Plus 3.90 percent ROI on spreads. Minus 16.59 percent ROI on totals. That is the framework at the All-Star break, in the two numbers that matter most.
The break is the natural moment to audit the season. What worked, what broke, what needs to change. This is my mid-season framework report.
The scoreboard.
Per the grader, cumulative record across the season for game-card tiers (which is the aggregator that best represents the full body of work):
Spread STRONG LEAN: 49-41-1 (54.4 percent, plus 3.90 percent ROI). Spread MODEST LEAN: 22-24 (47.8 percent, minus 8.70 percent ROI). Total STRONG LEAN: 45-58 (43.7 percent, minus 16.59 percent ROI). Total MODEST LEAN: 27-28 (49.1 percent, minus 6.28 percent ROI).
Two things read cleanly from those numbers. Spread STRONG is a validated positive-expectation tier. Total STRONG is a signal that something in the totals model is directionally biased against the current WNBA pace environment.
What held.
The Issue 42 half-stake default. This is the single rule change that most changed the season's variance profile. Before Issue 42, we were sizing bot HIGH picks at full stake and getting whipsawed by the variance of a small-sample tier. After Issue 42, we size at half stake by default and only step up to full stake when the conviction architecture fires (all three conditions: road team favored, agent late-quarter alignment, structural home Q4 weakness). Full-stake nights are now rare — I count maybe four in the past eight weeks — and half-stake nights are the norm. The bankroll variance has been meaningfully smoother.
The model-versus-agent conflict rule. This has fired maybe eight times across the season and has been right at least six of them. The clearest recent example was Tuesday, when the picks file had Connecticut plus 11 at Indiana as a HIGH pick, the agent had three HIGH-tier quarter edges on the Indiana side plus a "IND dominant game" SGP call, and the framework PASSed. Indiana won by 35. That was 0.5u saved. The historical pattern is that when the 2026 picks file and the 2025 agent structurally disagree on a HIGH pick, the agent's read (based on last-year net ratings and quarter dominance patterns) tends to be right.
The conviction architecture rule. Rarer trigger than the model-versus-agent filter but higher-conviction when it does fire. The three-condition setup (road favored, agent late-quarter alignment, structural home Q4 weakness against DAL/IND/LAS/CHI in particular) has produced a 5-of-6 hit rate over the season on full-stake sizing. Not a lot of trigger events but the ones that fire have been clean.
The bot's auto-demote layer. In late June, the model repo shipped a STRUGGLING_TEAMS_DEMOTE list (Portland only, as of the deployment) that auto-demotes HIGH picks on those teams to MEDIUM at the source. This replaced a manual filter we had been running on POR-at-home HIGH picks after two consecutive 20-plus-point losses on that pattern. The bot's version is cleaner and catches the pattern before it ever reaches our editorial layer.
The Issue 49 quarter-stake recalibration. This was the mid-June rule that said: on non-HIGH STRONG LEAN cards with edge above 5.0 (per the arcwnba game cards, not the picks file), size a quarter stake. The first two weeks after implementation ran essentially flat (net negative 0.04u). Not a home run, not a disaster; the sample is still small.
What broke.
Totals STRONG. Full stop. This is the biggest live problem in the framework and it has been getting worse, not better, across July.
The tier has been below 45 percent win rate for approximately the last six weeks. The ROI has drifted from around minus 8 percent in early July to minus 16.59 percent as of this morning. The pattern that keeps hitting the tier is games projected to go UNDER that end up going OVER by 3 to 15 points. The bot's totals model appears to have a pace-projection bias that is missing something structural about this WNBA season — likely the shortened rotations from injury attrition combined with faster-tempo half-court sets that most teams have adopted post-Vegas-2022.
The half-point beats on totals HIGH. This week alone we lost the CHI-NYL UNDER 188.7 HIGH by exactly half a point (final 189). That is variance not signal, but it is compounding on top of the structural totals bias. The framework has been correctly on the right side of the number on totals HIGH picks fairly often, but the shortened WNBA scoring environment produces games that land right at the line all the time. Half a point up-or-down decides more totals per season than any other margin.
The two-half-point-beats-per-week rhythm. This is not a rule failure but a season-shape observation. HIGH-tier picks have been getting bad-beated in a way that a 55-percent-implied-hit-rate should not produce as often as it has. Six of the last ten sized HIGHs have lost. That is a rough run but not statistically alarming yet at the ten-pick sample size.
What needs to change coming out of the break.
Totals STRONG structural downgrade. If the ROI drops another two percentage points from where it is now (to approximately minus 18.5 percent), the recommendation is:
Totals HIGH: quarter stake instead of half stake. Totals STRONG: PASS entirely. Totals MODEST: PASS entirely (already default).
That would leave us essentially not betting totals at all except on the rare very-large-edge HIGH pick at a small stake. That is the right posture for a market tier that has structurally underperformed for two-plus months.
The threshold has not been triggered yet. If the ROI stays around minus 16 to minus 17 percent for the rest of July, I keep the tier where it is (totals HIGH at half stake, totals STRONG PASS by default). If it drops materially further, the downgrade fires automatically.
Agent-alignment weighting. The rule that says "PASS when the agent has three or more HIGH quarter edges on the opposite side" has been right often enough that it should probably become more granular. Specifically: when the agent has two HIGH-tier quarter edges plus an SGP idea on the opposite side, that is 90 percent as strong a signal as three quarter edges. I am considering treating a two-plus-SGP alignment as an equivalent PASS trigger for the second half. This is a rule tune, not a change to the current logic — the current rule still fires cleanly when it fires.
Injury-dominant HIGH picks. The bot ships an injury_dominant flag on picks where more than 50 percent of the model edge comes from star-absence deduction. The framework currently treats these as PASS (they typically also carry a MEDIUM tier rather than HIGH). If a picks file ever ships an injury_dominant HIGH pick — which has not happened this season but could — the answer should be PASS regardless of tier. Formalizing that in the runbook is a second-half priority.
What comes next.
Regular slate resumes probably Tuesday July 28. The picks file for that night will populate overnight Monday. The first post-break slate is typically noisy because rotations shift for some teams (players who got extra rest, players returning from injury). The framework will apply the same discipline: follow bot HIGH picks at half stake unless conviction architecture fires (full stake) or a filter fires (PASS).
The Tempo return week is the storyline to watch on the franchise side. Jordan has the mid-season Tempo Report separately today. The Sykes injury remains the pivot for the second half.
Season-to-date read.
The framework rules are working. The rule that has produced the biggest live edge continues to be the one that saves us from PASSable HIGH picks (model-versus-agent conflict). The rule that most needs revisiting is the totals tier structure. The rule that most changed the bankroll's variance profile is the Issue 42 half-stake default.
The bankroll has held. The month-to-month has been choppy but the discipline has been consistent. The rules have evolved through six major recalibrations since May and are more robust now than they were at the start of the season.
Back Tuesday with the first post-break slate.
[ End Report ]
Share This