Targetread

W/P TR-7 · Method

What the desk reads, and what it refuses

A code diligence memo reads the software you are buying the way an accountant reads the books, before the money moves. This page names the instruments, the limits, the clock and the refund, and prints every number the free readout rests on.

Which instruments read the target?

A person runs three engines on the desk's own machine, and a reviewer grades what they find.

The three engines live in private repositories. The memo names the commit of each one that read the target, so a later reader knows which version said what.

What does the memo refuse?

The memo also never calls the target's code secure, compliant, clean or safe. It reports findings against named checks, and a check not run is printed as a gap.

How does the five-day clock run?

Five business days, Monday to Friday, from the day access is granted: the day the desk accepts the collaborator invitation, or the day a zip arrives and passes its first-bytes check. Public holidays are not skipped. The order form computes the earliest delivery from today and refuses payment when your deadline falls before it, because a memo that lands after the deal closes is worth nothing to you.

When is the price refunded?

In full, if the desk cannot read the repository within 48 hours of payment, or if the band was wrong on our side. Refunds are issued by a person, through Stripe. Write to [[CONTACT EMAIL NEEDED]].

How is the free line estimate made?

GitHub publishes, for any public repository, the bytes written in each language. The readout divides each language's bytes by an average line length for that language, which gives physical lines, blank and comment lines included. The line lengths below were measured by scripts/fetch-rates.mjs, which counted every file in the named public repositories at the commits shown. A language with no measurement carries an ASSUMED length, and the readout prints the red mark beside every line it moves.

Bytes per line, by language, retrieved 2026-10-10. A rate measured on public repositories names them; an ASSUMED rate says so in its last column.
MarkLanguageBytes per lineMeasured on
C#38.3JamesNK/Newtonsoft.Json@52fa3ae
CSS34.1shadcn-ui/taxonomy@298a885, pallets/flask@d086db8, vuejs/vitepress@8cc9e34, picocss/pico@a07563d
Go29.7gin-gonic/gin@0f09c3a
HTML59.3colinhacks/zod@284243f, expressjs/express@a419aa7, axios/axios@f694ecd, pallets/flask@d086db8, twbs/bootstrap@7771f16
Java36.3google/gson@845664b, square/okio@a6ae13e
JavaScript31colinhacks/zod@284243f, pmndrs/zustand@d7a5583, shadcn-ui/taxonomy@298a885, expressjs/express@a419aa7, axios/axios@f694ecd, twbs/bootstrap@7771f16, vuejs/vitepress@8cc9e34, nvm-sh/nvm@b1b3a00, picocss/pico@a07563d
Kotlin31.8square/okio@a6ae13e
PHP33.1slimphp/Slim@3675bf6
Python34pallets/flask@d086db8, psf/requests@611c616, vuejs/vitepress@8cc9e34
Ruby29.3sinatra/sinatra@cb22afd, Alamofire/Alamofire@bda9ed5, BurntSushi/ripgrep@3fce3b5
Rust34.8BurntSushi/ripgrep@3fce3b5
SCSS29.1sinatra/sinatra@cb22afd, twbs/bootstrap@7771f16, picocss/pico@a07563d
Shell33.6colinhacks/zod@284243f, pallets/flask@d086db8, square/okio@a6ae13e, BurntSushi/ripgrep@3fce3b5, twbs/bootstrap@7771f16, nvm-sh/nvm@b1b3a00
Swift42.5Alamofire/Alamofire@bda9ed5
TSX31.8pmndrs/zustand@d7a5583, shadcn-ui/taxonomy@298a885
TypeScript34colinhacks/zod@284243f, pmndrs/zustand@d7a5583, shadcn-ui/taxonomy@298a885, axios/axios@f694ecd, twbs/bootstrap@7771f16, vuejs/vitepress@8cc9e34
Vue23vuejs/vitepress@8cc9e34
Astro34ASSUMED
Blade40ASSUMED
C30ASSUMED
C++32ASSUMED
Dart34ASSUMED
Dockerfile40ASSUMED
Elixir32ASSUMED
Handlebars38ASSUMED
Jupyter Notebook120ASSUMED
Less26ASSUMED
Lua30ASSUMED
Objective-C38ASSUMED
Perl32ASSUMED
PowerShell36ASSUMED
R34ASSUMED
Sass24ASSUMED
Scala36ASSUMED
Solidity34ASSUMED
Svelte34ASSUMED
Twig38ASSUMED
Any other language35ASSUMED

The estimate is checked against five public repositories that were never used to set a rate. Each was counted file by file, and the estimate from GitHub's bytes has to land within the tolerance below. The same comparison runs in the site's tests on every build.

Five repositories never used to set a rate: lines counted file by file, against the estimate from GitHub's bytes. Tolerance 25%, fixed before the first comparison ran.
MarkRepositoryCountedEstimatedError
sindresorhus/ky@354188836,23131,858−12.1%
pallets/click@2247b3529,23128,593−2.2%
labstack/echo@388226644,03744,656+1.4%
rack/rack@a9833c822,42425,634+14.3%
tj/commander.js@ba6d13d19,25719,862+3.1%

What has this desk done so far?

Zero orders have been fulfilled by this desk. The engines have read other repositories, the fleet's own among them; none has yet read a target under a letter of intent. This sentence changes when the first memo ships.

What else could a buyer buy?

Prices as each seller prints them, with the address and the day they were read. Centurica's reports, sold through Flippa, cover the financials; code appears on that list only as Tech DD, priced "Custom".

Read 2026-10-09. Sources in sources.json.
SellerWhat is soldPrice, as printed
Centurica, via Flippa (investors.club)Red Flag Report: "1 year", "below $30,000"$1,500
CenturicaStandard Report: "2 years", "$30,000 to $200,000"$2,000
CenturicaEnhanced Report: "3 years", "over $200,000"$2,500
CenturicaTech DD"Custom"
Kolonell (kolonell.com, dated 2026-10-01)One app under 100,000 lines, about two weeks"8,000 to 12,000 EUR"
Kolonell100,000 to 400,000 lines"12,000 to 18,000 EUR"
KolonellAbove 400,000 lines"18,000 to 25,000 EUR"
Rockingtech (rockingtech.co.uk)"Technical Review", "10 working days"; unverified, from a search summary"GBP 15,000"
Scrums.com (scrums.com)Tech diligence, "approximately 2 weeks from signed agreement to first deliverable""Sign in for pricing"
TargetreadOne repository up to 200,000 lines, five business days after access$1,900
TargetreadOne repository up to 1,000,000 lines, five business days after access$3,900

An acquirer's exclusivity window commonly runs two to four weeks after the letter of intent. The memo is priced and timed to land inside it.