Not sure how much I want to let folks know, but you can scrape the routes from https://eddm.usps.com/eddm/select-routes.htm (backend you can just check out the ESRI server it hits in DevTools, it is just an open endpoint).

So those are not "polygons", but the actual polyline routes. But if you get that far you can probably figure out the GIS operation you want to do for your use case.

Those are the more specific routes as well (so more specific than the census ZCTA), although ZCTA should be good enough for most folks anyway to be honest.

I am currently using Perplexity for https://veruscite.com/. Verify citations actually exist -- so get in a list of 100, half I can check with crossref, but the other half I need to do a web search and fetch to make sure they exist.

So I need higher rate limits to not make the tool come to a crawl. I probably will not consider DeepSeek API here (I have tried DeepSeek hosted at US providers, like Perplexity). The other search/fetch tools I try do not have great web coverage of Asian journals, but the fact DeepSeek API is not zero data retention I cannot get over.

I am maybe not using copy-editing the way you are (feel free to correct me what is more typical in the field). My prompt is not much different than "review the text for minor grammar and spelling errors, and list the corrections you would suggest". I accept the large majority of the suggestions.

For technical review of the content, I just send it out to other experts.

Assuming that the result you get is reliable (I would not assume that), maybe that is enough. I had one book here where the author frequently repeated herself--sentences that basically said the same thing she'd just said in a previous paragraph. I had to cut all that. One with a lot of useless filler phrases that added nothing. Another where the author wrote whole paragraphs in the form of rhetorical questions. It was much better rephrased as statements. I did a lot of fact-checking on one book. Jargon that can be translated into English. Tables and charts where the data is not well organized. I mean, it goes on and on.

I know what you mean by "cannot write out of a paper bag". Agree AI won't help those folks (I honestly don't know how anyone has enough patience to help them).

My above prompt does indeed work well for me, but agree would likely not catch those example errors you list.

I am confused by this as well, shouldn't it be all white listed sites at the start?

So I wish I could use Google for https://veruscite.com/, but the number of Google searches are a hard cap on the account! So yes that is fine for agentic coding, but for an app that relies on web-search is not sufficient.

I am currently using Perplexity fast search and fetch, and I am happy with that. I would try our Ceramic.ai, but I need to be able to fetch the pages as well (I do not want summaries).

(I work at Linkup.) We do both: search returns raw results, no summaries, and there's a separate fetch endpoint that returns the full page as markdown, with optional JS rendering. You should compare us against your Perplexity setup - you might some value in switching

Put in the todo list to check out. If you cut web search costs even further for flash/fast search (similar to Perplexity, which is $1 per 1000 for fast search) would make it more competitive (at least for my app!)

If you check out https://veruscite.com/hallucinations there are much more banal examples than that article.

Part of the forensic work with this, I won't ever be able to tell the difference between "I as a human messed up" vs "LLM go brr and filled in everything". So I think editor discretion when to correct vs when to retract is fine. (I do think that PLoS One correction should have been retracted as well though.)

While this is true, it is still possible to take the probabilities and make determinations of recall and false positives using conformal sets, https://crimede-coder.com/blogposts/2026/ConfClassification

(Or just use a model to re-calibrate the probabilities, I like the conformal approach though as those rates are what I often care about.)

You just need some labelled data to generate the "corrected" probabilities (or thresholds to meet the specified error rates).

They are relevant and courts have considered the whole of a persons movements standard in terms of whether querying ALPR data constitutes a search, https://andrewpwheeler.com/2026/08/12/license-plate-reader-s...

Under current case law (Carpenter and recently affirmed in Chatrie) it will definitely be a search, IMO it is just when the sensors become dense enough according to the court (absent states do not make regulation themselves to require a warrant for historical data).

I sent it to the moderation email this morning. Will update if I hear back.

Great; it would be good to have a precedent. (The one you picked, [6], is indeed a blazing case.) And good luck with your business; it is very much welcomed!

This is what the majority of mine look like as well. Just simple instructions for things I need to do repeatedly.

I have not even really needed a formal memory system. If I see an error happen more than once, I just say "hey add a note on this to agents.md". Tends to be verbose but overall works quite well for the projects I am doing.