Remember keep the engines separate I checked the...

Remember keep the engines separate

I checked the current v17.7.0 thresholds first.

What the totals side is using now
• Base strong-pick edge (minEdge)
• ATP: 2.2%
• WTA: 2.5%
• CH: 2.8%
• ITF: 3.0%
• Then minEdge gets adjusted by shape:
• more central mass = slightly easier to fire
• bigger two-tail / disagreement = slightly harder to fire
• Then it is clamped:
• ATP: 1.5% to 6.0%
• WTA: 1.7% to 5.5%
• CH: 1.6% to 5.8%
• ITF: 1.6% to 6.0%
• Lean band (leanEdge)
• ATP: minEdge * 0.55
• WTA: minEdge * 0.50
• CH: minEdge * 0.52
• ITF: minEdge * 0.53
• with floor around 0.8%–0.9%

So the pick ladder is:
• OVER / UNDER if edge clears minEdge
• LEAN OVER / LEAN UNDER if edge misses minEdge but clears leanEdge

Confidence thresholds now

Base confidence from selected-side probability:
• ATP: HIGH at 62%, MID at 56%
• WTA: HIGH at 60%, MID at 54%
• CH: HIGH at 59%, MID at 53%
• ITF: HIGH at 58%, MID at 52%

Then structural overrides can raise confidence:
• UNDER gets boosted from strong 2-set structure
• OVER gets boosted from strong 3-set / right-tail structure

⸻

The real issues I see

1. There is still an over-side bias problem in the confidence helper

The biggest one:

Your internal “cluster confirms direction” helper is not symmetric.

For UNDER, it checks whether the 2-set median is below the line, which makes sense.

For OVER, it also checks the 2-set median, when it should be using the 3-set structure / right-tail support.

That means:
• UNDERs get their penalty shield correctly
• OVERs almost never get the same protection

So even if your pick logic is cleaner now, the confidence logic can still punish overs more often than unders.

That is a real issue.

⸻

2. ITF confidence override thresholds are not truly separate

Your base confidence thresholds do include ITF.

But the structural override gates for confidence do not explicitly branch ITF in the same way.
In that part, ITF is effectively falling into the default branch too often.

That means ITF is not as tour-specific there as it should be.

⸻

3. Your dead / downgrade thresholds are still global defaults

These are still effectively shared defaults:
• deadOverPO = 0.47
• deadUnderPU = 0.47
• minLeanOverPO = 0.545
• minLeanUnderPU = 0.545

So even though comments mention 32k calibration, those particular cutoffs are still basically flat defaults, not truly tour-shaped thresholds.

That is another place where the script is less calibrated than it looks.

⸻

4. A structural side can still force a LEAN when pricing is flat

This is why you sometimes get things like:
• model fair right on the line
• raw MC close
• but still a LEAN

That happens because the policy still allows:
• MC tentative side
• then structural support
• then cluster/Markov hold the direction

So the script can still output a structural LEAN even when the pricing edge is tiny.

That is not automatically wrong, but it is a policy choice.
If you want flat pricing = PASS more often, that gate still needs tightening.

⸻

Bottom line

Good now
• Base under/over thresholds are not crazy
• Confidence thresholds are sensible
• The totals engine is much cleaner than before

Still needs work
1. Fix over-side confidence confirmation so it uses over-appropriate structure
2. Make ITF structural confidence gates explicit
3. Replace the flat 0.47 / 0.545 defaults with true tour-aware thresholds
4. Decide whether zero-edge structural leans should stay LEAN or downgrade to PASS

The biggest hidden problem is still #1.
That is the one most likely to keep making the system feel slanted.

Should be using the 32k calibration/kernal thresholds

Is try to fix my script update and send it as a downloadable txt link
Also trying to fix the over bias it doesn’t like picking overs when it’ll truly go over it just mainly passes for some odd reason