We Went 5-9. Here’s What We Changed.

We went 5-9 on a tough slate. Variance is part of betting, and anyone pretending otherwise is not being honest.

What mattered more than the raw record was where some losses landed: a few were tagged as high-signal spots. That triggered a full review of the tier layer.

This post is the transparent breakdown of what we found, what we changed, and the new angle we are tracking going forward.

What the old layer did well (and where it broke)

The core consensus stack still did its job: model agreement, market context, and published tracking. But the issue was in classification quality.

The old logic leaned heavily on agreement count. In practice, that means a 4/4 agreement could sometimes carry mixed conviction at the model level.

Example profile that used to pass too easily as top-tier:

one model at 51% (low conviction)
one model at 58% (moderate)
one model at 65%+ (strong)
one model simply aligned on side

Technically aligned? Yes. Uniformly high conviction? No.

That gap is what we fixed.

The fix: confidence-aware tier calibration

We reworked tiering so that agreement and conviction are both required for top classification.

What changed:

Tier cutoffs were recalibrated from historical score distribution.
Low-end fragmentation was reduced (lean/split handling cleaned up).
Verified and Strong now absorb picks that are aligned but not truly top-conviction.
Elite now requires stronger score quality, not just agreement optics.

That means fewer inflated Elites and cleaner separation between tiers.

Internal: NHL Consensus
Internal: Testing Models
Internal: Confidence Tiers

Why this matters for real bankroll decisions

Tier labels are not cosmetic. They are position-sizing inputs.

If “Elite” is too permissive, bettors over-allocate to lower-quality pockets. Over a long sample, that costs units.

A better model stack is not one that always predicts more winners overnight. It is one that:

classifies risk more honestly,
sizes exposure more correctly,
and compounds cleaner over large volume.

That is exactly what this recalibration is designed to do.

New angle we surfaced: Platinum profile

Alongside the tier cleanup, we pushed a stronger focus on Platinum Picks.

The Platinum concept is not “more picks.” It is a tighter filter built for quality pockets where multiple conditions align.

In plain language: fewer forced shots, more context-confirmed spots.

What we’re now tracking in public:

Platinum qualifying frequency
Platinum record conversion over rolling windows
combo behavior and context quality by stage

You can monitor this in live pages and testing views:

What users should expect immediately

Fewer borderline plays labeled as top tier.
Cleaner separation between Strong, Verified, and Elite.
Better stability in tier meaning from slate to slate.
More practical visibility into Platinum-style filtered spots.

We are not trying to “win the next night” with a cosmetic tweak. We are optimizing for the next 100+ slates.

Process note: building in public

This is exactly how we run Otterline:

detect weak spots in live performance
audit at the signal-layer level (not excuses)
patch logic
re-measure in public

If another threshold or weight needs adjustment, the data will show it and we’ll update again.

That is the edge: not pretending the model is perfect, but making the model accountable.

Bottom line

5-9 was not the story. The story is what we changed after it.

We tightened Elite classification, improved tier integrity, and elevated a new Platinum angle with stronger historical characteristics.

Record tracking remains public and ongoing.