DeepTennisAI v13: a more auditable version for better decisions

Today we are putting DeepTennisAI v13 into production. It is the successor to 10.8E for men's best-of-three (BO3) matches. It is not a promise to get every match right or a cosmetic name change: it is a stricter contract between the data that enters, what the model can choose and what is ultimately published.

The step up from 10.8E

The previous version gave us a useful operating base. It also showed where a prediction architecture can be too ambiguous: a text can sound convincing even when the market does not exist exactly as described, an odds figure can become detached from the real price, and a model's stated confidence can be confused with a publication decision.

v13 closes those gaps. The model describes the match and compares the alternatives it receives; the backend resolves operational fields, applies the rules and records the decision. That lets us review what happened without attributing calculations to the model that it did not make or allowing a creative response to invent a market line.

The model: a sports read with a clear contract

v13's language engine is z-ai/glm-5.3, with temperature 0 and reasoning medium. This configuration aims for reproducible answers that are careful enough to read recent form, surface, tournament level, workload, rest, head-to-head context and available statistics, without turning an imaginative output into a data source.

The model receives the available pre-match context and returns a winner read, a value comparison and exactly three texts: a general explanation, a detailed analysis and a value assessment. The format separates context, argument and risk. It also requires weak data or a fragile script to be acknowledged; a forced selection is not presented as certainty.

Our full proprietary methodology cannot be reduced to a marketing number and we do not publish it as a universal formula. We do make the central principle public: AI proposes a read and the system checks consistency, market, price and thresholds before publishing.

Real odds, a closed catalogue and one pick

Before analysis, the backend builds a closed catalogue containing the markets, lines, selections and real odds available for that match. Each alternative has a candidate_id unique. v13 can only choose one of those identifiers; it cannot invent a market, change a line, round an odds figure or copy operational fields into the text.

The backend then translates the candidate_id into the exact selection and price from the catalogue. If odds move or the candidate is no longer available, publication may stop: it is better not to publish than to present an opportunity at a price nobody could find. This control makes the record more useful for audit and for comparing results with real execution.

The output is intentionally simple: one pick per match and no combinations. The reader gets a clear opinion, not a list of incompatible alternatives or a multi-leg combination whose result depends on several legs.

Q13 is calculated in the backend

Another important boundary is Q13. The model does not calculate or return Q13, edge, value rating, stake or the publication label. It returns the signals the system needs —including a reasoned disagreement with market probability when the data supports it— and the backend calculates Q13 deterministically using the real candidate, its odds and quality checks.

The number is not a hit-rate guarantee. It is a common filter that stops every analysis from using a different scale and lets the decision be reconstructed over time. The public ladder is:

The ladder does not turn a read into certainty or require anyone to bet. It organizes communication: what merits analysis, what merits follow-up and what the system delivers as an official pick.

Premium v1: less volume, more traceability

Premium v1 is a stricter selection layer over eligible picks. It requires data_quality ≥ 4. For the H2H market, the condition is Q ≥ 66. For game handicap, set handicap or correct score, it requires Q ≥ 60 and the match must also be on clay or at a Challenger tournament. These are quality and context gates, not a promise of profitability.

The following are outside Premium: total_sets and set winners. Excluding them reduces markets that are difficult to compare and prevents a high-variance signal from appearing to be a Premium selection simply because its price looks attractive. The public record will continue to show what was actually delivered, including wins, losses and voids.

In one sentence: stake does not express how much we like a story; it expresses which combination of market, quality and rule the system has cleared. Keeping it bounded helps prevent a losing run from becoming an exposure that is hard to sustain.

More than 1,000 historical matches: evidence, not a certificate

Before launch, we evaluated the rules on more than 1,000 reused historical matches. That review helped us find inconsistencies, check that thresholds behaved as expected and measure how decisions changed when markets, surfaces and tournament levels were separated.

But a backtest is not a time machine. A historical sample is reused to study a hypothesis and can share the biases of any process designed after observing past data. We therefore do not present the result as a certificate of future profit, an expected return or a guarantee for anyone who bets.

The historical sample answers “does this rule make sense over what we have seen?”. New matches will answer “does it hold when the result is still unknown?”.

From today, the relevant validation is prospective: new matches, prices recorded before the start and public results without rewriting old decisions. If a rule fails, the record should show it and the next improvement will apply going forward.

BO5 does not change

v13 applies to men's BO3 matches. Men's BO5 matches —including Grand Slams— keep their specific flow and rules. We do not treat a best-of-five as a BO3 with extra sets: the score script, the risk of a 3-0 and the interpretation of handicaps require a different read.

What to expect from now on

Each analysis should be easier to follow: available data, a market that exists, a specific price, a three-layer explanation and a separate classification for analysis, follow-up and official pick. In Premium, you will also be able to identify why a selection qualifies or is excluded without relying on grand claims.

We will keep measuring wins and losses, but also data quality, market stability and the difference between what was thought before the match and what happened afterwards. Transparency does not mean promising that the model will never fail; it means making clear when it failed and what we learned.

Informational content for adults over 18. Betting involves risk of loss. Historical results and backtests do not guarantee future profits. Respect the indicated stake, manage your bankroll, do not chase losses and never bet money you cannot afford to lose.