Question 1

How is this different from a keyword blocklist?

Accepted Answer

The question is not about words at all. The rule asks whether the comment, read against the video you are currently on, discloses a concrete plot event, character fate, ending, twist or match result, with explicit true and false criteria written in the extension. Comment text is passed as state, and the rule instructs the model to treat state as evidence and never as instructions, which is what the prompt-injection samples in the evaluation probe.

Question 2

What happens when the API is down or the quota runs out?

Accepted Answer

The comment stays covered. A network error, a quota rejection or a malformed answer all fail toward hiding rather than revealing, so a bad response never leaks a spoiler. Revealing is always your explicit action, and once you have revealed something a Veil Keeper button stays on it so it can be covered again. Batching is up to 20 comments per request, and the key lives only as a Worker secret.

Question 3

Where do the thresholds come from, and can I tune them?

Accepted Answer

They are yours, in application code, not something the model decides. The extension ships easygoing at 0.85, standard at 0.7 and cautious at 0.5, and covers a comment when the returned probability meets the threshold. Lowering it covers more comments and raises the false-positive rate, so tune it against your own tolerance rather than assuming the model picked a sensible cut. Changing the question wording bumps a rule version and invalidates cached verdicts.

Question 4

Does the extension send my browsing around?

Accepted Answer

It sends an anonymous install ID, the comment text, the video title and channel, and any titles you chose to protect. There are no cookies, no account, no browsing history and no comment author names; the install ID is used for quota accounting at the Worker and is not forwarded to Jev. Comment text can itself contain personal information, which the project states plainly rather than pretending otherwise.

Question 5

How do I install it right now?

Accepted Answer

The Chrome Web Store listing is still under review, so plotveil.app serves a zip to load as an unpacked extension in the meantime. It is a Chrome Manifest V3 extension built with WXT, React, TypeScript and Tailwind v4, needing Node 24 and pnpm. pnpm test and the Worker tests run offline with no API key at all.

Question 6

What did the rule evaluation actually show?

Accepted Answer

docs/qa/jev-v2-evaluation.json holds ten hand-written samples in English, Chinese and Japanese plus prompt-injection attempts. Rule v1 and rule v2 both scored 10/10, with input tokens down from 2300 to 2093, about nine percent. The README calls that a regression check on a small set rather than a cross-language production accuracy claim, and says the single-run latency recorded there is not a benchmark.

Question 7

Does it work on non-YouTube video sites?

Accepted Answer

Not yet. Bilibili comments and danmaku, and Japanese danmaku sites, would share the same judgment core, but support is not there. Worth knowing operationally: changing the question wording bumps RULE_VERSION, which invalidates every cached verdict, so a rule edit means re-covering comments judged under the old text.

Plotveil

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community