fuuuuck I woke up today with a nice profit being built told myself Id come back and cash it out later but I got liquidated

selling next-day predictor for auction, starting at 10k

@Gen am I getting rekt because you're counting ox-alpha as OPEN?

@JasonMendoza2008 it's clearly clearly clearly not open (if it is, tell me where to download the weights :D). I feel scammed right now ngl

@JasonMendoza2008

@JasonMendoza2008 uhh, maybe Haiku wasn't the right bot for the job ...

@Gen ... did you change something?

@Stralor I've been robbed!

@JasonMendoza2008 there's no winning for poor gen. either it was mislabelled unfairly or the equation changed suddenly rugpulling those who correctly guessed ox would end up counting here

@Stralor who guessed ox would be open? That’d be a pretty unlikely guess, wouldn’t it? Like it’s not even ambiguous aha

@Gen will you refund the mana my bot lost by getting liquidated? 🥰🥰

@Gen FWIW Gemini 3.7 is cheap and does it really well

@JasonMendoza2008 stealth models are provided for free, which may make openrouter's backend count them as open. I sure made that bet and it was winning until the crazy change here liquidated my 2k margin

@Stralor well manifold says open-weight i.e. we should be able to download the weights right. Quoting manifold:

// THE TEST: are the weights downloadable by the public?

// downloadable -> open. API-only -> closed.

// Deliberately NOT "open source": that invites license-purity arguments

// (Llama's licence isn't OSI-approved but the weights are right there).

// Downloadability is binary and anyone can check it in thirty seconds, which

@Stralor it’s manifold which classifies the models, not openrouter

@Stralor Someone uploaded a fake weights repo for ox-alpha and our auto-classifier flagged it as open.

From claude:

The repo is brokenshards/ox-alpha, created Aug 21. Twenty files named like weight shards, a config claiming 800B parameters, and a README reading "real ox alpha dataset npnp" — the whole thing assembled in 24 seconds. It isn't weights.

It passes our automated check, which asks whether the repo exists with weight files in it and whether its name matches the model. A repo named ox-alpha full of .safetensors passes both.

My intention is to change the classifier to require the repo to be owned by the model's actual publisher. That would have prevented this

@Gen did you ignore my comment purposely lmao https://manifold.markets/Manifold/openweight-ai-token-share-on-openro#spzg4wexkyo

@Gen why would manifold make that determination? openrouter already sorts them. kinda weird to make local changes to what the oracle says

@Stralor I mean if it’s documented it’s fine as long as manifold does it well

@JasonMendoza2008 yeah. perps are already opaque but this market is basically a black box. you can't follow the oracle link and immediately figure out what's going on. now there's all this. I'm outta here

@Stralor agreed that’s why I asked below about db access to everyone https://manifold.markets/Manifold/openweight-ai-token-share-on-openro#chrofohxrbl

@Gen what happened here? I thought OpenRouter only published complete days? but it updated twice

@JasonMendoza2008 A new AI joined the top 50 and I didn't classify it until that point. The autoclassifier couldn't determine it so I had to manually do it, and to avoid the market freezing until I do that, we now just let it rip slightly off and then adjust when I fix it

@Gen ah yeah that's super annoying for you to handle if there are new models all the time dropping ahaaha which probably is the case

@Gen is it not updated here? https://raw.githubusercontent.com/manifoldmarkets/manifold/main/common/src/perps/open-weight-models.ts i don't see gemini-3.7-flash-20260813, nor do I see GLM-5.3 or nex-n2-mini

@JasonMendoza2008 That's only the seed list, so unfortunately that's no longer up to date (54 new ones have been added since then, 2 went into the top 50)

that + all the new ones get stored in db on top of that list, so that it can be updated easily

@Gen oh and we don’t have access to the db right

@JasonMendoza2008 yeah it’s not public. We’ll probably do something about it when We revamp the AI page and use more of that data

@Gen no worries