I continued through the actual v1163 root. We foun...
Erstellt am: 15. September 2026
Beantwortet mit GPT-5.6 Thinking von Chat01
Frage
TennisLocks_FINAL_v1164_BO3_NEUTRAL_BAND_OUTPUT_CLEANUP_20260914.txt
I continued through the actual v1163 root. We found something important, and I’m not going to touch Sets Played again.
The next issue is the relationship between:
point-strength → game hold → set winner → P(3 sets)
For the Janvier example, the math is internally coherent:
- SPW: 60.5% vs 58.2%
- Point/game/set tree gives roughly 48.8% P3
- Match Winner is independently 61.7%
- Final PMF preserves that Winner probability while preserving the set-count margin.
- Sets Played then simply settles that final PMF.
So the UNDER 2.5 is genuinely coming from the length root, not the Sets Played publisher.
But I also found the important architectural distinction: the script contains a predictive 3×3 point-state quadrature that can produce substantially lower P3 than the central point/game/set tree. That route is used for the Winner/profile forecasting machinery, not as the normal final Sets Played root. So removing its variance would be the wrong repair.
What I’m targeting next
I’m going after the actual BO3 set-length geometry, specifically whether the current stationary service-point/game construction is systematically producing too little set stealing.
That means checking:
- SPW → hold conversion
- alternating server state
- set-win probability
- P3 = both players win a set
- Winner/IPF reconciliation
- Whether the current domain geometry tables are actually being applied or are just stale comments
And that last one is interesting because v1163 still says the BO3 geometry authority was recalibrated for ATP Hard, CH Hard/Clay and WTA Hard/Clay, but the live setDist() path itself is basically the structural Markov tree. I don’t want a supposedly “calibrated” table sitting in the script while the actual live tree ignores it.
So I’m moving into v1164: BO3 LENGTH-ROOT GEOMETRY AUDIT/REPAIR, with one rule:
No arbitrary P3 boost.
If the chronology supports changing the geometry, I’ll change the actual owner. If it doesn’t, I’ll leave the root alone and show exactly why.
The recent Sets Played fixes remain intact. We are digging underneath them now. 🔬
There’s probably more problems I haven’t caught that u need to identify and correct and removed old code and wire code correctly
Gedanke
Gedanke
Gedanke
Gedanke
Gedanke
Antwort
I found and repaired the deeper issue. The biggest finding is that in this uploaded v1164, the separation you intended was not actually enforced: the exact-profile path assigned the 3×3 posterior-predictive state as _mssPreviewRoot, and that root’s score PMF then flowed into the official exact-score/length construction. (TennisLocks_FINAL_v1164_BO3_NEUTRAL_BAND_OUTPUT_CLEANUP_20260914.txt) (TennisLocks_FINAL_v1164_BO3_NEUTRAL_BAND_OUTPUT_CLEANUP_20260914.txt)
That matters because the 3×3 machinery explicitly averages complete score PMFs across the nine point states, not just Winner uncertainty. (TennisLocks_FINAL_v1164_BO3_NEUTRAL_BAND_OUTPUT_CLEANUP_20260914.txt) And the point-root configuration itself says those predictive-state parameters were not fitted to set count outcomes. (TennisLocks_FINAL_v1164_BO3_NEUTRAL_BAND_OUTPUT_CLEANUP_20260914.txt)
For the Janvier regression, I reproduced:
- Central SPW 60.5% / 58.2% → holds 74.606% / 69.700% → P3 = 48.7526%, P(A wins) = 61.7476%.
- The same central logits passed through the existing BO3 FUSED 3×3 predictive state produce P3 = 36.3605%.
- That is a 12.39 percentage-point shortening before Sets Played ever sees the PMF. So yes, it was fully capable of manufacturing the strong Under behavior you were tracing.
- I did not remove the predictive variance. I split ownership: predictive/profile state remains for Winner and conditional game/service machinery; BO3 set count now gets a dedicated central matchup SPW → hold → alternating-server set tree.
- Winner remains independently finalized upstream, and the existing IPF then preserves that Winner margin while preserving the selected P2/P3 margin. That reconciliation logic was sound and did not need a P3 correction. (TennisLocks_FINAL_v1164_BO3_NEUTRAL_BAND_OUTPUT_CLEANUP_20260914.txt)
- I also confirmed the supposed historical BO3 “domain geometry authority” was effectively a ghost. The live standard
setDistCal()simply delegates to structuralsetDist(), whose transitions are stationary hold/break probabilities with alternating server state. (TennisLocks_FINAL_v1164_BO3_NEUTRAL_BAND_OUTPUT_CLEANUP_20260914.txt) So I removed the stale authority claim instead of inventing a calibration. - I removed dead
scorePmfIid/source bookkeeping, unused V995 posterior flags, the orphan BO3 tour/surface helper, obsolete pricing-root source plumbing, and dead preview recovery locals. - I fixed another bug you hadn’t mentioned:
_tlLengthEvidence_could declare SETS canonical authority before validating the hold pair. In the original ordering, missing holds were checked only after the SETS early return. (TennisLocks_FINAL_v1164_BO3_NEUTRAL_BAND_OUTPUT_CLEANUP_20260914.txt) It now fails closed. - I added two hard anti-regression protections: the final BO3 length root rejects a posterior-predictive state even if its source gets renamed, and a stamped central root rejects any separately supplied score PMF that does not exactly match its own central score seed.
- I added
_tlBo3LengthRootIsolationTripwireV1165_()with the Janvier 60.5/58.2 geometry. It must recover roughly 48.5–49.0% P3 while keeping the deliberately low predictive P3 diagnostic isolated. - Sets Played itself was left alone. Its final-PMF settlement and v1164 55%/60% neutral-band publication rules remain intact.
I bumped the substantive repair to v1165 — BO3 LENGTH-ROOT GEOMETRY AUDIT/REPAIR. Full-source node --check passes, the isolated Janvier geometry regression passes, and the removed-symbol/static route scans pass. I could not execute the bound Google Sheets/Apps Script environment here, so the included tripwire is also there for an in-runtime check after deployment.