You’re right, the prompt needs to make the AI unde...

You’re right, the prompt needs to make the AI understand that a 3-set match is not automatically an Over and that the engine should not be collapsing the total-games prediction into “2 sets = Under, 3 sets = Over.” That was one of the problems your previous rebuilds were trying to fix.

Use this:

Perform a full audit of the current match-length, set-count, and total-games interaction.

The recent rebuild appears to have increased Under selections, but the main concern is not the number of Unders. The concern is whether the engine has developed a structural bias where it is overvaluing straight-set outcomes and suppressing legitimate 3-set Under scenarios.

A critical modeling rule needs to be enforced:

A 3-set match MUST still be able to produce an Under 22.5, Under 23.5, etc. selection when the simulated score distribution supports it.

The engine cannot operate with the hidden assumption:

2 sets = Under
3 sets = Over

That is incorrect tennis modeling.

Examples of valid 3-set Under paths:

• 6-3, 3-6, 6-2 = 26 games (Over example)
• 6-4, 6-3, 6-4 = 29 games (Over example)
• 6-2, 6-3, 6-1 = 18 games (Under example)
• 6-4, 6-2, 6-3 = 25 games (depends on line)
• 7-5, 6-3, 6-2 = 27 games (depends on line)

The model must evaluate the actual score tree, not simply classify by number of sets.

Current output issue:

Example:
Match forecast:
2 sets / Under 2.5 sets (63.4%)

Totals:
Under 22.5 (53.7%)

Under probability composition:
53.2% from straight-set paths
0.5% from three-set paths

This suggests the engine is barely allowing 3-set paths to contribute to the total-games forecast. Audit whether the total-games PMF is incorrectly inheriting the match-length decision instead of independently calculating from legal score states.

Required fixes:

1. Separate match winner, set count, and total games into independent but connected distributions.

The flow should be:

Player strength/service/return inputs
↓
Set probability model
↓
Match tree simulation
↓
Actual score distribution
↓
Total games distribution

Do not force totals from the match-length forecast.

2. Allow the model to select:

• Under with high 2-set probability
• Under with meaningful 3-set probability
• Over with high 2-set probability
• Over with high 3-set probability

All four outcomes must remain possible.

3. Rebuild the Sets Won display.

Current issue:
The model can show:

Player A:
P(0 sets): 28.7%
P(1 set): 16.4%
P(2 sets): 54.9%

Player B:
P(0 sets): 34.7%
P(1 set): 20.2%
P(2 sets): 45.1%

while simultaneously displaying a 2-set match forecast.

These are marginal probabilities and are confusing because they are not tied to actual match paths.

Replace with conditional match-path probabilities:

Player A 2-0: XX%
Player B 2-0: XX%
Player A 2-1: XX%
Player B 2-1: XX%

Then derive:

• expected sets played
• expected games
• total line probabilities

from those same paths.

4. Audit recent Under increase.

Compare current version against previous calibrated versions.

Measure:

• Under/Over hit rate
• calibration by total line
• 2-set frequency
• 3-set frequency
• average games in 2-set matches
• average games in 3-set matches
• percentage of Unders coming from 3-set matches

If the model is correctly calibrated, 3-set matches should still meaningfully contribute to both Under and Over outcomes.

5. Investigate these possible causes:

• Terminal score calibration shifting too many outcomes downward.
• Hold probability adjustments creating unrealistic short matches.
• Tiebreak probability being suppressed incorrectly.
• PMF reconciliation forcing totals to match the winner model.
• Score root → final PMF shift removing legitimate longer outcomes.
• Three-set branch probability being reduced after winner reconciliation.

Do not tune the engine to produce more Overs or fewer Unders.

The objective is calibration:
When the engine selects Under, it should be because the full simulated score tree supports the edge, including possible 3-set paths.

A correctly functioning tennis total model must allow a close clay match with a high chance of going 3 sets to still land Under when the individual set scores are low enough.

That wording should push it toward fixing the actual architecture instead of just nudging percentages. The key sentence is the one about not inheriting totals from match length. That has been the recurring trap in your engine.