce went unmeasured.

06 The ruling

Multipics goes to lengths so that the reader does not have to. The presets are written as tendencies rather than fixed recipes, so the console composes finished images by itself, and the export works to match the screen at up to four times its size without breaking Safari's canvas limits. Your own pictures never leave the browser.

It has the lowest Proof and Candour bars on the bench, because our copy could not show a single blend and we found no stated limits. For a quick, striking composite it asks no skill at all; for a picture you mean to control, open Photopea or Pixlr.

On the docket: 13 of 20 checks met. Sixteen presets compose finished blends on their own; no layer control, no undo, PNG only.

Try Multipics Back to the docket

Live Intelligence · Internet · on the bench

A readable window on BGP whose count matched Hurricane Electric's exactly. No RPKI, and one source only.

Tested 27 September 2026 (UTC) against the local copy of labs.llc build 537 · Comparators checked the same day

16/20 checks met

Scorecard · 16 of 20 checks met

How the checks work Sources 4/4

All four: RIPEstat is the RIPE NCC's own service, called direct, keyless, with no relay.

Candour 4/4

All four: method named; failures and gaps in the wire are said, not filled.

Proof 4/4

All four: its address count for AS15169 matched Hurricane Electric's exactly.

Footprint 2/4

Missed F3 and F4: every lookup is a call to stat.ripe.net.

Reach 2/4

Missed R3 (no RPKI, IRR or exchange-point data) and R4 (no export in our evidence).

Open Routing on labs.llc A LABS tool, reviewed by LABS. See the disclosure.

01 What it is

Routing is a plain-language window on the internet's routing table. Enter an AS number, a prefix or an IP address and it shows who holds it, what it announces, its upstream and downstream neighbours, how many of RIPE's route collectors can see it, and a drawing of its announcement history. It opens by showing where your own connection sits.

It is for curious non-specialists and people who work near networks (support staff, journalists, students) who want to see who is behind an address without learning BGP tooling.

Exhibit: Routing, captured from the local copy of build 537.

02 How it works

The browser talks straight to one public API: the RIPE NCC's RIPEstat Data API at stat.ripe.net, keyless and open to cross-origin reads, fed by the RIS route collectors (routing.js:1-10, 36). For a network it calls as-overview, routing-status, announced-prefixes, asn-neighbours and routing-history (routing.js:159-162, 261); for a prefix, prefix-overview, routing-status and rir (routing.js:313-315); for an address, network-info (routing.js:140); and whats-my-ip for yours (routing.js:80).

Each call has a fifteen-second deadline, so a hung request becomes a stated failure. The limit was set after the source recorded AS15169 answering in 0.31 s while AS13335 was still open at 45 s (routing.js:36-58). Named shortcuts cover networks people have heard of, from Cloudflare to Comcast, each with a one-line explanation (routing.js:363-369).

03 The test

On 27 September 2026, between 20:02:18 and 20:02:24 UTC, we made the page's four network calls for Google's AS15169 directly to stat.ripe.net as the page does, spacing them a second and a half apart. We then set one figure beside Hurricane Electric's page for the same network.

Measurements for Routing, 27 September 2026, times in UTC

UTC	What we did	What came back
20:00:18Z	Page on the local copy	HTTP 200, 41,291 bytes; routing.js 21,652 bytes.
20:02:18–24Z	as-overview	HTTP 200 in 0.31 s: holder GOOGLE - Google LLC, announced.
20:02:18–24Z	routing-status	HTTP 200 in 0.30 s: IPv4 seen by 325 of 325 RIS peers, IPv6 by 319 of 319; 1,234 IPv4 prefixes covering 2,615,808 addresses; 176 IPv6 prefixes.
20:02:18–24Z	announced-prefixes	HTTP 200 in 0.73 s: 1,415 prefixes.
20:02:18–24Z	asn-neighbours	HTTP 200 in 0.32 s: 328 neighbours, 139 left, 20 right, 169 uncertain.
20:05Z	bgp.he.net, AS15169	IPs originated (v4): 2,615,808, the same count RIPEstat gave.

04 The comparison

Hurricane Electric BGP Toolkit

fetched 2026-09-27

What it is

A long-standing BGP lookup site with per-network pages of prefixes, peers, graphs, whois, IRR, IX and RPKI data.

What it does better

RPKI validity counts (1,420 originated valid for AS15169), IRR, exchange-point presence, whois and RDAP, and traceroute on one page; peer counts from its own vantage points (679 observed) and AS-path statistics.

Where Routing goes further

Routing explains in plain words, with named shortcuts and one-line descriptions of well-known networks (routing.js:363-369). It shows visibility as collectors seeing the network out of the total, 325 of 325 in our test, and places the reader's own connection with whats-my-ip (routing.js:80, 160).

bgp.tools

fetched 2026-09-27

What it is

A modern BGP and network-intelligence site with per-network pages of prefixes, upstreams, RPKI status, anycast detection and rankings.

What it does better

RPKI status per prefix, anycast detection, network-type tags, named upstream providers and rankings.

Where Routing goes further

Routing reads an open, documented public API straight from the browser and names it on the page, so any figure can be checked at the source (routing.js:1-10, 36), and it draws an announcement history for any network (routing.js:256-303).

05 Where it falls short

• No RPKI validity. routing.js makes no RPKI or ROA call, while both comparators show it.
• One data source. Everything comes from RIPE RIS via RIPEstat, so there is no IRR, exchange-point presence, whois or RDAP tab, and no traceroute.
• The history chart plots the cumulative count of announcement segments opened each month (routing.js:262-280): an envelope of growth, not the number of prefixes announced on a date. Very large networks can exceed the history call (routing.js:303).
• Neighbour labels are incomplete for big networks: 169 of AS15169's 328 neighbours came back uncertain.

06 The ruling

Routing's effort goes into being readable and checkable. It uses one open wire, names it, prints only what that wire says, and gives up after fifteen seconds with a plain message rather than spinning for ever. Our test gave it full marks on Proof: the address count it shows for Google, 2,615,808, is exactly the figure Hurricane Electric publishes.

What a network engineer needs next (RPKI validity, IRR, exchange points) is not here, and the history chart needs reading with care. To learn who is behind an address, start here; to operate a network, use bgp.he.net or bgp.tools.

On the docket: 16 of 20 checks met. A readable window on BGP whose count matched Hurricane Electric's; no RPKI, and one source only.

Try Routing Back to the docket

Live Intelligence · Utilities · on the bench

Its script makes no network request, and its CRC-32 checks out. No paste box, no side-by-side view, ZIP only, and a page that overstates its own silence.

Tested 27 September 2026 (UTC) against the local copy of labs.llc build 537 · Comparators checked the same day

18/20 checks met

Scorecard · 18 of 20 checks met

How the checks work Sources 4/4

All four: it reads nothing but your own files.

Candour 3/4

Missed C4: the page says it makes zero network requests, yet it loads the Google tag and shared site scripts.

Proof 4/4

All four: its CRC-32 returned the standard check value.

Footprint 4/4

All four: no network call exists anywhere in the tool's script.

Reach 3/4

Missed R3: no paste box, no side-by-side view, no formats beyond ZIP.

Open Diff & Crate on labs.llc A LABS tool, reviewed by LABS. See the disclosure.

01 What it is

Diff & Crate, at labs.llc/diff/, is two local desks on one page. DIFF compares any two files: texts line by line with word marks, images by slider, onion skin and a computed pixel-heat map with a difference figure, and anything else byte by byte in hex. CRATE opens a ZIP, lists every entry, previews texts and images, extracts single files and checks each against its CRC-32.

It is for anyone who needs to compare two versions or look inside an archive without sending the files to a website: writers, developers, people handling confidential documents.

Exhibit: Diff & Crate, captured from the local copy of build 537.

02 How it works

Entirely in the browser, by construction: diffcrate.js contains no fetch, XMLHttpRequest, sendBeacon or WebSocket at all. Text uses a hand-written Myers O(ND) line diff with word-level marks inside changed lines (diffcrate.js:116-170). Images are drawn to canvases and compared pixel by pixel into a heat map (diffcrate.js:220-261). Anything else gets a hex comparison over the first 8 MB (diffcrate.js:320-345). When two texts read the same but their bytes differ, it says only invisible bytes differ, encoding or line endings, and shows the byte view (diffcrate.js:180-185).

CRATE reads the ZIP's central directory itself, slices stored entries, inflates deflated ones through the browser's DecompressionStream, and verifies each entry with a hand-written CRC-32 table. Encrypted and ZIP64 entries are named and refused rather than guessed at (diffcrate.js:369-414).

03 The test

On 27 September 2026 we lifted the page's own CRC-32 and diff functions out of diffcrate.js, ran them in JavaScriptCore, and searched the script for every kind of network call.

Measurements for Diff & Crate, 27 September 2026, times in UTC

UTC	What we did	What came back
20:00:19Z	Page on the local copy	HTTP 200, 42,405 bytes; diffcrate.js 28,428 bytes.
20:00:58Z	CRC-32 of '123456789', the page's code	cbf43926, the standard check value; Python's zlib gave 0xcbf43926.
20:00:58Z	The page's Myers diff on two short texts	=alpha =beta −gamma +GAMMA =delta +zeta =epsilon: the minimal edit script.
20:00:58Z	Network calls in the script	fetch: 0. XMLHttpRequest, sendBeacon or WebSocket: 0.

04 The comparison

Diffchecker

fetched 2026-09-27

What it is

An online and desktop tool that compares text, documents, images, spreadsheets and, on the desktop, folders. Its image page says the comparison runs in your browser and nothing is published.

What it does better

Paste-in text, side-by-side and unified views, whitespace and custom ignore rules, a real-time editor; nine image modes including OCR and EXIF; PDF and spreadsheet comparison; shareable comparisons with comments; a fully offline desktop app.

Where Diff & Crate goes further

Its comparison script makes no network request at all, checkable in 28 KB, and has no saving or sharing path to switch on by mistake (diffcrate.js:3-5). It compares any two files, falling back to hex, flags texts that differ only in invisible bytes, and keeps a CRC-checking ZIP inspector on the same page.

ezyZip

fetched 2026-09-27

What it is

A browser-based tool to open, extract and create archives.

What it does better

Far more formats (ZIP, RAR, 7Z, TAR.GZ, ZIPX, ISO, JAR, APK), plus archive creation, password-protected ZIPs and ZIP repair.

Where Diff & Crate goes further

CRATE checks every extracted entry against its stored CRC-32, and its CRC code produced the standard check value in our test; ezyZip's page as fetched does not mention integrity checking. CRATE also previews texts and images inside the archive in place (diffcrate.js:509).

05 Where it falls short

• The page overstates its silence. Its introduction says the page makes zero network requests (index.html:413-415), but the page also loads the Google Analytics tag from googletagmanager.com and a dozen shared labs.llc scripts (index.html:208-216, 636-652). The files you drop still stay in the tab, and the page's own note says none of those scripts reads them (index.html:641-644); the sentence should say that instead.
• Files only. There is no paste box, since the inputs are two file pickers and a ZIP picker (index.html:457-458, 486), so comparing two snippets means saving them first.
• One unified view with two lines of context: no side-by-side view, no ignore rules, no share link.
• Hard caps: texts compared to the first 20,000 lines, at most 900 changed rows drawn, binaries scanned to 8 MB (diffcrate.js:117, 194, 324).
• CRATE reads ZIPs only (no RAR, 7Z or TAR, no ZIP64, no encrypted entries) and cannot create an archive.

06 The ruling

Diff & Crate earns its marks by the least glamorous route: it wrote its own machinery so it would never need a server or a library, then left the proof in a script small enough to read. Our run confirmed both halves of the promise: no network call anywhere in its script, and a CRC-32 that returns the textbook value. The one mark it drops is for saying more than that: the page claims zero network requests while it loads the Google tag.

The price is convenience. Diffchecker is quicker for pasted text and far richer in views; ezyZip handles every archive you are likely to meet. Choose Diff & Crate when the files must not leave the machine, and the others when they may.

On the docket: 18 of 20 checks met. Its script makes no network request and its CRC-32 checks out; no paste box, no side-by-side view, ZIP only.

Try Diff & Crate Back to the docket

Studio & Tools · Utilities · on the bench

Eleven calculators on one exportable tape, decimal-clean in our run. Its currency rates are examples, not live.

Tested 27 September 2026 (UTC) against the local copy of labs.llc build 537 · Comparators checked the same day

18/20 checks met

Scorecard · 18 of 20 checks met

How the checks work Sources 3/4

Missed S2: the currency rates are built-in examples, not read from a reference source.

Candour 4/4

All four: one label, 'Example rates', serves as both the stated limit and the flag.

Proof 4/4

All four: the example rate was checked against the ECB and Xe and differed as labelled.

Footprint 4/4

All four: the calculator's own code fetches nothing at all.

Reach 3/4

Missed R3: eleven calculators against about 200.

Open Paper Tape Calculator on labs.llc A LABS tool, reviewed by LABS. See the disclosure.

01 What it is

The Paper Tape Calculator puts eleven calculators behind one adding-machine tape: standard, scientific, programmer, units, currency, loan and mortgage, percentage and tip, dates, age, BMI and health, and statistics. Every result prints to the tape, where it can be reused, explained, printed, or exported as text, Markdown, CSV, JSON, PNG or PDF.

It is for people who calculate in sequences and want a record (bookkeepers, students, anyone checking a loan or a bill) and for keyboard users.

Exhibit: Paper Tape Calculator, captured from the local copy of build 537.

02 How it works

All in the browser, from one 873,685-byte script assembled from modules (calculator/assets/js/calculator.js). The tape keeps up to 2,000 lines in the browser's localStorage between visits (calculator.js:424-433, 4565-4570). Decimal answers are cleaned to fifteen significant digits so binary noise never shows (calculator.js:320-348). The exporters write text, a Markdown table, CSV to RFC 4180 with a byte-order mark, and JSON that can be imported again (calculator.js:5897-5912), plus PNG and PDF.

The currency mode lists thirty-one currencies with example rates the reader edits, labelled 'Example rates — edit to match today’s'; it fetches nothing (calculator.js:13122-13160). The whole calculator is built to load in JavaScriptCore with no page at all, so its libraries can be tested on their own (calculator.js:1-30).

03 The test

On 27 September 2026 we loaded calculator.js headless in JavaScriptCore, called its libraries directly, and set its built-in euro and yen rates beside the ECB reference rate and Xe's converter.

Measurements for Paper Tape Calculator, 27 September 2026, times in UTC

UTC	What we did	What came back
20:00:19Z	Page and script on the local copy	HTTP 200, 45,265 bytes; calculator.js 873,685 bytes.
20:01:16Z	Libraries, headless	Eleven present, from standard to stats.
20:01:16Z	exact.add(0.1, 0.2) and exact.mul(1.1, 1.1)	0.3 and 1.21; plain JavaScript gives 0.30000000000000004 and 1.2100000000000002.
20:01:16Z	loan.payment: $300,000 at 6% over 360 months	$1,798.65 a month.
20:01:16Z	stats on 2, 4, 4, 4, 5, 5, 7, 9	Mean 5; sample standard deviation 2.1381.
20:02:38Z	Built-in EUR rate against the market	Example 0.92 per dollar; ECB reference 0.87696 (25 Sep); Xe 0.87798 at 20:05 UTC, about 5% apart. Yen: example 150, ECB 157.59.

04 The comparison

Calculator.net

fetched 2026-09-27

What it is

A free collection of around 200 calculators across finance, fitness and health, maths and other topics, with no registration.

What it does better

Around 200 specialised calculators against eleven, each a focused page for one question.

Where Paper Tape Calculator goes further

One tape runs across every mode, survives between visits up to 2,000 lines and exports six ways (calculator.js:4565-4570, 5897-5912); the Calculator.net page as fetched mentions neither a history nor an export. Its decimal answers are clean, as measured above.

Xe Currency Converter

fetched 2026-09-27

What it is

A live converter quoting mid-market rates for 100+ currencies, with alerts, charts and money transfer.

What it does better

Live mid-market rates, 0.87798 euros to the dollar at 20:05 UTC on the day, with rate alerts and history. The labs converter has example rates only.

Where Paper Tape Calculator goes further

Each conversion prints to the tape with its working, alongside ten other calculators, and exports with everything else (calculator.js:13128-13130). Its currency code fetches nothing (calculator.js:13122-13145), so a conversion never waits on a rate service or sends your figures to one.

05 Where it falls short

• Currency is not live. The thirty-one currencies carry fixed example rates you must edit; on the day, the euro example was about 5% off the ECB reference. The label says so, but a hurried reader may not look.
• Eleven calculators against Calculator.net's roughly 200: nothing for retirement, pregnancy or concrete volumes.
• Double-precision arithmetic cleaned to fifteen significant digits; it is not an arbitrary-precision calculator.
• A large single script, 873,685 bytes, to load for a calculator.

06 The ruling

The Paper Tape Calculator goes to lengths over the record rather than the sum. Every result from every mode lands on one tape that outlives the visit and leaves in six formats, and the arithmetic behind it held up headless: 0.3 where plain JavaScript says otherwise, and a 30-year loan payment of $1,798.65.

Its currency mode is honest about being an example, and our check showed why that honesty matters: the euro figure was five per cent from the market. Use it when you want the working kept; use Xe when you need today's rate, then type it in.

On the docket: 18 of 20 checks met. Eleven calculators on one exportable tape, decimal-clean in our run; its currency rates are examples, not live.

Try Paper Tape Calculator Back to the docket

Markets & Capital · Markets · on the bench

Primary sources, named, and a working Form 4 decoder. No portfolios, and a yield curve without the two-year.

Tested 27 September 2026 (UTC) against the local copy of labs.llc build 537 · Comparators checked the same day

12/20 checks met

Scorecard · 12 of 20 checks met

How the checks work Sources 3/4

Missed S2: quotes are Yahoo's unofficial endpoints; the debt, rates and filings are primary.

Candour 4/4

All four: wires named, re-read times shown, the debt clock marked extrapolated.

Proof 3/4

Missed P3: no second feed was checked against the quotes.

Footprint 0/4

None: a live desk fetches, and lookups pass through the labs relay.

Reach 2/4

Missed R3 (no portfolios, screeners or options) and R4 (nothing to keep).

Open Finance on labs.llc A LABS tool, reviewed by LABS. See the disclosure.

01 What it is

Finance is a live, keyless desk that names every wire it reads. A tape runs across indices, rates, commodities, currencies and crypto; a symbol lookup brings charts, news, SEC filings and decoded insider Form 4 trades; and the page adds a Treasury yield curve, the national debt from the Treasury's own daily figure, and the day's headlines.

It is for retail investors and curious readers who want market data and primary filings on one page, with the source of each figure named.

Exhibit: Finance, captured from the local copy of build 537.

02 How it works

The work is split between the browser and a relay on the same origin. Straight from the browser come Frankfurter (the ECB's reference rates), CoinGecko and Coinbase, the US Treasury's Fiscal Data API and alternative.me's crypto Fear & Greed index (finance.js:8-19, 548-587). Through finance/proxy.php, a fixed allow-list relay with a cache time per host, go Yahoo Finance's public endpoints, SEC EDGAR and Google News (proxy.php:1-35, 116-118).

The relay's insiders action fetches the latest Form 4 filings from EDGAR and parses each filing's XML into one row per transaction, sending the named User-Agent the SEC asks for (proxy.php:17-18, 124-160). The debt clock starts from the Treasury's latest debt-to-the-penny figure and labels its ticking extrapolated (finance.js:521-527). Hidden tabs do not fetch (finance.js:21-23). If the relay is missing, reads fall back to the public relay api.allorigins.win (finance.js:65-67).

03 The test

On Sunday 27 September 2026 we read the yield curve and the insider feed through the relay on the local copy, tried to push a foreign address through it, and called two of the direct wires ourselves.

Measurements for Finance, 27 September 2026, times in UTC

UTC	What we did	What came back
20:00:20Z	Page on the local copy	HTTP 200, 61,527 bytes; finance.js 43,517 bytes.
20:01:40Z	Yield curve via the relay	HTTP 200 in 0.08 s: 3-month 4.07, 5-year 5.007, 10-year 5.184, 30-year 5.504; Friday's close, 25 Sep 18:59:55 UTC.
20:01:43Z	The relay asked for example.com	HTTP 403: host not allowed.
20:01:45Z	The Form 4 decoder	40 filings decoded into 124 transaction rows, from the relay's cache built at 19:58:16Z. First row: a GigaCloud Technology director sold 4,497 shares on 24 Sep at $53.58 ($240,966), holding 20,950 after.
20:02:37Z	Treasury and Frankfurter, direct	Treasury: 31 rows, latest 24 Sep, total public debt $40,068,807,991,924.84. Frankfurter: rates dated 25 Sep; EUR 0.87696, GBP 0.75458, JPY 157.59.

04 The comparison

Yahoo Finance

fetched 2026-09-27

What it is

A full consumer finance portal: quotes, charts, news, portfolios, screeners and premium tiers.

What it does better

Far broader coverage (stocks, options, futures, ETFs, mutual funds) with portfolios, stock screeners and comparison, advanced charts and an earnings calendar.

Where Finance goes further

Finance names every source on the page and reads primary data directly (the Treasury's debt figure, the ECB's rates, SEC filings) alongside Yahoo's quotes (finance.js:1-19). It decodes Form 4 filings on the same desk, with no account, no premium tier and no ads on the desk.

Finviz, Insider Trading

fetched 2026-09-27

What it is

A table of recent insider transactions from Form 4 filings, with buy and sale filters.

What it does better

Buy-only and sale-only views, a weekly top list, and a longer, filterable history.

Where Finance goes further

The same Form 4 fields (issuer, insider, relationship, date, code, shares, price, value, holdings after) decoded by the desk's own relay from each filing's XML, beside quotes, filings, rates and news (proxy.php:124-160).

SEC EDGAR full-text search

fetched 2026-09-27

What it is

The SEC's own search across the full text of electronic filings since 2001.

What it does better

It is the primary source, searchable across every filing's text since 2001, with filters for filing type, date range and location.

Where Finance goes further

Finance puts a company's latest 8-K, 10-Q, 10-K and Form 4 links, with a one-click full-text search, next to its price chart and news (finance.js:419-437).

05 Where it falls short

• The yield curve has four points (3-month, 5-year, 10-year and 30-year, from Yahoo's index symbols; finance.js:528-530). There is no two-year point.
• Quotes come from Yahoo Finance's unofficial public endpoints through a relay, not from a licensed feed; if the relay is absent, lookups go to a third-party public relay.
• No accounts, portfolios, watchlists, alerts, screeners or earnings calendar.
• The insider feed shows the latest filings market-wide or for one company, with no filter for buys only, size or clusters.

06 The ruling

Finance goes to the trouble of reaching primary sources and naming them, rather than repackaging one feed. Its relay refused a foreign address with a 403 in our test, and its Form 4 decoder turned forty fresh filings into 124 readable rows. It scores zero on Footprint, and that is structural rather than careless: a live market desk has to fetch, and a relay does the heavy lifting.

It is not a portal: no portfolios, no screeners, a curve missing its two-year. Use it to see where a number came from; keep Yahoo for tracking what you own.

On the docket: 12 of 20 checks met. Primary sources, named, and a working Form 4 decoder; no portfolios, and a yield curve without the two-year.

Try Finance Back to the docket