# Honest Dou Sha Bao, a browser extension that rates the sponsored content in review videos > changshuangdaijing/integrity_baozi- is a Chrome extension that sends the subtitles of a Bilibili review video to a large language model and returns a four-tier judgement on how far the creator's opinion tracks the sponsor's, plus timecoded summaries you can click to. It ships as source zips, needs your own API key, and the repository holds no web presence beyond the README. **changshuangdaijing/integrity_baozi-** — 快速鉴定测评视频的情商高低，并帮你转化成低情商话术辅助决策。 - Repository: https://github.com/changshuangdaijing/integrity_baozi- - Stars: 18 · Forks: 0 - Language: Unknown - License: GPL-3.0 - Published: 2026-10-20 · Updated: 2026-10-20 · Language: en - Canonical page: https://hysenlabs.com/projects/changshuangdaijing-integrity-baozi ## What the extension actually analyses on a Bilibili video Honest Dou Sha Bao is a Chrome extension aimed at one specific question: is this review video soft advertising, and which parts of it can be trusted. The README describes the flow as three steps. The extension pulls the current video's subtitles, title, uploader name, description, pinned comments and the list of related videos. Those are handed to a large language model for detailed analysis. The result is a four-tier emotional intelligence index, S, A, B or C, where the scale is explicitly inverted, higher scores meaning heavier sponsorship, together with timecoded paraphrases that jump to the matching moment in the video when clicked. The whole run takes roughly 20 seconds according to the README, and the translation of the key idea is that the tool provides a set of mouthpieces, or substitute phrases, for parts of the review that were skipped or softened. What makes the analysis input interesting is that it is not only the spoken content. The description, pinned comments and related video list all feed into the model, which means a creator who links a competitor in the description is part of the evidence. The README is explicit that reading happens only when you click the analyse button yourself, and that the data stays in your own browser except for the subtitle text sent to the model provider you configured. The uploader feedback channel is a personal page on Bilibili, and the README asks readers to treat the result as cross-verification input rather than as an accusation against any creator or brand, since misjudgements are described as unavoidable. ## Installing the unpacked extension and wiring up a model key This is the GitHub open source version, version v10.4.9-github according to the README, and it is installed as an unpacked extension rather than from a store. The guide asks you to download and extract the archive shipped with this version into a permanent directory, and it warns not to put it in a temporary folder because deleting that folder kills the extension. Then you open the extensions page in Chrome: ```bash chrome://extensions ``` From there you enable developer mode in the top right and choose to load the unpacked extension, selecting the extracted directory. A bun icon appears in the top right corner of any Bilibili video page, which the guide treats as the success signal. It also suggests pinning the extension from the toolbar puzzle icon so the side panel is easy to reach. The key configuration step is not optional in this build. The open source version ships no shared key at all, so you need an account with a model provider. The panel defaults to DeepSeek, described as metered per use at a very low cost per analysis, and the guide also lists compatibility with OpenAI, Anthropic, Google Gemini and Alibaba Qwen. You create a key on the provider's own site, paste it into the field, save, and optionally press a test connection button to confirm it works. Two details are worth understanding before you paste a key anywhere. The key is stored only in the browser's own local storage through chrome.storage, and the request goes straight from your browser to the provider's official domain with no third-party server in between, including the author's. Custom base URLs are restricted to an allowlist of each provider's official domains so a key cannot be redirected elsewhere. ## Reading the S, A, B and C scale correctly The scale is the part most likely to be misread, so the README states the mapping plainly in a table. C means clean: a genuine use experience where strengths and weaknesses are both covered. B means basically credible: the content is largely objective with some leaning, there are traces of a collaboration, but the weaknesses are genuinely discussed. A means emotional intelligence is high, that is the creator's stance sits very close to the manufacturer's point of view. S means the sponsorship is at full strength, with sideways comparisons, tailored benchmarks and script-like traces. So the ordering runs from C up to S, and the label used for the top band is the one that means worst for the independence of the review. The tool is not a quality score for the video, it is a sponsorship-intensity score. For B and C results, the tool attaches a fixed line asking viewers to like, favourite and subscribe, on the reasoning that commercial collaboration is normal and healthy while hiding it is not. That framing sits alongside a disclaimer section which states that legitimate commercial collaboration is a reasonable and healthy behaviour, that the plugin does not promote an anti-commercial stance, and that the results are reference information for cross-checking rather than evaluations of any creator. The README also asks explicitly that the results not be used in other creators' comment sections as evidence for criticism, and not for personal attacks or commercial disputes. The privacy section makes matching claims: no accounts, no telemetry, no statistics, no advertising and no tracking, with results and cache kept locally and disappearing completely when the extension is uninstalled. ## Caching, the floating panel and everyday use The daily workflow is short. Open a video, click the bun, wait about 20 seconds. The floating bun can be dragged anywhere on screen and its position is remembered. The toolbar icon opens a side panel, which the guide suggests for reading the verdict in detail next to the video. Timecodes in the result, in the form 03:12, jump the player to that point. Moving to another video resets the state automatically, so you click the bun again on the new page. The one detail that affects cost is caching. Results for the same video are cached to avoid paying twice for the same analysis. For a user with their own metered key, that is the difference between trying the tool on several videos and running up a bill, and it is worth knowing that a cached result will not reflect a video being edited or re-uploaded. The repository itself is thin in a way that matters for evaluation. Its top level holds a LICENSE, the README, a second document titled as a usage guide and disclaimer, and two zip archives, one named as source code for version v10.4.11 and one as the extension package for the same version. So the archives are there, but the code is not laid out in the tree for review, and the README text describes v10.4.9 while the archives are labelled v10.4.11, which is a version mismatch worth noting before you install. ## Licence, limits and what it replaces The licence field on the repository says GPL-3.0, which matters in a specific way here: the extension is distributed as a zip archive rather than as an installed package, and distributing a modified version of GPL-3.0 code carries the source-sharing obligation that licence implies. For personal use the question does not arise, but anyone thinking of forking the extension and passing it on should read the LICENSE file rather than assuming the same terms as an MIT project. The real limitations are more practical. The tool only works on Bilibili video pages, and only Chinese-language review content is what its prompts and scale appear tuned for. The output is model-generated analysis of text, so a reviewer who sounds scripted and a reviewer who was actually paid can look identical from a transcript, and the README says as much when it warns that misjudgements happen. The analysis sees the subtitles and the metadata, not the picture, so a video that shows a sponsored product without saying much about it will not be flagged. The alternative is simply reading with more care, which is what the README asks you to do when it says to combine the result with information from several sources. A commercial approach exists too: sponsorship disclosure databases and the Chinese advertising law requirement that commercial content be labelled. Those give a factual, verifiable signal, whereas this extension gives an inference that a model draws from text. Using the extension as a pointer to the timestamps worth watching is a defensible use. Treating its letter grade as a verdict is not, and the project's own disclaimer says so. ## Conclusion Adopt Honest Dou Sha Bao if you watch a lot of Bilibili product reviews and want a second opinion on how sponsored a video reads, and if you are comfortable bringing your own model API key and paying per analysis. Do not adopt it if you need a checked-in, auditable tool or a source repository you can review, because the repository contains only a README, a usage guide and two zip archives rather than browsable code. Verify first by loading the unpacked extension in developer mode on a single video and pressing the test connection button, and treat any output as a cross-check rather than a verdict on the creator. ## FAQ ### What does Honest Dou Sha Bao do on a Bilibili video? It collects the video's subtitles, title, uploader, description, pinned comments and related video list, sends that to a large language model and returns an S, A, B or C sponsorship level plus timecoded paraphrases. The README says the whole analysis takes about 20 seconds. ### How do I install the Honest Dou Sha Bao extension? The GitHub version is installed as an unpacked extension. Extract the provided archive into a permanent directory, open chrome://extensions, enable developer mode, choose load unpacked and select that directory. A bun icon then appears in the top right of any Bilibili video page. ### Do I need my own API key for Honest Dou Sha Bao? Yes. The open source version ships no shared key, so you create one with a provider such as DeepSeek, which is the default, and also works with OpenAI, Anthropic, Google Gemini and Alibaba Qwen. The key is stored in your browser's local storage and requests go directly to the provider's official domain. ### What does an S grade mean in Honest Dou Sha Bao? The scale runs from C to S, and higher means more sponsorship rather than better quality. S marks the strongest sponsorship with sideways comparisons, tailored benchmarks and script-like traces, while C marks a genuine review where strengths and weaknesses are both covered. ### Is the analysis of Honest Dou Sha Bao stored on a server? The README states there is no account system, telemetry, statistics, advertising or tracking, and that results and cache are kept in the local browser and vanish when the extension is uninstalled. Only the subtitle text is sent to the model provider you configured, using your own key. ## Sources - [changshuangdaijing/integrity_baozi- on GitHub](https://github.com/changshuangdaijing/integrity_baozi-) - [Issues](https://github.com/changshuangdaijing/integrity_baozi-/issues) - [License: GPL-3.0](https://github.com/changshuangdaijing/integrity_baozi-/blob/main/LICENSE) - [README](https://github.com/changshuangdaijing/integrity_baozi-/blob/main/README.md) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/changshuangdaijing-integrity-baozi