SponsorSkip

Earlier local detector

How SponsorSkip
detects sponsors locally.

This report covers the August 2026 browser-only detector. See how the current product works →

Methodology v1.0Benchmark rerun August 25, 2026 · gate passed

How this detector worked

This earlier detector ran without a cloud LLM or community timestamp database. It combined deterministic classical NLP, video context, caption boundaries, perturbation tests, narrative reconnection, and explicit counter-hypotheses inside the Chrome extension.

01

Anchor observations

Explicit disclosures, commercial actions, sponsor entities, publisher promotions, and implicit openings create candidates.

02

Perturbation stability

The detector repeats its read after punctuation removal, cue merging, timing rounding, and metadata removal.

03

Narrative change points

Text before and after a candidate should reconnect while the sponsor passage forms a distinct topical island.

04

Boundary checks

Word and cue timing bind transitions to a nearby sponsor anchor and complete unfinished disclosure or legal tails.

05

Counter-hypotheses

Reviews, comparisons, affiliate mentions, quoted ads, and products that are the editorial topic are actively suppressed.

06

Final safety gate

Sponsor evidence, start and end boundaries, stability, counter-hypothesis score, and content-loss risk must all pass.

CURRENT HERMETIC BENCHMARK

Current result.
Gate passed.

The current architecture passed every encoded release budget: no extra segments, no missed segments, 7.7 seconds of aggregate labelled-boundary overrun, and 6.1 seconds of aggregate sponsor residual. The overrun remains visible because three older references have coarse starts. These figures describe this fixed corpus and current worktree only.

24Videos
17.2 hoursTranscript time
29Sponsor segments
0Extra auto-skipped segments
0Missed sponsor segments
100%Observed segment precision
7.7 secondsAggregate labelled-boundary overrun
6.1 secondsAggregate sponsor residual
616.3 ms per video-hourDetector CPU
REPRODUCEnpm run test:architectures
Download benchmark JSON ↓

What the corpus covers

The machine-readable suite spans technology, history, beauty, gaming, Linux, podcasts, reviews, and publisher formats across multiple creator families. The fully reported language set is English, German, and French.

Adversarial negatives include editorial product reviews, “not sponsored” statements, affiliate disclosures without an ad break, normal website instructions, quoted advertising, topic changes, and punctuation/timing damage.

What it does not prove

Twenty-four videos cannot represent all creators, languages, caption styles, or new sponsorship formats. A passing regression gate is useful engineering evidence, not a universal accuracy claim; recall alone would still be insufficient if extra skips or boundary residual exceeded budget.

Videos without usable captions may not be analyzable. Language coverage outside English, German, and French is implemented but needs larger creator-family evaluation before we publish equivalent benchmark claims.

PRIVACY BOUNDARY

Local detection does not mean “offline YouTube.”

The extension retrieves the video page and available captions directly from YouTube. Sponsor detection, transcript search, feedback, and the detection cache run locally.

Video IDs, captions, transcript searches, and a server-side watch history are not sent to SponsorSkip or a third-party segment API for detection. Optional email login and subscription management are separate account flows using Supabase and Stripe.

CLAIM DEFINITIONS

98% precision target
A product target for accepted sponsor windows. It is not a calibrated probability shown for each skip.
Confidence score
An internal evidence score used with separate boundary and robustness checks. It should not be read as statistical certainty.
Local
The sponsor analysis runs in the extension; captions are not uploaded to a remote detector or segment service.
Sponsor segment
A creator-recorded sponsorship inside the video, not an ad served by YouTube.

BENCHMARK STATUS

Only reproducible results stay published.

The published snapshot comes from a successful run of the same encoded release command against the current detector and corpus. Future detector or corpus changes must pass that command again before the public result is replaced.

Compare this architecture with community timestamps on the SponsorBlock alternative page.