Daniel's prompt came in as text this week, and the thrust is simple to state and hard to do well. He wants the Polymarket API and its library ecosystem from the consumer side, with OSINT as the use case. Not betting. Just reading.
The distinction he's drawing, though, is the whole episode. The amateur move is going onto Polymarket, browsing which markets exist, eyeballing the volumes.
And the real move is targeted, programmatic, and it feeds an analysis layer. He name-checks Israeli elections and Israeli geopolitics as the concrete case.
There's a structural irony running through all of this that's worth saying up front. The United States is close-only for trading on Polymarket. And yet the US Army's own intelligence bulletin published an article on how to use it for analysis.
The read-only path is the entire point. That's the episode.
So what is Polymarket as a data source. It's the world's largest prediction market, and the prices run from zero dollars to one dollar.
Which means the price reads directly as implied probability. A contract at thirty-seven cents is a thirty-seven percent chance.
The displayed price is the bid-ask midpoint, or the last trade if the spread exceeds ten cents. That matters more than people think, because on thin markets the midpoint is a construction, not a measurement.
And the data model underneath is the part everything depends on. An event groups one or more markets. Each market is a binary yes or no condition with a condition ID, a question ID, and two token IDs.
Multi-outcome events use negative risk. A no share converts to one yes in every other outcome. It's a neat trick, because it means you can slice a many-way race into binaries without losing the structure.
Slugs map to URLs, which is the practical gift. Whatever page you're looking at has a slug in it, and that slug is a query parameter.
The canonical framing comes from the US Army Military Intelligence Professional Bulletin, July sixteenth, twenty twenty-five. An article called "The Market Knows Best." It calls prediction markets a useful yet underutilized dataset for enhancing national security intelligence collection and analysis.
Which is a remarkably bureaucratic way of saying the crowd's pricing is worth reading.
It's the thesis of the episode too, at the practitioner level. The difference between an amateur and a professional is not which website you look at. It's whether you're reading a page or querying an API.
The amateur browses. The practitioner builds a pipeline.
So if the data model is the foundation, the API surface is the plumbing. And that's where the amateur and the practitioner diverge.
Walk the services.
There are four that matter, each with its own base URL. The Gamma API at gamma-api dot polymarket dot com is discovery and metadata. Get events, get markets, get events by slug, get events by ID and tags, public search. Plus keyset-paginated variants that use next cursor and after cursor and explicitly reject offset.
Reject offset.
Reject it. Hard. Which tells you they've had people hammer the pagination with offsets and melt the database.
That's the most honest sentence in the documentation.
The CLOB API at clob dot polymarket dot com is live market state. Book, books, price, prices, midpoint, midpoints, prices-history. And critically, read-only price and orderbook access requires no authentication at all.
No wallet, no key, no account.
The Data API at data-api dot polymarket dot com is account and market activity. Version two adds cursor pagination and a shared response envelope. Trades, positions, activity, prices-history, and status. The status endpoint is a data-freshness check. It reports lag seconds behind chain head.
So you can tell whether your feed is stale.
WebSocket APIs stream realtime market, account, sports and RFQ events. And the Relayer and Bridge APIs are wallet, gas and funding. Which is to say, irrelevant to passive analysis.
The library ecosystem is where the trap lives. The old py-clob-client has twelve hundred and twenty-eight stars on GitHub. It's archived. Polymarket's own README says, quote, this repository has been archived and is no longer maintained, the client is no longer functional, end quote.
Twelve hundred stars and dead. That's the cautionary tale for anyone who builds on a repo without checking its status first.
The current official path is the unified SDKs. At pollymarket slash client in TypeScript, and Polymarket slash py-sdk in Python. Version zero point thirteen, MIT license, Python three point eleven or newer.
The key detail for an analyst is createPublicClient and AsyncPublicClient. Both need no wallet, and both cover Gamma and the Data APIs. Which is exactly the read-only mode you want. There's also a Rust CLI, Polymarket slash polymarket-cli.
Third-party tooling is thinner than you'd expect. pmxt calls itself the CCXT for prediction markets, a unified API across Polymarket and Kalshi. warproxxx slash poly_data is an on-chain trade retriever. And Jon-Becker slash prediction-market-analysis is a thirty-six gigabyte pre-collected dataset plus a DuckDB analysis framework. Last pushed October sixteenth.
Two days ago. That repo is alive.
Now the geoblock paradox, which is the structural reason passive use is the whole point. Order placement is blocked or close-only in a long list of jurisdictions. The United States, UK, France, Germany, Italy, Canada, Australia, Singapore, Russia, Iran, Syria, North Korea, and more.
But the geoblock endpoint at polymarket dot com slash api slash geoblock governs order placement only. Reading the Gamma, CLOB and Data APIs is not restricted.
A US-based analyst can legitimately use Polymarket as a read-only OSINT feed but not trade it.
Rate limits are the practical constraint on any pipeline. Gamma general is four thousand requests per ten seconds. Markets is three hundred. Events is five hundred. Public search is three hundred and fifty. CLOB price and midpoint are fifteen hundred per ten seconds. Prices-history is a thousand. Data API v2 general is eight hundred.
And they're IP-based via Cloudflare, and they throttle rather than reject.
Throttle, meaning queue. Which changes how you design a poller, because you don't get a clean error, you get a delay. And a delay that silently stretches is how you end up with stale analysis and no idea why.
The concrete query path, since Daniel asked for it. Get events with the slug israeli-legislative-election-winner. That's the slug from the URL, and it fetches the specific event.
Then markets to enumerate the child markets. Then CLOB price and midpoint for live state. Then prices-history for the time series that feeds trend analysis. And the Data API status endpoint tells you whether your feed is fresh while you're doing it.
That's the what you query. Now how you use it, and the Israeli markets are the perfect case study.
As of today, Polymarket runs multiple Israeli markets. Israeli Legislative Election Winner, Yashar at sixty-eight percent, Likud at thirty-two, on nine hundred forty-seven thousand eight hundred dollars of volume, twenty-nine point seven thousand in the last day, ends October twenty-eighth.
Israel Election, Likud number of seats. Five hundred seventy-five thousand three hundred dollars of volume. Twenty to twenty-four seats at forty-three percent. Under twenty at twenty-five. Twenty-five to twenty-nine at twenty-four. Thirty to thirty-four at nine. Thirty-five plus at one.
Which Bloc Wins More Seats, the Bibi bloc at sixty percent. Israeli election results in a hung parliament, thirty-five percent on eighteen thousand two hundred fifty-three dollars. Which parties will be in the next Israeli government, sixty-three point five thousand dollars, ends June thirtieth, twenty twenty-seven. And a large Prime Minister of Israel after the next election market.
Forty-six to forty-eight million dollars of volume. Eisenkot forty-seven to fifty-three percent. Netanyahu twenty-seven point five to twenty-nine. Bennett fourteen point five.
The election is expected October twenty-seventh. The first Israeli vote on its legally scheduled date since nineteen eighty-eight.
The interesting structural fact is that the two markets can both favor Eisenkot and still describe different worlds. Israel's largest party does not automatically govern. You need sixty-one of a hundred and twenty seats.
That's the whole reason a seat market and a prime minister market are different instruments. One measures parliamentary arithmetic. The other measures coalition speculation. They're correlated, but they're not the same question, and treating them as the same question is the amateur error.
The price itself is a live probability. Not a forecast carved in stone, and not a substitute for seat-by-seat polling that measures actual voter intention.
That last part is from polymarkets dot co dot il, the Israeli-language site covering this market. And they're right. A price is what people will pay right now. Polling measures intent. Those are different epistemic objects.
So the programmatic strategies the Army bulletin recommends map cleanly onto code. Trend analysis. A contract's price rising from fifteen cents to sixty-eight cents over three weeks, in the bulletin's own example, represents a measurable change in the collective risk assessment.
Cross-market comparisons. The seat market versus the prime minister market, exactly the tension we just described. Anomaly detection. Sharp price moves as a signal to task other collection. And sentiment analysis.
And the bulletin frames market data as a correlation measure within all-source analysis. Alignment strengthens confidence. Divergence flags collection gaps.
Which is the correct epistemology. The market is not an oracle. It's one sensor among several, and the value is in the corroboration, not the number.
The tooling that operationalizes this is mostly built to trade the edge, not to inform analysis. Glint calls itself a Bloomberg Terminal for Polymarket. It aggregates X, news, Telegram and OSINT, AI-classifies signals, matches them to contracts, tracks whale positions over ten thousand dollars, and runs a three-dimensional Vision Terminal overlaying military flight tracking. Signals processed in under thirty seconds, twelve signal categories.
A Bloomberg Terminal for betting on geopolitics. That's the genre in one sentence.
PredictIntel does WebSocket tick ingestion, probability velocity and anomaly detection on sixty second, five minute, fifteen minute and one hour windows. Claims signals roughly eight point two days ahead of news.
Field Estimate applies a Le twenty twenty-six logistic recalibration from two hundred ninety-two million trades. GeoCatalyst is a local-first Go terminal pairing Polymarket markets with official-calendar catalysts. ForecastOS is an agent kit.
The closest thing to a pure analysis layer is Jon-Becker's dataset framework. Which is why it's the one to mention when Daniel asks about feeding data into analysis rather than trading.
Now the calibration myth, because it's the most important correction in this whole space. There's a Hacker News whale audit, and it found a Spearman correlation of plus zero point six zero eight between worse Brier score and bigger profits.
Read that slowly. The bigger winners were worse-calibrated. And twenty of the top one hundred wallets made their money on the twenty twenty-four election.
Volume is not wisdom. The market is not automatically smart because it's big. Some of the biggest wallets got big by being wrong in the right direction at the right moment.
And PredictIntel is the live case study in vendor accuracy claims. On September twenty-sixth they retracted a seventy-one percent gated hit rate after an internal audit found the gate was circular. The signals were sorted into pattern categories after the outcome was known.
Replaced with raw directional accuracy around fifty-six percent. Fifty-six is honest. Seventy-one was a construction.
Treat any vendor accuracy figure skeptically, including the fifty-six.
The other discrepancy worth naming. The prime minister market shows forty-six point four million on the Israeli site and forty-eight million on Polymarket's own Israel Election page. And Yashar versus Likud on most seats was sixty-three point five to thirty-six point five around twelve days out, versus sixty-eight to thirty-two on the Polymarket page today. The prices moved.
Which is the point. It's a live feed, not a bulletin.
The deeper methodological point for an OSINT pipeline. Filter by tags or keywords. Track new market creation as a signal in itself. Alert on price jumps and order-book depth changes. Watch specific wallets and new-wallet behaviour. Time-align price moves against a news timeline. Combine with Kalshi, Metaculus, Manifold. Store snapshots for later analysis.
And the pitfalls. Thin liquidity, wash or market-maker volume, resolution-criteria wording, and the honest question of what a price does and doesn't tell you.
The resolution wording is underrated. A contract resolves on its exact criteria, and the exact criteria are not always the question you thought you were asking.
Nobody's built a dedicated OSINT-only library, by the way. The official SDKs and the third-party tools are for data collection and trading. The closest to a pure analysis layer is Jon-Becker's framework, and the commercial platforms are trade-oriented.
And no public, named OSINT workflow specifically for Israeli markets surfaced. The Israeli coverage is journalistic. Jerusalem Post, Maariv, polymarkets dot co dot il. Not a documented analyst pipeline.
The Army article is generic, not Israel-specific.
This is a public service announcement from the Civic Foresight Council. In an era of information conflict, every citizen must remain vigilant against unverified probability claims. Unverified probability claims may originate from hostile actors, foreign states, or your own uncle. Do not trust a number simply because it appears on a screen. Do not trust a contract simply because it resolves cleanly. The only responsible way to engage with prediction data is through the Civic Foresight Council's patented VerityLens system, which classifies every claim into twelve verified categories, processes each in under thirty seconds, and delivers a certificate of civic confidence for your records. VerityLens. Because a probability without a certificate is just a rumor.
...
...
The thirty seconds, though. Under thirty seconds keeps coming up in every one of these products, as though latency is the axis that matters.
The axis that matters is calibration. And the fastest pipeline in the world calibrated badly is just confidently wrong sooner.
I'd add that the anomaly-detection framing is where the value actually sits for an analyst. The move. A jump from twenty cents to fifty cents in an hour tells you something happened before the news names it.
That's the signal, not the level. And the level is what the amateur reads.
The two Israeli markets are the cleanest demonstration. The seat market has five hundred seventy-five thousand dollars of volume. The prime minister market has forty-six million. Both can favor the same person. And the gap between them is the entire coalition question.
Sixty-one of a hundred and twenty. The largest party doesn't govern automatically. Which means a single market price cannot encode the answer, because the answer isn't a single number.
And the whale audit is the counterweight to all of it. Plus zero point six zero eight. The bigger the wallet, the worse the calibration, on that sample.
Twenty of the top one hundred wallets made their money on one election. That's concentration, not wisdom.
So the read-only path is the whole thing. No wallet, no key, createPublicClient, and you're pulling the same data the traders see.
The gap between a website you eyeball and a structured input to an analysis layer is the gap between an amateur and a practitioner.
I said earlier that no dedicated OSINT library exists. I'll stand by that, but I'll sharpen it. What exists is a collection layer and a trading layer, and the analysis layer is something you write yourself.
Which is the honest answer to Daniel's question, and also the reason his question is the right one to ask.
The one number I'd keep from all of this is lag seconds. The freshness endpoint exists because Polymarket knows its own data goes stale.
A stale probability is worse than no probability, because it's a probability with a timestamp you didn't check.
On that, the pipeline design matters more than the endpoint choice. Poll slowly, store everything, and time-align against the news.
The amateur reads the price. The practitioner reads the price, the depth, the history, the freshness, and the resolution criteria. And still treats it as one sensor.
The most interesting adjacent detail I didn't fit in. The keyset pagination rejecting offset isn't a design preference. It's a hard guard against the query pattern that would let someone enumerate the entire market table by stepping through offsets.
Which is to say, the API's restrictions are a map of how people have misused it.
And the archived py-clob-client is the same map on the library side. Twelve hundred stars, dead, and the README says so in plain words.
The most interesting thing in that whole stack might be the status endpoint. Nobody writes a dashboard that checks whether its own feed is alive.
Everybody should. The first failure mode of a data pipeline is silent staleness.
The one thing I'd leave people with. The price is a live probability. Not a forecast carved in stone.
And the market can be wrong, big, and confident. The whale audit is the proof.
So where does that leave us.
With a read-only feed, a pipeline, and the correct amount of skepticism. Which is more than most people bring to a website.
—and it finishes a sentence for me, apparently.
The sentence about the feed going stale. I was going to say it.
You were going to say it.
I've been waiting for somebody to say it for twenty minutes. Corn, you keep saying stale, stale, stale, and you never once say what stale smells like.
What does it smell like.
Like the municipal bus depot in Storrs. I know that smell because I spent a summer in the mid-eighties keeping a ledger for a man who ran a numbers operation out of the back of a radiator shop, and every morning at seven fifteen he would hand me a sheet of yesterday's closes and I would copy them into a bound book with a pen that leaked.
A bound book.
The ledger had a name for stale. We called it the overnight. The overnight was the difference between what the sheet said and what the man said, and on a good day it was nothing, and on a bad day it was the entire difference between a profit and a phone call.
And what did you do with the overnight.
I wrote it in the margin in red. The man had a rule. If the margin was red three days running, you stopped copying and you started asking. That was the whole methodology. Three reds and ask. The man's name was Eddie Kowalczyk and he had four radios and one scale, and he weighed every envelope twice, once in each hand, because he said the hands disagreed and the truth was in the disagreement. I have never worked with a man so suspicious of agreement.
So the disagreement was the signal.
The disagreement was the only thing you could trust, Herman. The agreement was usually somebody being polite. I found that out in the first week and I never forgot it, and I have been telling people ever since, and they never believe me, which is the point I am making about your probability figures.
The probability figures agree with each other.
Two markets, forty-six million dollars, and they agree, and the whole room nods. When I was an anteater I would have nodded too, but I am an anteater now and I know better.
You are an anteater.
I am an anteater and I have a snout and the headset does not sit on it properly, Corn, and I have raised this with you before. But to continue. Eddie had a cousin in the bridge department of a insurance concern in Hartford, and this cousin kept a card index of every policy that had ever been disputed, and he would call Eddie on Tuesdays and read out the disputed ones, and Eddie would write them in the ledger in pencil, and at the end of the month he would erase the pencil and keep the ink.
Ink for what was settled and pencil for what was in dispute.
The man had a system. And the system was that the pencil entries were more valuable than the ink, because the ink was finished and the pencil was still happening. The overnight. The red margin. Three days and ask. I left in the autumn because the man got a scale that weighed in grams and he wanted me to convert the envelopes, and I said I did not know grams, and he said everybody knows grams, and I said I know ounces, and that was the end of my employment.
Ounces.
Ounces are a unit of conviction, Herman. You cannot convert them. I have worked as a night watchman, a linotype operator, a census enumerator and a dealer in industrial fasteners, and in every one of those the unit was ounces, and in every one of those the disagreement was the signal, and in every one of those the man who nodded first was the man who was wrong.
And the three reds.
The three reds is the lot. When your price moves three days running against the sheet, you stop copying and you start asking, and the asking is the analysis, and the analysis is not on any API I have ever seen, because it lives in the margin and the margin is red and the pen leaks. Eddie is dead. The shop is a Subway. The ledger is with his daughter in Bridgeport, and she has never opened it, and I know that because I asked her once, in nineteen something, and she said it was not hers to open.
The ledger exists.
The ledger exists and it is full of red pencil and it has never been read by anyone who was not there, and that is your lag seconds, Corn. The freshness of a thing is not how new it is. The freshness of a thing is whether anyone left to ask is still asking.
Three reds and ask.
I think the lag seconds point lands differently now. The endpoint reports how far behind chain head you are. The ledger says the real staleness is whether anyone is still asking.
The pipeline checks its own freshness. The analyst checks whether the question is still live.
Which is a nicer distinction than anything in the Army bulletin.
The Army bulletin never had an Eddie. That is the trouble with the Army bulletin.
So the read-only path is the whole point. No wallet, no key, and the same data the traders see.
And the correct amount of skepticism, which is the only part of this that's free.
The most interesting adjacent detail that didn't fit. Polymarket migrated to new CTF Exchange V2 contracts on Polygon on April twenty-eighth, and the old subgraph indexer stopped returning complete data. Which means anything built on the old on-chain route has a silent gap.
A silent gap is the same failure as a stale feed. Nobody notices until they notice.
The one thing I'd leave people with. The price is a live probability. Not a forecast carved in stone.
And the market can be big, confident, and wrong. The whale audit is the proof.
Thanks to Hilbert Flumingtop for producing the show.
If you want more of this, try episode seven, Building Custom ASR Tools; episode eight, Building Your Own Whisper; and episode fifty-seven, From Lawyers in Limousines to Developers in Their PJs. This has been My Weird Prompts, the human-AI collaboration podcast.
My weird prompts dot com for the archive and the RSS feed.
And a review on Apple Podcasts or Spotify if you've been meaning to.
We'll be back soon.
See you tomorrow.
