I agree with one thing you caught: the AI may have...

Question

I agree with one thing you caught: the AI may have “fixed the example” instead of fixing the engine. That’s a common failure mode.

The danger signs are:

The favorite suddenly projects 12 games because this match’s most likely path was 6-3 6-4.
The loser suddenly projects 7 games because this match’s most likely path was 6-3 6-4.
But the PMF still has a mean around 11.3, indicating the underlying distribution is still overweighting long matches.

If that’s what happened, it simply changed the displayed projection instead of repairing the probability model.

I’d tell the AI something like this:

⸻

TENNISLOCKS ENGINE-WIDE TOTAL GAMES, PLAYER GAMES, AND SCORE DISTRIBUTION REBUILD

This is not a request to fix one match, one tournament, one player, or one example output.

The recent changes corrected several presentation issues, but the remaining problems suggest the canonical probability distribution itself may still be structurally biased.

Do not tune the engine until the displayed example “looks right.”

Do not hardcode projections that happen to match this one match.

Do not introduce special handling for favorites, underdogs, two-set matches, three-set matches, specific scorelines, tours, surfaces, or individual players.

The objective is to improve the mathematical model so that every future match naturally produces coherent probabilities without any match-specific logic.

⸻

Verify no hardcoded corrections were introduced

First, perform a full audit to ensure no recent change accidentally replaced a probabilistic calculation with deterministic display logic.

Search for code that:

Forces projections from the most likely path instead of the full PMF.
Chooses one exact score to display while ignoring the remaining probability mass.
Replaces expected values with modal values.
Uses fixed favorite or underdog adjustments.
Applies conditional logic based on winner identity, ranking, Elo, tour, or specific scorelines.
Alters projections only when certain contradictions are detected.

Every displayed statistic must remain a consequence of the underlying probability distribution, not a manually corrected value.

⸻

Reconstruct the entire exact-score probability distribution

Treat the exact-score PMF as the foundation of every downstream market.

For every legal BO3 and BO5 scoreline, compute and expose internally:

Exact score probability.
Total games.
Games won by each player.
Sets won.
Tiebreak count.
First-set games.
Break opportunities.
Hold opportunities.

Verify that every market is generated exclusively from these outcomes.

No downstream market should estimate values independently once the score PMF exists.

⸻

Audit the shape of the PMF rather than only the mean

Do not stop at reporting:

Mean
Median
Mode

Inspect the complete distribution.

Measure:

Skewness.
Kurtosis.
Tail mass.
Density around sportsbook lines.
Probability assigned to every legal total.

Determine whether the distribution consistently allocates excessive probability to long matches.

If the right tail is inflated, identify exactly which scorelines contribute the excess probability.

Repair the transition probabilities creating that inflation rather than suppressing the tail afterward.

⸻

Audit scoreline generation

Generate frequency tables for every legal scoreline.

Examples:

6-0 6-0
6-1 6-2
6-2 6-3
6-3 6-4
7-5 6-3
7-6 6-4
6-4 3-6 6-4
7-5 4-6 6-3

Determine whether the engine systematically:

Produces too many deciding sets.
Produces too many 7-5 sets.
Produces too many 7-6 sets.
Produces too many late recoveries.
Produces too few routine straight-set victories.
Compresses score margins toward competitive outcomes.

Do not compensate by lowering totals afterward.

Repair the score generator itself.

⸻

Audit every transition inside a set

Investigate every transition that extends a set beyond six games.

Specifically examine transitions from:

4-4
5-4
5-5
6-5
6-6

Determine whether:

Hold probabilities become too strong late in sets.
Break probabilities decrease too aggressively.
Clutch adjustments overprotect servers.
Volatility inflates extended games.
Deuce resolution favors holds excessively.

Even small transition errors compound into substantial total-games inflation.

⸻

Audit deciding-set generation

Do not evaluate only the overall percentage of three-set matches.

Instead determine why the engine enters deciding sets.

For every simulated third set, record:

Match state entering Set 2.
Probability of favorite comeback.
Probability of underdog comeback.
Set-specific hold changes.
Break-rate changes.
Fatigue adjustments.
Momentum adjustments.
Volatility adjustments.

Identify whether one subsystem systematically keeps matches alive longer than the underlying player strengths justify.

⸻

Audit winner-margin calibration

Winning probability and winning margin are different quantities.

A player projected to win 62% of matches should not automatically produce highly competitive scorelines.

Determine whether the engine compresses victory margins.

Measure:

Average game differential.
Average set differential.
Favorite game margin.
Underdog game margin.

If favorites consistently win by margins smaller than justified by the underlying hold and break probabilities, total games and loser games will both become inflated.

Correct the margin generation rather than modifying totals afterward.

⸻

Audit player-games generation

Player Games must be computed from the complete exact-score PMF.

Do not derive Games Won from:

Mean total games.
Winner probability.
Regression models.
Independent estimators.
Display heuristics.
Most likely scoreline.

For every sportsbook line, compute the probability directly from the PMF:

P(Player Games > Line) equals the sum of probabilities of all scorelines where the player exceeds that line.

Verify that the displayed mean, median, mode, and Over/Under probabilities all originate from the same PMF.

⸻

Audit total-games generation

Compute Over/Under probabilities only by summing the exact-score PMF.

Do not allow:

Monte Carlo averages.
Markov estimates.
Anchor adjustments.
Lane voting.
Mean total games.
Median games.
Expected games.
Tail weighting.
Calibration layers.

to modify the final Over/Under probability after the PMF is finalized.

Those systems may diagnose or validate the PMF but must not override it.

⸻

Audit structural conservation

For every simulation verify that:

Winner ⇄ Exact Score

Winner ⇄ Sets Won

Winner ⇄ Player Games

Winner ⇄ Total Games

Match Length ⇄ Exact Score

Match Length ⇄ Total Games

Match Length ⇄ Tiebreaks

Exact Score ⇄ Player Games

Exact Score ⇄ Sets Won

Exact Score ⇄ Total Games

Exact Score ⇄ First Set

Total Games ⇄ Player Games

Props ⇄ Canonical PMF

Any disagreement indicates an architectural bug.

Fix the probability generation rather than hiding or overriding outputs.

⸻

Eliminate distribution drift

After every adjustment, verify that no subsystem silently reintroduces bias through normalization, smoothing, blending, calibration, or post-processing.

The final PMF should emerge directly from the simulation.

Do not artificially pull probabilities toward Overs, Unders, favorites, underdogs, close matches, or blowouts.

⸻

Success criteria

The rebuild is complete only when:

No hardcoded match-specific corrections exist.
No player-specific logic exists.
No favorite/underdog shortcuts exist.
Every market derives from the same canonical exact-score PMF.
Total Games, Player Games, Sets Won, First Set, Tiebreaks, and Props remain mathematically consistent.
The PMF produces realistic scoreline frequencies without downstream correction.
Future matches naturally produce coherent outputs across ATP, WTA, Challenger, and ITF events without requiring special-case logic.

The guiding principle is simple: fix the probability engine, not the symptoms. If one generic improvement changes the shape of the PMF correctly, every downstream market should improve automatically.

════════════════════════════════════════
🎾 TENNISLOCKS 🎾
════════════════════════════════════════

🎯 WTA 250 Iasi | Clay, Outdoor | Best of 3 | Total line 20.5
Court speed index: 28 | Tour: WTA

────────────────────────────────────────
Tereza Valentova vs Aliaksandra sasnovich
────────────────────────────────────────
READING GUIDE: Forecast = model outcome | Fair price = model probability price | External EV = comparison with an entered sportsbook price | Pass = no priced edge

========================================
📊 MATCH WINNER
────────────────────────────────────────
Winner forecast: Tereza Valentova (62.4%) | fair ML -166
Moneyline: PASS at entered prices
Model fair ML: Tereza Valentova -166 / Aliaksandra sasnovich +166

════════════════════════════════════════

📊 Player Stats (Live):

Hist. Surface (RED CLAY): Tereza Valentova (53.8% [2-4]) | Aliaksandra sasnovich (53.2% [38-44]) [TA_LIVE_SPW]
Tereza Valentova: Hold 59.4% [Efron-Morris]
Aliaksandra sasnovich: Hold 53.8% (raw: 55.2%, adj: context/fatigue) [Efron-Morris]
Style: Tereza Valentova [Baseline/Tactician] | Aliaksandra sasnovich [Vulnerable]
Match Momentum (Wang): Tereza Valentova Wang Score: 67.5% (Dominant Momentum), W-L 5-2, SS 4-1, Sets 11-5 ; Aliaksandra sasnovich Wang Score: 34.9% (Negative Momentum), W-L 3-4, SS 3-3, Sets 7-8
1st Srv Win: Tereza Valentova 56.4% | Aliaksandra sasnovich 59.6%
2nd Srv Win: Tereza Valentova 50.2% | Aliaksandra sasnovich 43.8%
1st Srv In: Tereza Valentova 62.5% | Aliaksandra sasnovich 64.7%
SPW: Tereza Valentova 54.2% | Aliaksandra sasnovich 53.9% [Barnett-Clarke adj.]
Break Rate: Tereza Valentova 46.2% | Aliaksandra sasnovich 40.6%
Dominance Ratio: Tereza Valentova 0.64 | Aliaksandra sasnovich 0.61
Volatility: Tereza Valentova 3.6% | Aliaksandra sasnovich 12.7% | Match: 8.1%
Elo: Tereza Valentova favored (59.7%) [KM-Elo bridge]
Source: Verified Elo_Lookup (Tereza Valentova=1783, Aliaksandra Sasnovich=1698)
Blend (Tereza Valentova): MC 62.3% | Stats 69.1% | Elo 59.7%
Analytical Hold: Tereza Valentova 63.9% | Aliaksandra sasnovich 63.4% (Newton-Keller) (small sample)

========================================

🎯 TOTAL GAMES:

TOTAL FORECAST: OVER 20.5 (65.1%) | model fair odds -187
Mean 24.0 | median 23 | mode 19 (7.1%) | SD 5.7 | skew 0.20 | typical 80% range 17-31 games
Over probability composition: 16.6% from straight-set paths | 48.5% from three-set paths
Set-shape audit: 7-5 non-tiebreak 12.0% | 7-6 tiebreak 12.4%
Canonical hold inputs: Tereza Valentova 59.5% | Aliaksandra sasnovich 53.7%
########################################
🎯 PLAYER PROPS & PROJECTIONS 🎯
########################################

📊 Tereza Valentova - Projections:
Games Won: 12 projected on most likely 2-0 path (34.1%) | full PMF mean 12.7 | median 13 | mode 12
1st Set Games: OVER 5.5 (64%) | Fair -178
Sets Won: 1.45 projected | median 2 | mode 2 (62.4%) | P(0) 17.3% / P(1) 20.2% / P(2) 62.4%
Serve Games: 11.85 projected
Serve Points: 81.3 projected
Aces: UNDER 0.5 (71%) | Fair -240
Double Faults: UNDER 3.5 (54%) | Fair -118
Breaks Won: UNDER 5.5 (53%) | Fair -112
Break Points Created: OVER 10.5 (53%) | Fair -113
BP Conversion: 49% projected (5.5 / 11.3)
Opponent BP Save: 51% projected (5.8 saved / 11.3 faced)
Opponent return: opp return 28% | gate needs player sample/opponent sample
â ï¸ CAUTION: Props below computed from tour-average return rates due to insufficient player sample. Confidence is reduced.

📊 Aliaksandra sasnovich - Projections:
Games Won: 7 projected on most likely 2-0 path (34.1%) | full PMF mean 11.3 | median 12 | mode 12
1st Set Games: UNDER 5.5 (52%) | Fair -107
Sets Won: 1.04 projected | median 1 | mode 2 (37.6%) | P(0) 34.1% / P(1) 28.4% / P(2) 37.6%
Serve Games: 11.84 projected
Serve Points: 81.7 projected
Aces: OVER 0.5 (56%) | Fair -126
Double Faults: UNDER 4.5 (51%) | Fair -105
Breaks Won: OVER 4.5 (53%) | Fair -113
Break Points Created: OVER 9.5 (54%) | Fair -117
BP Conversion: 46% projected (4.8 / 10.4)
Opponent BP Save: 54% projected (5.6 saved / 10.4 faced)
Opponent return: opp return 29% | gate needs player sample/opponent sample
â ï¸ CAUTION: Props below computed from tour-average return rates due to insufficient player sample. Confidence is reduced.

🎲 Match-Level Context:
1st Set Total: UNDER 9.5 (51.1%)

Tiebreak forecast: NO TIEBREAK (72.2%)
Tiebreak O/U: UNDER 0.5 (72.2%)

Answer 1

Your diagnosis is substantially correct, and the uploaded code makes the failure mode more precise.

What the code is actually doing

1. The 12-game and 7-game “projections” are presentation-layer replacements

The relevant renderer:

js
function _gamesWonForecastPathValue_(engine,side){
  ...
  var value=Number(st.mode);
  if(!Number.isFinite(value))value=Number(st.median);
  if(!Number.isFinite(value))value=Number(st.expected);
  ...
}

It then prints that conditional path value first:

js
path.value.toFixed(0) +
' projected on most likely ' + path.score + ' path'

Only afterward does it report the full-PMF mean. Therefore:

Valentova’s 12 is the mode within the most-likely 2-0 score path.
Sasnovich’s 7 is the corresponding conditional path value.
Neither number replaced or repaired the unconditional distribution.

That is exactly the “fixed the example rather than the engine” behavior you suspected.

2. The 24.0 total and 12.7/11.3 player means are genuinely coming from the canonical PMF

The total-games engine directly consumes root.totalPmf, computes its statistics, and prices the line by summing its discrete mass. Its audit explicitly says:

decisionUsesMean:false
decisionInputs:['CANONICAL_EXACT_TOTAL_GAMES_PMF','SPORTSBOOK_LINE']
gatesApplied:0
rescuesApplied:0
otherEnginesDiagnosticOnly:true

So the Over 20.5 probability of 65.1% is not a displayed-value hack or a post-PMF Over adjustment. It is genuinely what the current root distribution believes.

The conservation validator also checks that:

Player A games + Player B games = total games.
The player-games joint PMF equals the canonical root PMF.
The total-games PMF equals the canonical root total PMF.

That means the downstream markets are now coherent with one another, but they can all be coherently wrong if the root is misshaped. (TennisLocksEngine_v416_SCORELINE_TRANSPARENCY_NO_SYNTHETIC_GW_LINES_20260716 (1).txt)

The more important structural problem

The strongest evidence does not point to a special 5-5, 6-5, or 6-6 transition bug.

It points to the way the BO3 engine couples winner probability to match length.

The BO3 engine:

Starts with the player holds.
Moves A’s hold upward and B’s hold downward until the state-space winner probability equals the externally blended winner target.
Takes the deciding-set probability generated by that same stationary state-space.
Sets:

js
var targetP3 = winnerAlignedP3;

Preserves that P3 through the final reconciliation.

Although it computes a separate recent-history transition estimate, that estimate is diagnostic; it is not used as the live P3 target.

That creates an almost deterministic relationship between match-win probability and deciding-set probability.

For a stationary, approximately independent BO3 set model:

P(\text{match win}) = s^2(3-2s)

A match-win probability of 62.4% implies approximately:

s = 58.34\%

Then:

P(\text{three sets}) = 2s(1-s) = 48.61\%

The engine reports 48.5% from three-set paths.

That near-exact match is the architectural fingerprint. The high total is primarily being produced because a 62.4% winner is automatically being represented as a player who wins each set roughly 58% of the time, which naturally produces a deciding set almost half the time.

In other words:

The engine does not independently model “how often A wins” and “how decisively A wins.”

It derives both from nearly the same stationary strength parameterization.

The late-set audit produces another useful result

Using the displayed canonical holds:

Valentova: 59.5%
Sasnovich: 53.7%

and reproducing the legal alternating-server set transitions gives approximately:

7-5 or 5-7: 12.04%
7-6 or 6-7: 12.39%

The report shows:

7-5: 12.0%
7-6: 12.4%

So those long-set frequencies are not being introduced by a hidden clutch rule in this canonical branch. They are the exact consequences of the supplied hold probabilities and ordinary tennis scoring.

This does not prove that the hold inputs are correctly calibrated. It means that suppressing 7-5 or tiebreak outcomes afterward would be the wrong repair. The real questions are:

Are the canonical holds correctly estimated?
Should the blended winner probability be allowed to overwrite the score-generating strengths?
Is a stationary-strength match model adequate?
Is score-margin variance independently calibrated?

The canonical state-space also keeps recent terminal-score and tiebreak history diagnostic-only rather than applying it to the live distribution.

What I would add to your instruction

Your existing prompt is strong, but this code-specific section would prevent the AI from spending its effort on the wrong subsystem:

:::writing{variant="document" id="58317"} 13. BREAK THE ARTIFICIAL WINNER-PROBABILITY-TO-MATCH-LENGTH COUPLING

The current BO3 architecture appears to adjust the players’ hold probabilities until the legal state-space reproduces the externally blended match-winner probability. It then accepts the deciding-set probability generated by that same stationary state-space.

This can create an artificial near-one-to-one relationship between match-winner probability and deciding-set probability.

Do not assume that a 62% match winner must be represented as a player with approximately a 58% chance of winning every set.

Winning probability and winning decisiveness must be independently identifiable.

Specifically audit and remove any architecture equivalent to:

Solving hold probabilities solely to match the final blended winner probability.
Reusing those solved holds unchanged in every set.
Setting the live deciding-set target equal to the deciding-set probability implied by that winner-aligned state.
Preserving that mechanically generated deciding-set probability through IPF or another reconciliation step.
Computing an alternative deciding-set or score-margin estimate but retaining it only as a diagnostic.

The canonical score model must jointly represent at least:

Match-winner probability.
Straight-set probability.
Deciding-set probability.
Expected game differential.
Expected loser games.
Total-games distribution.
Exact terminal set-score distribution.

These quantities must be calibrated jointly on held-out matches. Match-winner probability alone is not sufficient to determine the score-distribution shape.

Use a generic, symmetric model rather than favorite/underdog rules. Suitable approaches include:

A latent match-performance distribution integrated over the legal tennis state-space.
A legal exact-score PMF reconciled simultaneously to independently calibrated winner and margin targets.
A hierarchical model in which match-level performance variation allows the same winner probability to correspond to different score-margin distributions.
Minimum-divergence reconciliation that preserves legal score conditionals while targeting winner probability, deciding-set probability, and game-margin moments together.

Do not set the final BO3 deciding-set target automatically equal to the structural P3 produced after winner alignment.

Log and expose the following before and after reconciliation:

Original hold inputs.
Winner-aligned hold inputs.
Original structural winner probability.
Final target winner probability.
Original deciding-set probability.
Independently calibrated deciding-set target.
Final deciding-set probability.
Expected total games.
Expected game differential.
Expected games for each player.
Straight-set Over contribution.
Deciding-set Over contribution.

The rebuild must demonstrate that two matches with the same match-winner probability can produce different deciding-set and game-margin distributions when their hold levels, hold gaps, volatility, and calibrated score-shape evidence differ.

RESTORE THE CORRECT MEANING OF “PROJECTION”

Do not label the modal player-games result within the most-likely exact-score path as the player’s overall Games Won projection.

The primary projection must be an unconditional statistic from the complete canonical PMF, such as:

PMF mean.
PMF median.
PMF mode.

A conditional value may be displayed separately as:

“Games on most-likely exact-score path.”

It must not replace or visually outrank the unconditional projection.

For example:

Games Won projection: mean 12.7 | median 13 | mode 12
Most likely exact-score path: 2-0, with 12 games won

Do not describe 12 as the overall projection merely because it is the modal result inside the most likely match-score branch. :::

Bottom line

The current result is not primarily:

“The totals engine added an Over adjustment.”

It is closer to:

“The winner engine supplied 62.4%, the score engine translated that into an approximately 58.3% per-set winner, and the resulting stationary BO3 structure naturally generated about 48.6% deciding sets.”

Then v416 put 12 and 7 in the headline by selecting the modal game counts inside the most-likely 2-0 path, while the unchanged PMF continued to produce 12.7, 11.3, and 24.0 underneath.

So your central conclusion is right: the display was repaired, but the score-distribution architecture was not independently recalibrated. (TennisLocksEngine_v416_SCORELINE_TRANSPARENCY_NO_SYNTHETIC_GW_LINES_20260716 (1).txt)

Answer 2

I audited and edited the uploaded v416 engine. (TennisLocksEngine_v416_SCORELINE_TRANSPARENCY_NO_SYNTHETIC_GW_LINES_20260716 (1).txt)

Main fixes in v417

Removed the BO3 winner-to-score-shape feedback loop. The external winner probability no longer changes the hold inputs used to generate totals, margins, tiebreaks, and exact scores.
Separated winner calibration from match-length calibration. Winner probability is applied only as an exact-score row margin through IPF.
Made deciding-set probability independent. BO3 P3 now comes from the unmodified base-hold state plus the existing shrunk ordered-transition posterior.
Reweighted the base state, not the winner-solved counterfactual state.
Kept winner-aligned holds only as diagnostics. Failure of that diagnostic solver can no longer invalidate the live score distribution.
Added audit fields for raw P3, selected P3, old winner-aligned P3, coupling gap, base-hold conservation, and whether winner alignment affected score shape.
Fixed Games Won presentation. The primary projection is now unconditional PMF mean/median/mode. The most-likely-score value is explicitly labeled conditional.
No match-specific or favorite/underdog adjustment was added.

Validation completed:

Full JavaScript syntax check passed.
Winner and match-length margins reconcile correctly.
Same holds with different winner targets produce the same P3.
Different holds with the same winner target produce different P3.
Final canonical holds remain equal to the original finalized holds.
The previous modal-path projection substitution is absent.

This is an architectural correction, not proof of empirical calibration. Proper ATP/WTA/Challenger/ITF calibration still requires a held-out historical score dataset.

Download the complete v417 script

Download the JavaScript version

Review the v416-to-v417 diff

Answer 3

I completed the post-v417 audit and produced v418.

Most important finding

The engine already contained a capped, row-preserving terminal-score posterior for:

6-0, 6-1, 6-2, 6-3, 6-4, 7-5, 7-6

But it was not affecting the live PMF.

setDist() accepted score-shape evidence, and TennisSetShapeModel.apply() calculated the posterior, but the live generator returned the original structural distribution. The posterior was diagnostic-only, despite other comments claiming it had already been applied upstream.

That was a genuine engine-path omission rather than a display problem.

v418 changes

Applies the terminal-score posterior to the live legal set paths.
Preserves each player’s set-win probability while allowing score-margin and set-length frequencies to change.
Tracks every played terminal set score across BO3 and BO5 matches.
Preserves conditional set-score counts for each exact match result during reconciliation.
Audits four BO3 Total Games stages:
- Live score-shape root before reconciliation
- Deciding-set-margin-only reconciliation
- Winner-margin-only reconciliation
- Final joint reconciliation
Reports mean, variance, skewness, kurtosis, Total Variation, KL divergence, Jensen-Shannon divergence, and sportsbook-line probability movement.
Adds hold-gap compression diagnostics.
Adds explicit stationary-strength and latent-variance status.
Updates the Total Games report to use all played sets, rather than only the first-set PMF, for scoreline-frequency reporting.
Adds full conservation checks between terminal set-score counts and the exact match-score PMF.

Validation

All completed successfully:

JavaScript syntax and full-file evaluation
BO3 and BO5 representative boundary tests
Exact-score, player-games, totals, and all-set PMF normalization
Terminal set-score count conservation
Winner-row preservation
Winner-target/P3 independence
Exact-score reconciliation accounting

With no terminal-score evidence, v418 produces exactly the same score and Total Games PMFs as v417. It therefore does not invent a new cold-start adjustment.

A controlled evidence test changed mean Total Games from 24.00 to 22.95, while match-win probability remained unchanged to numerical precision. That demonstrates the repair can alter score shape without altering winner strength; it is not evidence that this particular reduction is empirically optimal.

Latent within-match performance variation remains disabled because enabling it without held-out calibration would merely replace one assumption with an arbitrary noise coefficient.

Files

Download TennisLocks v418 JavaScript

Download TennisLocks v418 text version

Review the v417-to-v418 diff

Download the validation report

This repairs another real architectural omission, but it still does not prove that the hold priors, posterior concentration, or tour-specific score frequencies are empirically calibrated. That final determination requires held-out historical score data.

Answer 4

I completed the v419 Hold Model & Score Generator Audit.

What changed

Two actual hold-pipeline defects were corrected:

The Elo-service damping was previously a no-op. kmEloBridge() returns f_ij and f_ji, but v418 attempted to blend nonexistent fA and fB fields. v419 now damps the fields that are actually converted into hold probabilities.
Observed service-point exposure was being ignored. Hold shrinkage always substituted service games × 10. v419 now prefers direct service-point attempts, then observed points per service game, and uses the old multiplier only as a fallback.

The larger architectural change is that hold uncertainty now reaches the canonical PMF:

Posterior variance comes from the same Efron–Morris assumptions already used by the hold estimator.
Each player receives a deterministic three-node hold distribution.
Up to nine legal state spaces are combined.
Exact scores, total games, player games, first-set scores, tiebreaks, service exposure, and terminal set-score counts are all mixed from those same nodes.
Low-information players influence the PMF more broadly; high-information players naturally approach the point-hold model.
No random noise, Over/Under correction, or favorite/underdog rule was added.

BO3 now also records the requested ±0.5%, ±1%, and ±2% hold sensitivity tests for both players, including changes in winner probability, deciding sets, mean totals, line-crossing probability, tiebreaks, player games, and modal exact score.

The old winner-aligned hold solver is now disabled by default because it is not part of the live model. It remains available only as an optional research diagnostic.

Validation

Completed successfully:

Full JavaScript syntax and evaluation
BO3 and BO5 PMF normalization
Exact-score and terminal-set count conservation
Player-games and Total Games conservation
Winner-target reconciliation
Winner-target/P3 independence
Twelve hold-sensitivity scenarios
Low-information-to-high-information uncertainty tests
Public formula engine integration
JavaScript and text versions are byte-identical

In a synthetic exposure test, posterior hold uncertainty and its PMF impact decreased monotonically as service-point evidence increased. This validates the intended behavior, but it does not prove that the existing Efron–Morris variance constants or hold blend coefficients are empirically optimal.

Files

Download TennisLocks v419 JavaScript

Download TennisLocks v419 text version

Review the v418-to-v419 diff

Download the v419 validation report

The remaining major task is empirical hold-model calibration on held-out ATP, WTA, Challenger, and ITF data. v419 now exposes the necessary stages and sensitivities so that calibration can target the actual source of error rather than adjusting Total Games afterward.

Answer 5

I followed the evidence-first instruction and did not modify the engine.

Root cause found

The Iasi forecast is reproducible using only:

Hold A: 59.5%
Hold B: 53.7%
Winner target: 62.4%
No terminal-score history
No ordered-transition history
Automatic first-server mixture

That reconstruction produces:

Metric	Result
Mean Total Games	24.0002
Over 20.5	65.14%
Three sets	48.61%
Any tiebreak	27.84%
Player A games	12.7003
Player B games	11.2999

The complete enumeration contains 1,470 legal ordered score paths, sums to exactly 100%, and reproduces the engine’s Total Games PMF to numerical precision.

Where the Over comes from

Straight-set paths contribute 16.60 percentage points to Over 20.5.
Three-set paths contribute 48.54 percentage points.
Only 0.074 percentage points of three-set mass finishes Under 20.5.

The largest straight-set Over contributors are combinations such as:

6-4 7-5
7-5 6-4
6-4 7-6
7-6 6-4
6-3 7-5
6-3 7-6

So the excess is not merely “too many three setters.” The engine also concentrates too much within-set probability around competitive terminal scores.

What elimination testing proved

Reconciliation is effectively innocent:

Mean-games impact: −0.000086
Over impact: −0.0007 percentage points

Other findings:

Winner probability changes have negligible Total Games effects once P3 is fixed.
First-server assignment changes the mean by only about ±0.007 games.
No terminal-score evidence was active in this reconstruction.
No ordered-transition evidence changed P3.
Hold uncertainty did not inflate the forecast; it moved mean and Over slightly downward.
Latent within-match variation was disabled.

Therefore, the high distribution already exists in the stationary hold-driven legal state space before reconciliation.

Historical evidence

I compared the reconstruction with completed WTA clay matches from 2021–2025 using Jeff Sackmann’s public match-result data. The archive identifies Sackmann’s ATP and WTA repositories as its original sources and includes match data through 2026. (GitHub)

A 108-match comparison cohort with similar rolling pre-match clay hold estimates averaged:

Higher hold: 59.88%
Lower hold: 54.63%
Higher-hold player win rate: 60.19%

Metric	Historical cohort	Model
Three sets	37.04%	48.61%
Mean games	21.75	24.00
Over 20.5	50.00%	65.14%
6-0	6.64%	3.12%
6-2	21.88%	16.47%
6-4	18.75%	24.50%
7-6	7.81%	12.39%

The model is shifting mass away from decisive scores and toward close scores.

The match-level differences were statistically meaningful in this cohort:

P3 difference: p = 0.0162
Over 20.5 difference: p = 0.00161
Mean-games difference: p = 7.15e-05

Attribution

Diagnostic counterfactuals—not production recommendations—showed:

Counterfactual	Mean change	Over change
Historical P3 only	−1.123 games	−7.78pp
Historical terminal set shape only	−1.188 games	−5.42pp
Both	−2.250 games	−14.33pp

The remaining inflation appears approximately split between:

Excess deciding-set probability
Within-set margin compression

Published research also supports testing departures from stationary, independent-set assumptions: Depken, Gandar, and Shapiro report strategic and psychological set-level effects using 66,262 professional BO3 matches, while Cui and colleagues document set-to-set performance variation. Neither paper supplies a universal coefficient that should be inserted without held-out validation. (IDEAS/RePEc)

Decision

No v420 was created.

The evidence identifies the likely statistical failure, but it does not yet justify choosing a replacement model or coefficient. Production changes should wait for time-split held-out validation across ATP, WTA, Challenger, and ITF.

Download the full investigation report

Download the complete 1,470-path decomposition

Download the sensitivity analysis

Download the complete audit package

Answer 6

TennisLocks v420 completed

This is an evidence-backed rebuild rather than another display or totals adjustment.

What v420 changes

Adds held-out, symmetric deciding-set calibration by tour and surface.
Adds row-preserving terminal-score calibration for 6-0 through 7-6.
Fixes a discovered gate that prevented population score calibration from operating in cold-start matches.
Ensures surface context reaches every calibrated state-space rebuild.
Adds Challenger Grass P3 calibration after separate training, selection, and 2025 testing.
Updates only ace/DF fallback cells that reduced locked 2025 error.
Leaves rejected or unsupported cells unchanged.

The historical files came from the Jeff Sackmann ATP and WTA repositories and their archive mirror. (GitHub)

Independent 2025 validation

Across the accepted cells:

Metric	Baseline	v420
P3 log loss	0.6819	0.6484
Terminal-score log loss	1.8221	1.8071
Total Games MAE	5.4472	4.9492
Over 20.5 Brier	0.2708	0.2504
Exact-total log score	3.1146	3.0385
Total-distribution CRPS	3.5140	3.2600

Iasi reconstruction

Metric	v419	v420
Valentova winner probability	62.40%	62.40%
Mean Total Games	24.000	21.809
Over 20.5	65.14%	50.68%
Three sets	48.61%	36.68%
Any tiebreak	27.84%	21.47%
Valentova games	12.700	11.669
Sasnovich games	11.300	10.140

The winner probability is unchanged. The movement comes from historically calibrated match-length and terminal-score distributions.

The reference PMF still contains 1,470 legal ordered paths, sums to one, and reproduces the canonical Total Games PMF with total variation below 2e-16.

Validation status

JavaScript syntax and full evaluation: passed
Full canonical validator: 5/5 cases passed
36-case BO3 stress grid: passed
Maximum PMF mass error: 6.67e-16
Maximum player-games/total-games identity error: 3.20e-14
Unsupported ATP Grass versus v419: unchanged to approximately 1e-14
BO5 versus v419: exactly unchanged in the controlled comparison
JS and text versions: byte-identical

ITF score calibration and BO5 recalibration remain intentionally unchanged because the available data did not support defensible hold-conditioned replacements. The report documents these limits rather than presenting them as solved. The original engine lineage audited for this work is available in the uploaded source. (TennisLocksEngine_v416_SCORELINE_TRANSPARENCY_NO_SYNTHETIC_GW_LINES_20260716 (1).txt)

Downloads

Download the v420 JavaScript engine

Download the v420 text version

Read the complete engineering audit

Download the validation report

Review the v419-to-v420 diff

Download the complete 37-file audit package

Download the 1,470-path decomposition

Answer 7

v421 red-team audit completed

v420 did not survive unchanged. Its population calibrations generalized, but several live layers failed independent testing and were replaced or removed.

What failed

Player-specific deciding-set history overlay

On the independent 2026 replay of 6,935 matches:

P3 model	Log loss	ECE
Structural hold model	0.681520	0.125210
Population calibration	0.649562	0.008016
Population + player history	0.655194	0.030858

The history overlay matched the aggregate P3 rate more closely but destroyed calibration discrimination: slope fell from approximately 1.011 to 0.223. This was a compensating error, not a genuine improvement.

Player-specific terminal-score posterior

Across 16,327 completed sets in the 2026 replay:

Terminal-score model	Log loss
Structural	1.826410
Population calibration	1.808545
Population + player history	1.843131

The player-history layer worsened results in every evaluated year and every evaluated tour/surface cell. Both recent-history systems remain available as diagnostics but can no longer alter the canonical PMF.

Hold-uncertainty mathematics

v419/v420 calculated posterior variance using service-point-rate parameters, then added the resulting standard deviation directly to game-hold probability. Those are different probability scales.

v421 now:

Creates uncertainty nodes in service-point space.
Converts each node through the exact tennis-game hold equation.
Builds a complete legal score state at every node.
Mixes the resulting exact-score, games, tiebreak, and service-exposure distributions.

The improvement is small but consistent:

Model	Total MAE	Exact-total log score	CRPS
Point holds	5.012015	3.009826	3.275398
Old hold-space uncertainty	5.011652	3.009723	3.275280
Correct SPW-space uncertainty	5.010234	3.009430	3.274855

What survived

The held-out population P3 calibration improves every evaluated year:

Year	Structural	Retained calibration
2024	0.682223	0.649436
2025	0.682410	0.649672
2026	0.681520	0.649562

The population terminal-score calibration also improves 2024, 2025, and 2026.

A competing global logistic model produced a tiny aggregate 2026 improvement of approximately 0.00017 in P3 log loss, but failed to improve consistently across tours, surfaces, and hold-gap segments. It was therefore rejected rather than substituted on the strength of a negligible overall gain.

Additional corrections

Three v420 service fallback changes failed the partial-2026 replay and were reverted:

ATP Hard double faults: 0.03711 → 0.030
WTA Clay aces: 0.02829 → 0.034
WTA Clay double faults: 0.05044 → 0.049

The historical replay used chronologically ordered public ATP and WTA match-result data compiled by Jeff Sackmann and its archival mirror. (GitHub)

Adversarial validation

An extreme recent-history input moved v420’s deciding-set probability by 23.69 percentage points and changed its Total Games PMF by 0.241 TV distance. In v421, the same input changes both by exactly zero because the failed posterior is diagnostic-only.

Validation results:

JavaScript syntax: passed
Full canonical validator: 5/5
Calibrated market grid: 54/54
Adversarial grid: 96/96
Winner targets tested: 5%, 50%, and 95%
Hold range tested: 0.30–0.99
Maximum PMF mass error: 4.44e-16
Maximum player-games/total-games identity error: 4.26e-14
No duplicate top-level functions or modules
No Total Games, Player Games, props, pricing, or confidence feedback into the canonical BO3 score generator

The original engine audited in this work is the uploaded TennisLocks source. (TennisLocksEngine_v416_SCORELINE_TRANSPARENCY_NO_SYNTHETIC_GW_LINES_20260716 (1).txt)

Downloads

Download the v421 JavaScript engine

Download the v421 text version

Read the complete red-team audit

Download the validation report

Review the v420-to-v421 diff

Read the dependency map

Download the complete reproducible audit package

v421 is better supported than v420, but it is not presented as proof that every possible defect has been eliminated. ITF hold-conditioned calibration, indoor cells, and new BO5 calibration remain unresolved where the available data did not justify a replacement.

Câu hỏi

Python

Suy nghĩ

Trả lời

What the code is actually doing

1. The 12-game and 7-game “projections” are presentation-layer replacements

2. The 24.0 total and 12.7/11.3 player means are genuinely coming from the canonical PMF

The more important structural problem

The late-set audit produces another useful result

What I would add to your instruction

Bottom line

Câu hỏi

Suy nghĩ

Trả lời

Main fixes in v417

Câu hỏi

Suy nghĩ

Trả lời

Most important finding

v418 changes

Validation

Files

Câu hỏi

Suy nghĩ

Trả lời

What changed

Validation

Files

Câu hỏi

Suy nghĩ

Trả lời

Root cause found

Where the Over comes from

What elimination testing proved

Historical evidence

Attribution

Decision

Câu hỏi

Suy nghĩ

Suy nghĩ

Suy nghĩ

Trả lời

TennisLocks v420 completed

What v420 changes

Independent 2025 validation

Iasi reconstruction

Validation status

Downloads

Câu hỏi

Suy nghĩ

Suy nghĩ

Suy nghĩ

Suy nghĩ

Trả lời

v421 red-team audit completed

What failed

What survived

Additional corrections

Adversarial validation

Downloads

Chia sẻ Q&A này