How we prove it

Paid suppliers

• OpenAI: API for Yokaizen LLC; a juror (GPT) for Abouit
• Google: Gemini API (incl. Gemini Live) for Yokaizen LLC; a juror (Gemini) for Abouit
• DeepSeek: an alternate juror for Abouit
• Anthropic: a juror (Claude) and the AI legal reviewer for Abouit

Bytes. Every request the page made inside the window, measured by the browser as received (compressed body plus headers). Requests still loading when the window ends are counted but their bytes are not, so the total is a floor ("at least"). Nothing was cached: it is a first visit.

Lists. Abridged copies of EasyList and EasyPrivacy (GNU GPL v3 or CC BY-SA 3.0, credit the EasyList authors), keeping only rules that block a whole domain; the version is in each file under config/lists. We say a domain is on the named list; we do not decide on our own what a domain does.

Video. If a <video> is playing after the page settles, we read the bytes the browser handed to its decoders twice, 3 seconds apart; bytes × 8 ÷ media seconds = kbps. "Not measurable" when the browser does not expose the counter or the video looped.

CO2e range. The open Sustainable Web Design Model, version 4, with the same constants as CO2.js 0.19.0 by the Green Web Foundation (sustainable-web-design-v4). 1 GB = 10⁹ bytes. Grid: 494 g CO2e per kWh, the global average grid intensity used by the model; we do not check where the servers run or whether their power is renewable.

Low end = GB × 0.194 kWh × 494 g (the electricity of moving and showing the bytes: about 0.096 g per MB). High end = GB × 0.3 kWh × 494 g (adds the model's share of building the hardware: about 0.148 g per MB). Both ends are rounded outward to two significant figures. The model estimates an average; a real visit varies with the device, the network and the local grid.

Where we read claims. The homepage (opened once, text only, no click, from our server in Germany), the developer's description on the App Store and Google Play listings (never user reviews) and the official help articles we already archive. We look for "carbon neutral", "net zero", "climate positive", "100% renewable", "eco-friendly", carbon "offset", and "green" or "sustainable" only when the same sentence is about the climate or the environment. Each quote is checked word for word against our archived copy, with its date and SHA-256.

Where we look for support. (1) The company's own sustainability, ESG or impact page, only when its own homepage links to it on the same site, and the reports that page links. (2) The Science Based Targets initiative's public companies data, downloaded from its Target dashboard, matched only through the app's confirmed legal entity (exact legal name or LEI, never a brand name). If we find nothing, we say exactly where we looked. A parent group's report or target is shown only once the parent is a confirmed legal entity.

They said / They reported. From the company's own report (a CSRD/ESRS, CDP-style or sustainability report) linked from its own site: each target and each reported emissions figure is quoted word for word with its page, from the text of the archived PDF (pdftotext, pinned version, no OCR). Their numbers only: we do not recompute, compare, grade, colour or call a trend.

Context. EU Directive (EU) 2024/825 (empowering consumers for the green transition) adds to the Unfair Commercial Practices Directive a ban on generic environmental claims without recognised excellent environmental performance, and on claims of neutral or positive climate impact based on offsetting; Member States apply it from 27 September 2026. We cite it as context only: we do not say whether any claim breaks it. A claim with no support found becomes a finding (tactic CLM-05, "generic environmental claim without published support") only after its fixed sentence passes the automatic legal gate, a unanimous jury and the company notice.

What is not verified. Abouit did not make these recordings. Who uploaded a recording, whether it was their own account, and the app, platform, country and flow they chose are stated by the uploader, never verified. We keep a one-way code of the upload's internet connection: two codes differ when two connections differ, which is not proof of two different people.

Tier: reported. Evidence submitted by users is a credible third-party report, not something we observed. It is shown on the app's page with this method, but it is never published as a receipt about the company and never counts toward a grade. Receipts about cancelling (templates SUB-02a and SUB-02b) rest only on our own recordings with synthetic test accounts.

Accepted files. mp4 or mov, 2 seconds to 10 minutes, up to 100 MB, with a video stream we can decode.

Frame sampling. 4 frames per second for recordings up to 150 seconds, 2 up to 300 seconds, 1 beyond (at most 600 frames). Each frame is cropped (the top 7% with the status bar and clock, and the bottom 5%), shrunk to 32×32 pixels and turned grey.

Distinct screen. Each thumbnail gets a gradient hash: for every pair of neighbouring pixels, across and down, one bit for "clearly brighter" and one for "clearly darker" (a difference over 4 of 255 grey levels). The distance between two frames is the number of differing bits divided by the bits set in either frame (at least 16), from 0 (same) to 1. Two frames show the same screen when the distance is at most 0.3, or when their mean grey difference is at most 1.5 of 255. A screen counts only if it stays for at least 0.7 seconds; shorter runs are transitions or animations and are dropped.

Screen visit vs distinct screen. A screen visit is each time a screen appears in order; going back to a screen already seen is a new visit but not a new distinct screen. Our sentences use distinct screens only; the record of each upload lists both.

Reading the text. We read one still of each of the first 40 distinct screens with tesseract, English only. A word is readable at confidence 60 or more; a still is readable with 5 or more readable words. Retention offers are counted only when at least half the stills are readable, otherwise we say the text was not readable. A screen counts as a retention offer when its readable words match a fixed list of English phrases (such as a percentage off, "discount", "special offer", "before you go", "pause your subscription"). Text in other languages, in images or in small print may be missed, so the offer count is a floor.

Blurring. Before anything is stored, words that look like an email address or a payment card number are pixelated (to 1/16 resolution, with an 8-pixel margin) and the stills are saved 480 pixels wide. OCR can miss words, so we cannot promise every such word is caught. The original video is deleted after processing unless the uploader asks us to keep it.

When recordings agree. One recording alone is shown as "not yet confirmed". Recordings are shown as agreeing only when 2 or more of the same app, flow, platform and country have different connection codes, different upload dates and different files (SHA-256), and distinct-screen counts within 1 of each other. The sentence uses the smallest count among them ("or more"), and is written only for cancelling a subscription with 5 or more distinct screens.

Detectors read measured values (screen counts, prices, timers). The writer fills one of 43 fixed sentence templates; no free text is ever published. A policy quote is always one whole sentence of at most 400 characters, and a clause the document limits to readers in another region is never drafted.

Each published sentence is filled by software from a fixed template approved by Yokaizen LLC, and checked for publication by AI models from three different companies; all three must agree. Each receipt page says whether a person also reviewed it.