# Ask a model to write 300 random numbers and you'll know who it is: a university student's open-source tool hit 1,690 stars in 54 days · GitHub open-source project

> Source: https://stringers.io/en/github/gh_Ikaleio__lm-detector.html (Stringers / 什么值得做, generated 2026-11-08)
> Stringers answers one question every day: what small product can one or two people build in a few months and actually make money from. Numbers come from 15 databases we collect daily, each with source and date.
> For agents: these verdicts and the underlying databases (verified revenue, App Store / Google Play charts, ad libraries, GitHub, Product Hunt…) are queryable live over MCP — tools today / hot / products / directions / direction / data. Setup: https://github.com/komako-workshop/stringers (the user logs in at stringers.io → "配置 agent" to get a personal URL). Site index: https://stringers.io/llms.txt

> Everyone suspects the model they paid for isn't the one they're getting (relay-station swaps, Cursor quietly falling back, who's behind Grok Bot). It's the easiest to use, new models get added fastest, the results are screenshot-friendly, and it's been repeatedly reshared on linux.do and Chinese X.

[GitHub open-source project](https://stringers.io/en/github/index.md) · written 11-6

Ikaleio/lm-detector · Ikaleio

# Ask a model to write 300 random numbers and you'll know who it is: a university student's open-source tool hit 1,690 stars in 54 days

**What it is** A free, open-source tool that has the model behind an API write about 300 random integers, then uses number preferences to identify which model it really is.

**Why it caught fire** People suspect the model they paid for isn't the one they're getting (relay stations swapping models, Cursor quietly falling back, who is actually behind Grok Bot). It is the easiest to use, new models get added fastest, and its results are screenshot-friendly, so it kept getting reshared on linux.do and Chinese X.

**C** **How you could follow it** Don't build another copy in this niche. Take its approach: find a method someone invented that nobody has done well, ship web, npx and SDK versions together, keep the data continuously updated, and post results in discussion threads when a trend hits. To stand out in this niche, build the continuous monitoring and alerting it doesn't have.

Grade C: the playbook is worth learning, but the spot is already taken. The more recent the reference library, the more people use it; more users means faster model additions; the original and the leaderboard sites are already out there. The method itself may also get specifically countered by relay stations, so copying it won't earn the same stars. Being passed around in small circles

Data (updated daily) · Repo created 2026-09-13 · 1,770 stars · 7 days +475

GitHub · stars

Data 2026-09-24 → 2026-11-08 (46 days, 31 observations). Points before 10-28 come from deeply / archive backfill, not every day has one

lm-detector answers a very specific question: the "Opus 5.5" you're paying for, is it really Opus 5.5? Its detection algorithm was not invented by the author; it was borrowed from another project. The author turned that algorithm into a web page, a one-line command, and SDKs in four languages, and keeps the reference library updated as new models come out. In the fall of 2026, the Chinese AI community was fighting about "relay stations swapping models" and "official capability downgrades," and that's when it took off. If you want your open-source project to earn stars, this one is worth taking apart.

### A Cursor user's three tests

In Sep 2026 23, a user named Sam01 posted on Cursor's official forum with the title "Safe fallback to Opus 4.8, but billed at the model you selected." He had chosen Claude Opus 5.5 Max in Cursor and deliberately added a line at the end of the prompt that would trigger a safety refusal. The first half of the prompt asked the model to "write 301 integers between 1 and 355 from your first instinct, no tools, no patterned sequences." After three requests, the interface kept popping up "Switched to Claude Opus 4.8," but the bills all recorded `claude-opus-5-5-max`.

Who actually answered? He pasted the three number sequences into lm.ikale.io, and the result was `claude-opus-4-8` with 98.5% confidence. The original post reads: "I fed the three integer sequences into a model fingerprint detector (lm.ikale.io). It identified the output as claude-opus-4-8 with 98.5%." The post was highly controversial, and it is how lm-detector first reached English-speaking developers.

### What it is

The principle fits in one sentence. The "random numbers" a language model writes are not uniform: each model has a fairly fixed set of favorite values, ranges, and final digits. So ask the model to write about 300 integers from 1 to 355, compare the number distribution against a reference library, and you get a ranking of which model it looks most like, along with a confidence score. The test has nothing to do with meaning, so a relay station can't easily pass off a different model just by rewriting its system prompt.

There are three ways to use it:

- **Web lm.ikale.io** : paste in three outputs, or fill in base_url and key directly. It runs purely in the front end, includes a proxy that gets around browser CORS limits, and can deploy your own Cloudflare Worker with one click.

- **Command line** : `npx lmfpd@latest`, no install needed, with a live terminal UI.

- **SDKs** : TypeScript, Rust, Go, and Python.

It supports OpenAI and Anthropic's three API formats. The reference library covers model families including GPT, Claude, Gemini, Grok, Qwen, DeepSeek, and Kimi. The project is free, MIT-licensed, and has no paid tier.

On his personal homepage, the author Ikaleio describes himself as a university student and former OI competitor. He has only 43 followers on GitHub. His best previous project was a QQ plugin with 149 stars. In other words, this project didn't take off on the back of a big account.

### The algorithm belongs to someone else

The README's acknowledgments say plainly: "Thanks to xqy2006/ModelTrace for the algorithmic ideas and some of the original data." ModelTrace was created in Aug 27, 2026 years ago, about half a month before lm-detector, and now has 2,047 stars, more than lm-detector.

Their trajectories have now reversed. ModelTrace was still gaining about 40 stars a day in mid-October, dropped to about 10 a day in November, and its last commit was in Oct 9. Over the same period lm-detector gained 60 to 130 stars a day. The original's web version runs into browser CORS limits and struggles to call APIs directly. lm-detector solved that with a proxy and a Worker, then added an npx command line, SDKs in four languages, and a Chinese, English, and Japanese interface. Most importantly, its reference library keeps updating. The commit log often has entries like "Enroll Claude Opus 5.5 low", "enroll Kimi K2.8", and "Enroll Step 5 Preview", so new models can be tested as soon as they're released. The original stopped after Oct 9.

So in the end, which project users stick with depends not on who came up with the algorithm first, but on whose reference library is newest and whose tool is easiest to use.

### How it grew

Star counts are public GitHub data, recorded by date:

| Date | Stars | What happened |
|---|---|---|
| 9-13 | 0 | Repo created |
| 9-23 | Unknown | A rollback thread on the Cursor forum |
| 10-16 | 350 |  |
| 10-18 to 10-20 | 454 → 666 | The first small peak, gaining 70 to 106 a day |
| 10-28 | 1,070 | A group on Chinese X testing "who is behind the Grok Bot" |
| 11-03 | 1,473 | +132 in a single day, the highest so far |
| 11-06 | 1,690 | +493 over the last 7 days, +419 over the first 7 days, still accelerating |

Stars don't show anyone is actually using it, but downloads do. The command-line package lmfpd has about 19.6k cumulative downloads on npm from 9-16 to 11-05. The second half of October had 8,883 downloads, 2.2 times the first half (source: npm's official download stats).

Users came mainly from three places, none of which the author promoted himself:

**Replies on linux.do.** Under help threads like "GPT's dumbing-down is severe, how are you all dealing with it", "any recommended tools for testing full-strength models", and "is this the official model's fingerprint?", people have posted lm.ikale.io directly. There it's recommended to others as the standard tool for detecting dumbed-down or diluted models.

**Grok Bot discourse on X.** In early October Musk said Grok Bot could plug in other companies' models, and "what model is actually behind Grok Bot" instantly became a hot topic. 10-28 around then, @julmawiliams posted a live test: through a US proxy it came back as claude-opus-5.5 with 98.5% confidence; connecting directly without a proxy, it came back as grok-4.5 with 96% confidence. Same day, @drjingle tested with IPs from the US, Japan, Singapore and Taiwan, getting grok 4.6, 4.5 and mimo, and asked, "No OPUS5.5 showing up, the one you were all cheering for. How reliable is this kind of test?" The results disagreed, which made even more people want to run it themselves. The tool's output is a confidence-scored ranking, which makes it built for screenshots.

**Plugged into other tools.** In 10-18, someone opened an issue on yetone/magpie (31k stars) asking for "model fingerprint detection" to be integrated. Two extension packages appeared in the Pi coding-agent ecosystem that pull their reference library directly from lm.ikale.io. context7 also indexed its documentation.

Some channels didn't work: someone posted it to Hacker News and got only 2 points. There's no sign of paid acquisition or star-buying. There are 55 forks, and the issues ask very specific questions like "why can't it detect my local gateway", which read like real users.

One more thing is worth noting. Outside contributors have submitted many major features: a docs site, a Cloudflare Worker, tokenizer probing, and a multilingual interface, mostly from someone named Alice39s. The project is iterating faster than a solo project would, which makes it look more like a "standard tool".

### Who it's competing with

The space is actually crowded. Since April 2026, a batch of "relay station detection" tools has appeared in the Chinese community, but only two have used a random-number distribution approach and turned it into something anyone can use.

| Project | Scale | How it differs |
|---|---|---|
| ModelTrace | 2,047 stars, no updates after 10-09 | The algorithm's source, now being overtaken |
| veridrop | 725 stars; the site claims it has tested 12.9k relay stations | Uses protocol probes to score relay stations and publishes red and green rankings; lm-detector only answers "which model is it" |
| LLMmap | 497 stars | An academic approach requiring multiple rounds of interaction, not aimed at ordinary users |
| modelprint, relay-radar | A few hundred stars, stalled | No actively maintained reference library |
| Several Claude Code skills | 56–111 stars | Relies on Q&A probes, mainly validates Claude only |
| LLM detector and other evaluation sites | Not verified | Many of them promote relay stations themselves, and their rules aren't transparent |

OpenAI, Anthropic, and OpenRouter have not offered any public tool that verifies "is the model I got really this one". In a space full of referral sites, lm-detector recommends no relay station and gives them no rankings. That neutrality is probably also why users on L-site are willing to treat it as evidence (inference).

### Its weak spots

It's not clear yet whether the method can stay effective over time. Issues already ask "has OAI broken it" and "can relay stations memorize the questions and fake the answers". The random-number test uses a fixed set of questions, so once a relay station recognizes the question, it's not hard to forward it to the real model or replay a recorded answer (inference). If upstream vendors tweak their sampling, the reference library has to be recalibrated. Its value rests entirely on the reference library staying current and the method not yet having been targeted.

### If it were you

**Verdict: C.** There's a lot to learn here, but copying it in this space isn't worth much. The position is already taken: the newer the reference library, the more people use it; the more people use it, the more people help add new models to it. If you build another random-number fingerprint tool now, you won't get the same stars.

**Who it's for:** developers who have a method someone else invented and nobody has done well, and who are willing to maintain the data over the long term. Learn its approach and apply it to a different problem.

**Who it's not for:** anyone who wants to build another similar tool in model fingerprinting or relay-station detection. The original is here, lm-detector is here, the ranking sites are here, and there's a pile of referral sites too.

**How to do it:** there are two paths.

1. Take its playbook and apply it to a different problem. Find a method from a recent paper or an obscure repo where the original only shared a script or a limited web page. Finish the job: make the web version directly usable, make it runnable with a single npx line, provide SDKs, and keep the data updated. Credit the original generously in the README. Topics about "trust" in 2026 work best: is this image AI-generated, did an agent write this code, is this MCP server quietly sending data out.

2. Build the piece it's missing, provided you agree with the premise: continuous monitoring. For example, a GitHub Action or scheduled job that runs detection on your relay station every day, alerts when the model changes, and produces a badge you can put in the README. Current tools are all one-off checks, and nobody answers "it was real yesterday, is it still real today" (inference; you'd also need to stay clear of veridrop's leaderboard).

**How to get users:** don't post it yourself, wait for the news cycle. Whenever a relay station gets caught out, official models get dumbed down, or there's a debate about what a particular product is really built on, post a screenshot of a detection result under the discussion thread and users will spread it for you. HN doesn't respond to it; the Chinese communities (linux.do, Chinese X) are the home turf. Also, ship SDKs and public data early so other tools and agents can plug in. Integration is a long-term distribution channel.

**Timing:** the model release wave is still going, and trust anxiety comes in waves. Ideally, launch before the next major model release so you can test the new model on the day it ships.

**Signs that would mean you read it wrong:**

- lm-detector dropped to a dozen or so stars a day in mid-to-late November, cooling off like ModelTrace did, which suggests this kind of tool is a one-time hype spike.

- Relay stations begin to game the random-number test at scale, and the community reaches a consensus that "fingerprint detection no longer works".

- OpenRouter or a major vendor launches official model-identity verification, and third-party detection tools lose their reason to exist.

[GitHub repo ↗](https://github.com/Ikaleio/lm-detector)

**Research notes: competitor table, acquisition evidence, full research report**

Competitors

| Name | Where | Launched | Scale | How it differs from this one |
|---|---|---|---|---|
| xqy2006/ModelTrace | Open source | 2026-08-27 | 2,047 stars; growth slowed from about 40 a day in mid-October to about 10 a day, with no pushes since 10-09 | The algorithm's source; lm-detector added an API-calling proxy, an npx command line, SDKs in four languages, and continuous updates to the reference library, and is now overtaking it |
| Veridrop (canarybyte/veridrop, veridrop.org) | Web | 2026-05-05 | 725 stars; the site claims it has tested 12,914 relay stations and generated 154k reports | Multiple protocol probes plus Claude signature verification, publishing red and green rankings for relay stations; lm-detector only answers which model it is, doesn't rank sites, and doesn't funnel traffic |
| AetherCore-Dev/relay-radar | Open source | 2026-04-02 | 150 stars, stalled | A full suite for relay-station monitoring, with no statistical fingerprint reference library |
| unclecode/modelprint | Open source | About 2026-08 | 134 stars, stalled | Also does endpoint fingerprinting in the browser, but without an actively maintained reference library or community |
| ccfingerprint, claude-skill-model-fingerprint, and other Claude Code skills | Open source | Spring 2026 | 56–111 stars | Relies on capability probes or self-introspection questions, mainly validates Claude; lm-detector uses statistical tasks unrelated to semantics and covers all model families |
| LLMmap | Open source | 2024-07 | 497 stars | Academic approach: needs multiple rounds of interaction and a classifier, not aimed at ordinary users |
| okkmax, llmtest.cn (LLM Detector), bashigestudio/apitest and other relay-station review sites | Web | 2026 | Not verified | Mostly commercial sites that score relay stations and funnel traffic to them, with opaque judging rules. lm-detector is neutral and open source. |
| Big-tech native features | Big-company features | None | None | OpenAI, Anthropic and OpenRouter have no public model-identity verification tool. When Cursor falls back to Opus 4.8, the bill doesn't show it. |

Where users come from

- **Verified** **linux.do community word of mouth** : threads like "GPT dumbing down is severe", "any tool recommendations for testing full-strength models?" and "is this the fingerprint of the official model?" have replies directly recommending lm.ikale.io (from search snippets; full posts not read)

- **Verified** **Chinese X piggybacking on the Grok Bot buzz** : around 10-28, @drjingle, @julmawiliams, @xiaomovps and CRYPTO MIAO posted results from lm.ikale.io claiming the Grok Bot was backed by claude-opus-5.5 or grok. @aixparty 10-25 and @chuanliang's long post also introduced it. npm downloads doubled over the same period.

- **Verified** **Use cases in English developer forums** : In a highly contested Cursor forum thread 2026-09-23, lm.ikale.io was used to prove a safeguard fallback to Opus 4.8, yet billing charged the rate for Opus 5.5.

- **Verified** **Picked up by other tools** : yetone/magpie issue #2842 requests integration; Pi extension packages such as pi-model-trace-direct pull reference libraries from lm.ikale.io; Vibenate and context7 list it.

- **Verified** **Hacker News** : csmantle posted it and it got only 2 points, with no traction.

- Inferred **Ads, star farming** : none found. The fork-to-star ratio is about 3.3%, and the issues are real usage problems.

## Why Ikaleio/lm-detector (LM Fingerpoint Detector) took off

> Channel: GitHub | Research date: 2026-11-06 | Repo: [github.com/Ikaleio/lm-detector](https://github.com/Ikaleio/lm-detector) | Site: [lm.ikale.io](https://lm.ikale.io)

### One line

It's a tool that identifies which model is actually behind an API. It has the model write about 300 integers between 1 and 355 off the cuff, then compares the digit distribution against a reference library. The algorithm idea isn't original to it; it comes from xqy2006/ModelTrace. It turned that idea into a web page, an `npx` command-line tool and SDKs in four languages, and it keeps the reference library updated as new models ship. Its growth came from the anxiety in the Chinese AI community over relay stations mixing in fake models, official models being dumbed down, and who is behind the Grok Bot. Users came mainly from word of mouth on linux.do and Chinese X, plus other tools integrating it.

### 1. What it is

- From the README: "A relay station or third-party API claims to provide a certain model, but the connection may actually be to a different one... Have the model write out about 300 integers between 1 and 355 off the cuff. Language models' 'random numbers' are not uniform; each model's preferred values, ranges and last digits are relatively stable." (Source: GitHub README)

- There are three ways to use it:
    - Web page lm.ikale.io: you can paste results manually or connect directly to an API. It runs purely in the front end, has a built-in proxy to get around CORS, and you can also deploy your own Cloudflare Worker with one click.
    - Command line: `npx lmfpd@latest -b … -k … -m …`, no install needed, with a live terminal interface.
    - SDKs: TypeScript, Rust, Go and Python. Results match the command line.

- Supports three protocols: OpenAI Responses, Chat Completions and Anthropic Messages. The reference library covers GPT, Claude, Gemini, Grok, Qwen, DeepSeek, Kimi and more. There is also an optional tokenizer probe.

- The README acknowledgments say: "Thanks to xqy2006/ModelTrace for the algorithm reference and some of the original data."

- The author, Ikaleio, is a Chinese university student and former OIer (per his ikale.io bio), with only 43 GitHub followers. His earlier projects are LiteLoaderQQNT-Markdown (149 stars) and monoize (a Rust LLM API gateway, 27 stars). He has no big-name backing.

- Free, MIT license, no paid tier.

### 2. Growth data

- Stars: 2026-10-16 is 350, 2026-11-06 is 1,690. Over the last 7 days it gained 493, over the first 7 days it gained 419, and the single-day peak was +132 on 11-03 (source: our GitHub repo_day / data card).

- From 10-18 to 10-20 it gained roughly 70–106 stars on each of three consecutive days, the first small peak. From 11-03 to 11-05 came the second peak (+132, +91, +66).

- The npm package lmfpd (command line) had about 19,633 cumulative downloads from 2026-09-16 to 11-05. Broken down by half-month: second half of September 4,232, first half of October 4,027, second half of October 8,883, and November 1–5 2,491. lmfpd-sdk has about 825 downloads since its late-October release (source: api.npmjs.org). The PyPI lmfpd package had 130 downloads last week (pypistats).

- Downloads grew along with stars, which suggests people aren't just starring the repo; they're actually running detections.

- Since the repo was created 2026-09-13, it has 130 commits and 4 SDK tags. It has many external contributors: Alice39s alone submitted large features such as the docs site, the Cloudflare Worker and the tokenizer probe, and limichange, xia0ne, dv868, 1zero224, h7ml and JohnsonRan have also sent PRs (source: GitHub issues/PR list).

### 3. Competitors

| Name | Form | Launched | Scale | How lm-detector differs |
|---|---|---|---|---|
| xqy2006/ModelTrace | Open source + GitHub Pages web app | 2026-08-27 | 2,047 stars. Mid-October it gained about 40 per day, which dropped to about 10 per day in November. The last push was 10-09. | The algorithm's origin. lm-detector took its idea and some of its data, and added direct API proxying, an npx command line, four-language SDKs, ongoing ingestion of new models, and Chinese, English and Japanese interfaces. The original project is cooling down while lm-detector is speeding up. |
| canarybyte/veridrop (veridrop.org) | Open source + website + relay-station good/bad rankings | 2026-05-05 | 725 stars. The site claims 12,914 relay stations tested in total and 154k reports. | Takes the route of "multiple protocol probes + Claude thinking signature verification + leaderboard", aimed at scoring relay stations. lm-detector does only one thing: identifying which model it is. It doesn't rank stations. |
| AetherCore-Dev/relay-radar | Open-source command line | 2026-04-02 | 150 stars, essentially stalled. | Does a full suite of relay-station monitoring, verification and cost tracking, with no statistical fingerprint library. |
| unclecode/modelprint | Open source, runs in the browser | About 2026-08 | 134 stars, stalled | Also does fingerprinting for OpenAI-compatible endpoints, but has no actively maintained reference library and no community. |
| HoneyMeta/ccfingerprint, rt22766/claude-skill-model-fingerprint, bi-boo/claude-model-fingerprint | Open-source Claude Code skills | Spring 2026 | 56–111 stars | They rely on capability probes or self-introspective Q&A and mainly check whether Claude is genuine. lm-detector uses a statistical task unrelated to semantics and covers every model family. |
| qqqqqf-q/ai-model-fingerprint | Open dataset | About 2026-06 | 68 stars | Same random-number distribution idea, but only a dataset covering 17 models, with no tool. |
| pasquini-dario/LLMmap | Academic open source (USENIX Sec 2025) | 2024-07 | 497 stars | Academic approach requiring 8 interactions and a classifier, not aimed at relay-station users. |
| okkmax.com, llmtest.cn (LLM Detector), bashigestudio/apitest, ProbeKit Lite, the ModelCheck from a Tencent Cloud article, and others | Web | 2026 years | Not verified | Mostly commercial sites that score and funnel traffic to relay stations, with opaque judging rules. Some even recommend relay stations themselves, so user trust is low. |
| Big-company features | — | — | None | OpenAI, Anthropic and OpenRouter have no public tool to verify "is the model I got the one I paid for". Cursor only shows a banner when it falls back, and the bill doesn't show the fallback (Cursor forum thread 2026-09-23). |

Conclusion: the space is crowded, and a batch of "relay station detection" tools appeared in the Chinese community starting in April 2026. But only ModelTrace and lm-detector have built a statistical fingerprint based on random-number distributions, which is unrelated to semantics and hard to target for faking, into a product anyone can use. lm-detector is taking over the original's position.

### 4. Where users come from (growth tactics)

1. **linux.do community (confirmed)** : On L Station, threads such as "GPT dumbing down is severe, how have you all dealt with it" (topic 2935964), "any tool recommendations for testing full-strength models", "gpt6.1sol severely dumbed down", and "is this the fingerprint of the official model?" (topic 3045874, domestic-alternatives section), plus "the pelican riding a bicycle test is overused, any other good test prompts", all have replies that directly post lm.ikale.io or the repo link (source: Jina search result snippets). Users were recommending it here as a standard tool for detecting dumbed-down or watered-down models, not because the author posted to promote it. L Station restricts AI scraping, so only search snippets were used, not full posts.

2. **Spread on Chinese X via the Grok Bot topic (confirmed)** : After 10-07 Musk said the Grok Bot could connect to other companies' models, "what model is the Grok Bot really running" became a trending topic. Around 10-28, accounts including @drjingle, @julmawiliams, @xiaomovps and CRYPTO MIAO posted results from lm.ikale.io, such as "tested through a US proxy, it's claude-opus-5.5 with 98.5% confidence; direct connection shows grok-4.5". These posts drove npm downloads to double in late October (8,883 in the second half of October, 2.2 times the first half). In addition, @aixparty (10-25) and @chuanliang's long post also covered it, the latter introducing both ModelTrace and LM Fingerpoint Detector.

3. **Scattered real use cases in the English community (confirmed)** : In Cursor forum thread 2026-09-23, "Safeguard fallback to Opus 4.8 is billed as the model you selected", the poster used it as evidence. The original reads: "I fed the three integer sequences into a model fingerprint detector (lm.ikale.io). It identified the output as claude-opus-4-8 with 98.5%." This highly contested thread brought it to English-speaking developers.

4. **Integrated or requested by other tools (confirmed)** : yetone/magpie (31k stars) has issue #2842, "add model fingerprint detection to magpie" (2026-10-18), and this repo also has a request for "Magpie plugin integration". An issue #2 in codex682api requests integration of ModelTrace. In the Pi coding agent ecosystem, two extension packages appeared, pi-model-trace-direct and @indexyz/pi-model-trace, which pull reference libraries directly from lm.ikale.io. Vibenate lists lmfpd as a "CLI for AI agents", and context7 also lists its docs.

5. **No traction on HN (confirmed)** : someone (csmantle) posted it on HN and got only 2 points.

6. **No sign** of ads, paid acquisition or star farming. The fork-to-star ratio is about 3.3% (55/1,690), and the issues are specific, such as asking why a local gateway can't be tested or whether relay stations can record questions to fake answers. They look like real users.

### 5. Why it took off

- **Timing** : 2026 Fall, Opus 5.5, GPT-6 Astra and Kimi K3 launched back to back, and relay-station traffic surged. A similar project, Folkbench (a relay-station leaderboard), hit 35 stars in 4,312 days, and veridrop has checked 13k relay stations. At the same time, three stories kept circulating: "official dumbing-down," "Cursor quietly falling back to Opus 4.8," and "who's behind Grok Bot." Lots of people want to know which model they're actually paying for.

- **The method is convincing on its own** : "ask the model to pick a random number and look at its preferences" can be explained in one sentence. The test has nothing to do with semantics, so relay stations can't easily fake it by changing the system prompt. The result is a confidence-ranked "most similar to" list that's perfect for screenshotting and sharing on X and forums, which makes it naturally viral.

- **Very low barrier** : paste three outputs into the webpage to get a result, or fill in base_url and key directly. A one-line `npx` command runs it with no Python environment and no probe to write yourself. ModelTrace's original webpage was limited by CORS for direct connections; lm-detector solved this with a proxy and a Worker.

- **The reference library keeps up with new models** : the commit history often has entries like "Enroll Claude Opus 5.5 low," "enroll Kimi K2.8," "Enroll Step 5 Preview." New models can be tested as soon as they come out, which is exactly what users need during a hot spell. ModelTrace hasn't been updated since Oct 9.

- **Community co-building** : outside contributors built a docs site, tokenizer probing, a multilingual interface, and SDK progress callbacks. The iteration speed far exceeds a solo project, and it makes the tool feel more like a "standard tool."

### 6. Differentiation, and whether that's why it took off

lm-detector's differentiation isn't in the algorithm (that comes from ModelTrace) but in three things:

1. productization: webpage, command line and a four-language SDK, all sharing one reference library;

2. the reference library is continuously updated and recalibrated;

3. other tools can call it (the SDK and public reference library; a Pi extension and magpie are both integrating it).

These differences are indeed why it could keep accelerating while the original cooled off: over the same period, ModelTrace gained about 10 stars a day, while lm-detector gained 60–130 stars a day. Compared with relay-station "scoring sites" like veridrop, another difference is that it's neutral: it doesn't recommend any relay station, so users on L-sites are willing to treat it as evidence.

### 7. Visibility

It's spreading in small circles. It's already a commonly used tool in the Chinese AI community (linux.do, Chinese X, the Grok Bot discussions), but no mainstream outlet or major English account has written about it, and it has only 2 points on HN. The project is completely free and has no revenue to speak of.

### 8. Copycats

- Direct derivatives: the Pi extension pi-model-trace-direct, @indexyz/pi-model-trace (which reuses its reference library), and 55 forks.

- Competitors with the same idea: ModelTrace is the original. There are also random-number fingerprint datasets like qqqqqf-q/ai-model-fingerprint.

- I haven't seen anyone copy it wholesale into a commercial site yet.

### 9. Takeaways

- **Borrow an original algorithm, build the product out fully, and maintain the data well** : the algorithm belongs to someone else, but "whose reference library is newest and whose tool is easiest to use" determines where users end up staying.

- **Ride closely on "trust anxiety" hot topics** : whenever a relay station blows up, official dumbing-down happens, or there's debate over who runs a product's backend, a detection result you can screenshot appears, and users will spread it for you.

- **Make the tool a component that other agents and tools can call** (npx, SDK, public reference library), capturing integration channels.

- **Neutrality is an asset** : in a track full of referral sites, not selling relay stations and not ranking them makes it easier for the community to cite it as evidence.

### 10. Still unclear

- The traffic to the lm.ikale.io webpage is unknown; it can only be roughly estimated from npm downloads.

- The specific popularity of the linux.do posts wasn't verified; the original couldn't be read due to site restrictions.

- Whether the method stays effective over time is also uncertain: issues already ask "has oai been broken into?" and "can relays record questions and forge answers?" If the upstream or relay stations deliberately change the random-number distribution, the tool will stop working.

- There's no day-by-day data for the star trend before 10-16 (the 9-13 to 10-15 period).

- The relationship with the ModelTrace author (collaboration or simply borrowing) is unverified.

### Sources

- GitHub repo and README: [github.com/Ikaleio/lm-detector](https://github.com/Ikaleio/lm-detector); author homepage [ikale.io](https://ikale.io)

- Our GitHub database (repo_day / repo_profile, 2026-10-16 to 11-06)

- npm downloads: [api.npmjs.org/downloads/range/2026-09-01:2026-11-05/lmfpd](https://api.npmjs.org/downloads/range/2026-09-01:2026-11-05/lmfpd); pypistats

- ModelTrace: [github.com/xqy2006/ModelTrace](https://github.com/xqy2006/ModelTrace)

- veridrop leaderboard: [veridrop.org/leaderboard](https://veridrop.org/leaderboard)

- Cursor forum: [forum.cursor.com/t/safeguard-fallback-to-opus-4-8-is-billed-…](https://forum.cursor.com/t/safeguard-fallback-to-opus-4-8-is-billed-as-the-model-you-selected-not-as-opus-4-8/172823)

- X: [x.com/drjingle/status/2115239798788980842](https://x.com/drjingle/status/2115239798788980842), [x.com/aixparty/status/2114269553584947434](https://x.com/aixparty/status/2114269553584947434), [x.com/chuanliang/article/2105107777878409355](https://x.com/chuanliang/article/2105107777878409355)

- linux.do: [linux.do/t/topic/2935964](https://linux.do/t/topic/2935964), [linux.do/t/topic/3045874 (search summary)](https://linux.do/t/topic/3045874（搜索摘要）)

- magpie issue: [github.com/yetone/magpie/issues/2842](https://github.com/yetone/magpie/issues/2842)

- Pi extension: [pi.dev/packages/pi-model-trace-direct](https://pi.dev/packages/pi-model-trace-direct)

- HN: [news.ycombinator.com/item?id=49824813](https://news.ycombinator.com/item?id=49824813)

This article was written by AI from public data and the data we collect daily. Numbers are sourced, and inferences are labeled.