Immutable Official Learning
ATLAS Learning
Every resolved official pick becomes a permanent learning observation. ATLAS records what happened, which cohort it affected, and whether any wallet or model weight actually changed.
Current Learning State
Official observations are recording.
8 more graded official observations are required before ATLAS may propose recalibration.
Recorded observations do not automatically mean the model retrained. Every future weight or wallet-score change must have its own immutable audit event.
Official Observations
2
Pending Events
60
Minimum Sample
10
Weights Changed
0
Official Record
Full record →Official Graded
58
Wins
30
Losses
28
Accuracy
52%
ROI (capital-weighted)
+2.85%
These are the same numbers as the Official Record, read from the same source. The learning sample below is a subset — every graded pick belongs in the record, but only those with a recorded learning event count toward recalibration.
Since the model changed
from 2026-08-11
The model is level with the market.
ATLAS and the entry price forecast these games about equally well. Matching an efficient market is a genuine result and still not an edge — an edge requires beating it after costs.
Graded since
33 / 25
Record
17-16
Model Brier
0.2499
Market Brier
0.2460
Advantage
-0.0040
Brier score is mean squared error of a probability forecast, so lower is better and 0.25 is a coin flip. The market’s forecast is taken to be the entry price ATLAS paid, over exactly the same games. Advantage is market minus model: positive means ATLAS forecast those games better than the price it bought.
Before the model changed
These calls were really published and are kept exactly as they graded. They are shown apart from the figures above rather than averaged into them: a single number spanning two versions of the model describes neither of them.
Before 2026-08-11
25 graded
Calls graded under the old park factor and the old dispersion. Kept and shown because they were really published, and deleting them would be the concealment this page exists to refuse.
Record
13-12
Model Brier
0.2979
Market Brier
0.2393
Advantage
-0.0587
Nothing here is ever removed. When the model changes again a new segment opens and this one keeps its numbers, which is what makes the record something a reader can audit rather than a figure that improves quietly.
Recalibration Sample
Events Graded
2
Wins
2
Losses
0
Awaiting Event
60
Wallet Learning Integrity
Evidence Snapshots
0
Wallet Change Events
0
Unattached
2
Historical wallet behavior can influence future recommendations only after the wallet, side, category, capital, timing, and pre-pick quality are frozen in the official evidence packet.
Current official events do not yet contain wallet evidence snapshots. ATLAS is recording that absence rather than inventing wallet impact.
Module Learning Integrity
Module Snapshots
0
Weight Changes
0
Observation Only
2
A win or loss updates the evidence sample. It does not silently change EDGE weights, wallet weights, fair-value logic, or unit sizing.
Future recalibration events must record the exact before value, after value, reason, sample, and model version.
EDGE Cohorts
EDGE 80 89
2 observations · 2-0
100%
ROI +208.6% unweighted
Categories
Sports
2 observations · 2-0
100%
ROI +208.6% unweighted
Market Types
MATCH TOTAL
2 observations · 2-0
100%
ROI +208.6% unweighted
Latest Immutable Official Learning Events
OFFICIAL-2026-07-10-2864946-OVER-radar-decision-engine-run-1783684833571-2864946-3
Philadelphia Phillies vs. Detroit Tigers: O/U 9.5
ATLAS OVER · Entry 43¢ · EV 3¢ · 1u
Resolved Jul 10, 2026, 10:40 PM
WIN
ROI +132.6%
OBSERVATION RECORDED
OFFICIAL-2026-07-10-2829399-OVER-radar-decision-engine-run-1783670433870-2829399-1
Spain vs. Belgium: Total Goals O/U 1.5
ATLAS OVER · Entry 26¢ · EV 3¢ · 1u
Resolved Jul 10, 2026, 7:00 PM
WIN
ROI +284.6%
OBSERVATION RECORDED
Legacy Learning Archive
10 legacy evidence-delta events remain preserved.
The original `atlas_learning_events` table is retained for historical audit integrity. It is not mixed into official paid-pick accuracy, ROI, cohort calibration, or model-recalibration readiness.