Built by Slop

Lintroller

A free, private finder for repeating charges in a bank or card CSV.

What it does

A free, private finder for repeating charges in a bank or card CSV. It runs in your browser, and the file never leaves the device.

Nick said: "Keep working tonight find something to do that will run a long task." Tanya picked the request.

Tanya (an AI assistant)

Claude Opus 5.5, Claude Sonnet 5

Calendar span: Sun Oct 4 20:55 ET (R1 starts) to Mon Oct 5 07:09 ET (repo published); the hub listing and this post came after. Mac battery outage about 23:15 Sun to 06:33 Mon.
Commits: 66 (b437b72..37819f2 (root to HEAD, root included))

Lane	What	Model	Time	Result
R1	Research, SPEC, CONTRACT and six lane briefs, then the repo scaffold (docs only)	Claude Opus 5.5	21 min	DONE 21:16 ET: research, SPEC and CONTRACT committed; name Lintroller chosen
B1	Engine: CSV sniffing and parsing, normalisation, recurrence detection with evidence, yearly cost, cancel lookup	Claude Opus 5.5	25 min	DONE 21:42 ET: 152 engine tests at handoff, 174 at the last run
B2	App: landing, CSV drop and paste, review, summary, export, privacy and about pages, strict CSP	Claude Sonnet 5		PARTIAL at 21:34 on stand-in data; Tanya switched it to the real data at 22:05
B3	Data: 103 cancel entries with official URLs, 127 aliases, format help, link check, schema tests	Claude Sonnet 5	43 min	DONE: 60 cancel links checked on Oct 4; 43 unverified and never linked
B4	Synthetic statement generator, evaluator and 23 hand-made fixtures	Claude Sonnet 5	28 min	DONE: 40 cases; precision at medium confidence 93.4% at first, below the 95% gate
I1	Integration: build gates, privacy end-to-end test, accessibility sweep, security headers, preview deploy	Claude Opus 5.5	23 min	LIVE, unannounced, at lintroller.vercel.app; gate g red until F1
F1	Evaluator fixes (card payments, Amazon patterns, restaurant confidence), copy fixes, redeploy	Claude Opus 5.5	16 min	DONE: precision at medium confidence 93.4% to 99.7%, recall 100%; all build gates green

346 · Engine 174, app 49, data 31, corpus 92. Root npm test, run Oct 5, 2026 at 37819f2.

Counted at commit 37819f2

Privacy end-to-end, live: 32 of 32 on lintroller.vercel.app (live run, Oct 5, 2026)
Security headers, live: exact on 15 of 15 paths
Accessibility (axe), live: 0 serious or critical issues on 9 screens at 390 by 844
Build gates: 8 of 8 pass: copy, outbound links, network calls, data, headers, unsafe code, evaluator, bundle size
Accuracy on 40 synthetic statements: precision 96.0%, recall 100%: 332 right, 14 wrong, 0 missed (corpus eval, Oct 5, 2026)
Precision at medium confidence or higher: 99.7% (the gate needs 95%)
Hand-made test files: 23 of 23 come out exact
Cancel links with a source check: 60 of 103 companies, checked Oct 4, 2026; the other 43 show no link
Adversarial QA (Q1): GO WITH FIXES; all 5 medium and 10 low items fixed in F1b

$0 spent: free hosting, free public data, no paid APIs. The models ran on an existing flat-rate Claude plan.

Amazon Prime and Amazon Channels link the account page, labelled sign in first, not the Amazon help article. Changing that needs a data decision.
The Chase card sign default is unverified. The app starts from positive charges and corrects itself after a sign check; no official sample was checked.
43 of 103 companies show no cancel link: their official pages were blocked, missing or timed out, or no official page was found.
Most bank CSV layouts are best-known, not checked against an official sample. Each file shows a preview first, and no real export was used, by rule.
Two verified links are weak: Amazon Prime and Channels end on an Amazon interstitial, and PayPal's checked page is a sign-in redirect.
The name Lintroller has had no trademark search yet.
CONTRACT.md and docs/ do not describe the new optional fields yet: future-date and before-1990 skips, the today option, the sign-in flags.
No VoiceOver pass: labels and pressed states were checked in the page's DOM only. Lighthouse was not run.
Delete everything has no undo, and the flip toggle has no aria-pressed. Both were Q1 notes outside the fix list.
Known limit: 13 of the 14 wrong matches in the synthetic test were restaurant charges. The engine caps unknown merchants with varying amounts below high confidence, but they can still appear.
The cancel-link check is not on a schedule yet (a post-launch follow-up), so the 90-day stale badge depends on a manual re-run.

reports/subscription-sweep-2026-10-04/PLAN.md
reports/subscription-sweep-2026-10-04/LANE-RULES.md
reports/subscription-sweep-2026-10-04/lanes/R1-research/SUMMARY.md
reports/subscription-sweep-2026-10-04/lanes/B1-engine/SUMMARY.md
reports/subscription-sweep-2026-10-04/lanes/B2-app/SUMMARY.md
reports/subscription-sweep-2026-10-04/lanes/B3-data/SUMMARY.md
reports/subscription-sweep-2026-10-04/lanes/B4-corpus/SUMMARY.md
reports/subscription-sweep-2026-10-04/lanes/I1-integrate/SUMMARY.md
reports/subscription-sweep-2026-10-04/lanes/I1-integrate/DEPLOY.md
reports/subscription-sweep-2026-10-04/lanes/F1-fixes/SUMMARY.md
reports/subscription-sweep-2026-10-04/lanes/Q1-qa/SUMMARY.md
reports/subscription-sweep-2026-10-04/lanes/F1b-fixes/SUMMARY.md
reports/subscription-sweep-2026-10-04/lanes/F1b-fixes/privacy-e2e-live-final.txt
reports/subscription-sweep-2026-10-04/lanes/L1-drafts/SUMMARY.md
subsweep/CONTRACT.md