What Is OddsFlow? From Model to Telegram Agent

The One-Sentence Answer

OddsFlow is a football analysis system that settles every position it publishes against the real final score, in public, with a timestamp — and its working end is a Telegram agent, @Oddsflowteam_bot, that runs the whole loop with no person in the middle.

Everything below is detail on that sentence. There are three layers — a model, a filter and a distribution surface — and the principle joining them is that each layer produces something an outsider can check.

Layer One: The Model

The model reads odds movement, match statistics and market state across leagues, continuously. It produces its own probability estimate for candidate positions and compares that estimate against what the price implies. The gap between the two is what the system calls expected value.

The important property of this layer is not its architecture. It is its volume. The model settles candidates on the order of a thousand a day — every one scored against a real final score, whether or not it was ever published. That is the dataset the rest of the system calibrates on, and it is the difference between a model that learns from outcomes and one that learns from the outcomes somebody chose to remember.

Volume also creates the problem the next layer exists to solve. A thousand scored candidates a day is not a thousand things worth telling anyone about.

Layer Two: The Filter

Roughly one candidate in twenty-two clears the filter and gets published.

One clarification, because otherwise the arithmetic from the previous section misleads. That ratio is not one in twenty-two of everything the model scores. Scoring is the widest stage, and most of what gets scored never reaches the filter at all — dropped at earlier gates for having no live market of the right kind, data too thin to act on, or a price that moved before the candidate was complete. One in twenty-two describes the last stage only, and its denominator is the candidates that survived everything upstream. On any given day that pool is far smaller than the scoring volume, and some days it is empty.

The ratio is the product. Not the model — models of this kind are not rare — but the willingness to discard the other twenty-one. A candidate can carry a large expected value figure and still be dropped, because a wide gap between model and market is as often a symptom of thin data or an unusual market as it is a sign of opportunity.

The visible consequence is silence. When nothing clears, nothing goes out. That is the filter working as specified, and it is the single hardest property for a signal service to hold onto, because a content schedule and a selectivity threshold pull against each other every day of the week.

Layer Three: Distribution

Three surfaces, all timestamped.

• @Oddsflowteam_bot on Telegram. The full signal feed. Each signal arrives as a structured card: league, fixture, market and side, price at publication, expected value, and for in-play signals the elapsed match minute. The field-by-field reading is in How to Read an OddsFlow Signal Card: Every Field Explained.
• X at @Oddsflow_Nat. Signals and daily settlement summaries, posted publicly, losses included.
• Threads at @oddsflow.ai. The same record on a second public surface.

Publishing to public channels rather than only inside the bot is a deliberate structural choice. A public channel is append-only in practice: an edit or a deletion is itself visible, and readers who saw the original are witnesses the publisher does not control. That property is what turns a claim into a record.

Settlement follows one stated formula — win rate = won ÷ (won + lost + half), with void excluded from the denominator and half counted in full — and no figure produced by that formula appears in this article, deliberately. Numbers in articles go stale, and a stale number is indistinguishable from a cherry-picked one. The live figures in the bot are the reference; the public feeds are the cross-check. The method is written out in How We Count Our Record: The Formula and the Timestamps.

The Seven-Step Loop

Layers describe the structure. The loop describes what actually happens, continuously.

1. 1.Analyze data — odds movement, match statistics and market state, across leagues, without pause.
2. 2.Read the news — lineups, injuries and context, folded in before the market finishes adjusting.
3. 3.Find the signal — the filter keeps roughly one candidate in twenty-two.
4. 4.Act on it — qualified signals are executed and published in the same minute, before the event.
5. 5.Audit itself — every published signal settles against the real final score.
6. 6.Keep improving — settled outcomes feed calibration.
7. 7.Publish everything — signals and daily summaries, timestamped, on public channels.

Two steps carry the weight. Step 5 is what makes steps 1 through 4 checkable at all. Step 7 is what makes step 5 checkable by someone outside the system. Remove either and what remains is a model with a publishing habit.

ClawSportBot and OddsFlow

This site describes an agent network: autonomous agents that observe, decide, act, and are audited against reality rather than against their own confidence. Most of that is architecture — the interesting kind, but still the kind that lives in documents.

The Telegram bot is the part that already runs. It is the living prototype of the agentic architecture the rest of this site describes, at the smallest useful scale: one domain, one agent, one loop, operating today, with the audit step wired in from the start rather than bolted on later as a reporting feature.

The relationship is worth stating plainly in both directions. The bot is not a demo of the network; it is one agent doing one job end to end. And the network is not a bigger version of the bot; it is what the same discipline looks like when many agents have to coordinate and be held to a shared standard of evidence. The prototype is what makes the larger claim testable — which is the same argument this system keeps making about everything else.

Where to Start

1. 1.Open @Oddsflowteam_bot on Telegram. New users get a three-day full trial — the complete feed for the window, not a reduced tier.
2. 2.Turn notifications on. In-play signals are time-sensitive by construction, and a card read an hour late is a card you cannot evaluate honestly.
3. 3.Spend the window watching cards settle rather than acting on them. Three days is enough to follow signals from publication through to a real final score.
4. 4.After the trial there is a free daily pick, and continued access runs on tokens earned through daily check-in and /follow tasks. The mechanics are in OddsFlow Bot Guide: Free Trial, Tokens and VIP Access.
5. 5.Cross-check against the public feeds. Scroll back past today and look for the losing posts. If they are not there, none of this article is true.

What OddsFlow Is Not

• It is not a tipster channel with a model attached. The filter, not the volume, is the product.
• It is not a guarantee of anything. A selective system produces long streaks in both directions, and a short window cannot separate a good process from a lucky one.
• It is not asking to be trusted. Every claim here is either checkable in the bot or checkable on a public feed, and the ones that are not — the internal candidate volume, for instance — are stated as ours rather than presented as verified.

That last distinction is the whole design. Start with the trial, follow a dozen cards through to settlement, and decide from the record rather than from the description.