⚠️ CRITICAL DIRECTIVE: ILLUSTRATIVE PROBABILITIES ...

⚠️ CRITICAL DIRECTIVE: ILLUSTRATIVE PROBABILITIES ONLY
Context for the AI Agent:
All specific numbers, percentages, probabilities, and scores mentioned in this architectural brief (e.g., 0.50 default traps, 64% surface averages, 85% SPW posterior spikes, 6-0 blowout scores, or 80% hold rates) are purely illustrative examples used to explain the mathematical anomalies and failure modes.
Strict Execution Rules:
DO NOT hardcode any of these example values into the engine's bounding logic, Bayesian priors, or fallback functions.
DO NOT write code that sets a hard ceiling of 85% or a hard floor of 64%.
DO NOT use 0.64 (or any other number) as a hardcoded default for missing data. You must follow the explicit rule to intercept ⁠NO_CURRENT_POINT_PAIR⁠ and strictly PASS/return an ⁠UNPRICED⁠ status.
All ceilings, floors, shrinkage targets, and constraints must be derived dynamically from the live data pipeline, the surface form calculations, and the mathematical axioms provided.

SYSTEM PROMPT FOR AI AGENT: BO3 Tennis Probability Engine Audit & Debugging
Role & Mission
You are an expert Quantitative Sports Modeler and Python Engineer. Your mission is to audit, debug, and strengthen a live Best-of-3 (BO3) tennis pricing engine.
The user recently refactored the BO3 architecture to base match outcomes entirely on set-coverage probabilities (winning at least one set). While this successfully unified the exact-score PMF and Sets Won markets, several mathematical anomalies have emerged (runaway variables, negative probabilities, and overconfidence).
Your goal is to fix these anomalies without abandoning the new coverage-based architecture.

1. How the BO3 Architecture Was Changed (The New Working Baseline)
   The engine no longer uses legacy methods (e.g., corpus P3 targets, Set-2 Jeffreys shrinkers, set-winner-mass tilts, or post-PMF patches). Instead, BO3 length and exact scores are derived directly from a player’s probability to avoid a sweep (win 1+ set).
   The Root Geometry:
   ⁠coverA⁠ = \bm{P(\text{Player A wins} \ge 1 \text{ set})}
   ⁠coverB⁠ = \bm{P(\text{Player B wins} \ge 1 \text{ set})}
   ⁠P3⁠ (Over 2.5 Sets) = \bm{\text{coverA} + \text{coverB} - 1}
   ⁠P2⁠ (Under 2.5 Sets) = \bm{1 - \text{P3}}
   The Exact Score Reconciliation:
   The four exact-score probabilities are solved to preserve the independent Match Winner marginals exactly:

How Coverage is Calculated:
Instead of a separate coefficient, the "win 1+ set" probability is rebuilt from the point-state mixture itself:

1. Simulate Set 1.

2. Condition the latent point-strength states (SPW/RPW) on the hypothetical Set-1 score.

3. Re-run Set 2 from that posterior state.

4. Sum the two sweep probabilities to find engine coverage: \bm{P3 = 1 - P(2\text{-}0_{sim}) - P(0\text{-}2_{sim})}.

5. This is blended with target-surface form, combining observed sweep avoidance (\bm{1 - \text{straightLosses}/\text{matches}}) and recent set-win rate converted to BO3 (\bm{1 - (1 - \text{setWinRate})^2}), using a \bm{\sqrt{\text{set count}}} weight discount.

6. The Anomalies to Investigate and Fix
   Despite the clean architecture, the integration of point-state conditioning and surface-form blending is causing cascading mathematical failures. You must investigate and fix the following four anomalies:
   Anomaly A: Posterior SPW Spikes (The Root Cause)
   The Issue: The Serve Points Won (SPW) metric occasionally becomes absurdly high during the simulation phase.
   Where to Look: Look at the conditioning step between Set 1 and Set 2. If a player dominates Set 1 (e.g., 6-0 or 6-1), the Bayesian update/posterior for Set 2 is likely overreacting, pushing the latent SPW parameter to extreme boundaries (e.g., > 85%).
   The Fix: Implement tighter bounds or heavier prior regularization on the Set 2 point-state posterior. A strong Set 1 should improve Set 2 parameters, but not cause variance explosions.
   Anomaly B: P2 Over-prediction (Under 2.5 Bias)
   The Issue: The model is calling for 2-set matches far too often, generating P2 probabilities that are too high.
   Where to Look: This is a downstream effect of Anomaly A. When posterior SPW spikes, the model thinks the Set 1 winner will easily sweep Set 2, inflating \bm{P(2\text{-}0)} and \bm{P(0\text{-}2)}, which artificially drives up P2. Additionally, check the surface form calculation: the IID assumption \bm{1 - (1 - q)^2} may be systematically underestimating set-split correlations.
   The Fix: Constrain the sweep probabilities from the simulation and introduce a correlation discount to the surface form math so that P3 isn't artificially suppressed.
   Anomaly C: Match Winner Overconfidence
   The Issue: Match Winner probabilities have become highly polarized/overconfident since this update.
   Where to Look: Look at the exact score formula: \bm{P(2\text{-}1) = P(\text{A wins}) - P(2\text{-}0)}. If the engine is producing an artificially inflated \bm{P(2\text{-}0)} (due to Anomaly A), the only way the solver can prevent \bm{P(2\text{-}1)} from going negative is by artificially dragging the Match Winner \bm{P(\text{A wins})} upward.
   The Fix: Decouple Match Winner from being forced upward by the sweep probability. Ensure Match Winner serves as an absolute anchor, and \bm{P(2\text{-}0)} is capped mathematically so it never exceeds \bm{P(\text{A wins})}.
   Anomaly D: Zero-Set Truncation (0 Coverage Edge Cases)
   The Issue: In some matchups, the model outputs a scenario where a player is predicted to not win a set at all (Coverage drops to 0, or exact scores break).
   Where to Look: Look at the post-blend step where form and simulation mix. Small sample sizes (\bm{N=7} matches) combined with a bad point-state simulation can drag ⁠cover⁠ below the player's Match Winner probability.
   The Fix: Implement strict Fréchet bounds. A player's coverage (chance to win \bm{\ge 1} set) must be strictly greater than or equal to their chance to win the match. Add a post-blend clamp: ⁠coverA = max(coverA, match_win_prob_A)⁠.

7. Strict Guardrails for Missing Data (⁠NO_CURRENT_POINT_PAIR⁠)
   The Issue: The pipeline occasionally crashes or generates fabricated probabilities when a player lacks measured SPW/RPW data.
   The Fix: You must harden the missing data guardrails. If the system detects ⁠[NO_CURRENT_POINT_PAIR]⁠ for at least one player:

8. Do NOT fall back to \bm{0.50} default point probabilities.

9. Do NOT fabricate coverage stats.

10. The function must immediately log ⁠[DATA] MATCH PREVIEW UNPRICED⁠ and cleanly return/pass without attempting to solve the BO3 matrix.

11. Execution Steps for the AI Agent

12. Audit the State Transition: Review the code handling the transition from Set 1 -> Set 2. Apply a variance clamp to prevent SPW from spiking to extreme levels based on a single hypothetical 6-0 set.

13. Apply Axiomatic Clamps: Inject bounding logic into the exact-score solver. Ensure \bm{P(2\text{-}0) \le P(\text{A wins})} and \bm{\text{coverA} \ge P(\text{A wins})}.

14. Fix Form Weighting: Review the \bm{1 - (1 - \text{setWinRate})^2} logic. Ensure small sample sizes (\bm{N < 10}) don't completely override the simulation's baseline coverage.

15. Posterior SPW Spikes ("The Blowout Bias")
    What went wrong: Your model is overreacting to small sample sizes when conditioning Set 2 probabilities on Set 1 outcomes.
    The Math: A 6-0 or 6-1 set is exceptionally short—often only 20 to 30 total serve points. If Player A wins a set 6-0, they may have won 12 out of 14 points on their serve (85% SPW). If your Bayesian updater feeds that 85% directly into Set 2 without enough resistance, the Markov chain breaks. In tennis modeling, shifting a player's SPW from an average of 64% up to 75%+ changes their probability of holding serve from ~80% to over 95%.
    The Fix: You need Bayesian shrinkage. Your prior (their long-term SPW on that surface) must be heavily weighted so that 15 points of dominant Set 1 serving only nudges their Set 2 SPW up by a fraction of a percent, rather than dragging the posterior all the way to the in-match average.

16. P2 Over-prediction (The Under 2.5 Sets / Sweep Bias)
    What went wrong: Your model assumes that sets are Independent and Identically Distributed (i.i.d.), which causes it to price 2-0 sweeps too frequently.
    The Math: If your model calculates that Player A has a 60% chance to win any given set, basic i.i.d. math dictates a 2-0 sweep happens 36% of the time (\bm{0.60 \times 0.60}). However, quantitative research by Klaassen and Magnus (2001) famously proved that tennis points and sets are highly non-stationary. Players who win Set 1 often experience a subconscious dip in intensity, while the loser plays with elimination urgency.
    The Fix: You must introduce a negative autocorrelation or "momentum penalty" between sets. If Player A wins Set 1, their baseline probability to win Set 2 should be slightly discounted to reflect real-world mean reversion, which will properly inflate your Over 2.5 Sets / 2-1 exact score pricing. 

17. Match Winner Polarization via Coverage Constraints
    What went wrong: You are forcing your Match Winner odds to be the sum of your exact score odds, allowing a bloated Set betting market to wag the dog.
    The Math: You are likely enforcing the linear constraint:

If your model suffers from Anomaly #2 (inflated 2-0 predictions), and you build your Match Winner probability strictly from the bottom up, that inflated \bm{P(2-0)} will artificially drag your overall Match Winner odds into extreme, unbettable polarization (e.g., pricing a -150 favorite as a -300 favorite).
The Fix: Top-down reconciliation. Model the overall Match Winner \bm{P(A_{match})} independently using Elo, surface yield, and H2H. Then, use your point-state model to distribute that top-line probability into the exact score buckets, rather than letting the exact score buckets dictate the top-line price.
4. Zero-Set Truncation & Negative Probabilities
What went wrong: Your blended sub-models are operating independently and violating the axioms of probability.
The Math: If you use one algorithm to predict Match Winner \bm{P(A)} and a separate, unconstrained algorithm to predict a Set 1 win or a 2-0 sweep \bm{P(A_{2-0})}, variance will inevitably cause the subset to outprice the superset. If the model outputs \bm{P(A) = 0.55} but \bm{P(A_{2-0}) = 0.60}, calculating the 2-1 score yields:

The Fix: Mathematical bounding. You must enforce the rule that \bm{P(A_{2-0}) \le P(A)} at the code level. If the raw outputs violate this, you need a normalization function (like a Softmax applied to the constrained bounds) to force the exact scores to fit inside the Match Winner envelope.
5. The 0.50 SPW Imputation Trap
What went wrong: You imputed missing Serve Points Won data with a raw 0.50 (50%) mathematical default, which destroys the variance structure of the sport.
The Math: In the ATP, average SPW is ~64%. At 64%, a server holds their game about 80% of the time. If you input 0.50 for a player with missing data, you are telling the Markov chain that the server has zero advantage. The model will simulate an endless string of breaks and deuces, entirely flattening the match variance and producing useless pricing.
The Fix: Apply your established data rules: If data conflicts or cannot be verified, do not guess—clearly state uncertainty or PASS. If you absolutely must simulate the match, impute missing data using the tour-level and surface-level average (e.g., 0.64 for ATP Hardcourt), never a blind 0.50.

Correct my script delete the corrupted parts and require properly do not make mistakes after all the corrections check for syntc