Dual Conviction Engine

Both Models Agree — Confirmed Picks

Highest Confidence

Team / Pick	Dual Score	Linear vs Vector	Status	Best Archetype	Key Edge
Houston Championship	96	L 94 V 99	Confirmed	Average Champion	KenPom dual + Sampson 3x FF + 7¢ on Poly vs 9% book = biggest edge. Vector confirms profile matches 22/23 champs.
Arizona Championship	96	L 94 V 99	Confirmed	Average Champion	Closest vector match to the "average champion" of all 12 teams. Balanced profile, 1-seed, KenPom #2. Nearly identical to past winners.
VCU ML vs UNC	92	L 89 V 95	Confirmed	Grinder	Vector matched VCU to "Grinder" archetype — experience-heavy, low public attention, hot momentum. Same profile as Villanova '16. Wilson OUT for UNC seals it.
Santa Clara ML vs Kentucky	89	L 85 V 93	Confirmed	Grinder	Classic 5v12 profile. Vector confirms the "nobody picking us, experienced grinder" pattern that produces 35.6% upsets over 40 years.

Flagged Disagreements — Vector Says More, Linear Says Less

Review Required

Team	Dual Score	Linear vs Vector	Delta	Status	Archetype
Iowa State Championship	94	L 88 V 99	+11	Flag	Grinder
Duke Championship	90	L 82 V 97	+15	Flag	Balanced Powerhouse
S. Florida ML vs Louisville	89	L 82 V 96	+14	Flag	Grinder
Michigan Championship	88	L 78 V 98	+20	Flag	Balanced Powerhouse
Florida Championship	88	L 76 V 99	+23	Flag	Average Champion

Iowa State — L:88 V:99 (+11)

Vector matched Iowa State to the "Grinder" archetype at 0.998 cosine — the highest archetype match of any team. Their profile (elite defense, slow tempo, experienced, under-the-radar) is nearly identical to Villanova '16/'18 and Baylor '21. The linear model dinged them for being a 2-seed with moderate coaching score, but vector says their overall shape is champion-tier.

VERDICT: Upgrade to STRONG BUY. Vector insight is compelling — this is the "quiet champion" archetype.

Duke — L:82 V:97 (+15)

Vector loves Duke's overall profile (Balanced Powerhouse, 0.991 cosine), but linear correctly penalizes Foster OUT, freshman-heavy roster, and massive public backing (fade signal). The disagreement is meaningful: Duke's ceiling is champion-tier, but the floor dropped with Foster's injury.

VERDICT: HOLD. Both models are "right" — elite ceiling but real injury risk. Don't add to Polymarket at 21¢.

Michigan — L:78 V:98 (+20, biggest gap)

Largest disagreement in the system. Vector says Michigan's overall profile is nearly identical to Duke '10/'15 and UNC '09 (Balanced Powerhouse). Linear hammered them for being the most public-backed team (12.1% tickets — our "fade public" factor). This is the key debate: does profile trump public sentiment?

VERDICT: HOLD/SMALL BUY. Michigan at 19¢ isn't cheap enough for the public fade risk. But if they drop to 15¢, vector says pounce.

Florida — L:76 V:99 (+23, second biggest gap)

Vector gives Florida a 0.997 cosine to Average Champion — the defending champion's profile just looks like winners. Linear was cautious because Florida's seed (#1 South) may be slightly inflated, and their ATS trend is weak. But the Tampa venue boost (+0.40 above avg) is massive in the vector model.

VERDICT: Upgrade to BUY. At 11.3¢ vs 13% sportsbook implied, there's a 1.7¢ edge. The venue boost is real and linear under-weights it.

S. Florida ML — L:82 V:96 (+14)

Vector matched to Grinder archetype at 0.978 cosine. Same pattern as VCU — experienced, low public attention, conference tourney momentum. The gap from linear is because S. Florida's raw talent metrics (KenPom, conference strength) are weaker than VCU's opponent situation.

VERDICT: SMALL BUY if line stays at 2.81x. The archetype match is strong but the confidence band is wider than our confirmed picks.

Both Models Say Pass — No Edge

Avoid

Team	Dual Score	Linear vs Vector	Why
Connecticut	56	L 65 V 47	Fails KenPom dual filter. Stewart injury uncertain. Vector distance to all champion archetypes is 2.06+ (huge). Despite Hurley's pedigree, this team doesn't look like champions.
Illinois	54	L 62 V 47	Fails KenPom dual filter. Low coaching tournament experience. Vector shows this profile is distant from every champion archetype. 2-seed is a mirage.
Purdue	53	L 58 V 48	Fails KenPom dual filter. Painter's tourney history is mixed. Last year's runner-up curse. Vector confirms: not a champion profile.

How the Dual Engine Works

Methodology

Linear Model (L)

Weighted sum of 7 factors: KenPom dual (25%), seed history (20%), ATS trends (15%), coaching (15%), injuries (10%), venue (10%), liquidity (5%). Each factor scored 0-100, multiplied by weight, summed. Simple, transparent, manually adjustable. Good at capturing individual edges (injuries, public fade). Weakness: treats all factors independently.

Vector Model (V)

Each team is a 17-dimensional vector. Cosine similarity measured against 3 historical champion archetypes (Defense-First, Balanced Powerhouse, Grinder) built from 23 champions (2003-2025). Weighted by dimension importance. Good at detecting "this team looks like past winners." Weakness: can over-index on profile similarity and miss context (injuries, public betting).

When Linear > Vector

Linear is better at: context-dependent edges (a star player's injury won't show up in the vector until we manually adjust it), contrarian signals (public fade data), and platform-specific pricing gaps. Trust linear when the edge is about market mispricing rather than team quality.

When Vector > Linear

Vector is better at: holistic profile matching (does this team "feel" like a champion?), detecting hidden similarities to past winners, and catching teams where individual factors are mediocre but the overall shape is elite. Trust vector when evaluating championship potential over a 6-game tournament.

The Decision Rule: When both models agree (within 8 points), conviction is highest — size up the position. When they disagree, investigate why. The linear model's context-sensitivity usually wins on single-game upset picks. The vector model's pattern-matching usually wins on championship futures. The dual score (average of both) provides the final ranking, but the agreement flag is what determines position sizing.

Final Action Summary — Dual Engine

Execute

Action	Pick	Dual	Agreement	Sizing
STRONG BUY	Houston Championship @ 7¢	96	Confirmed	Full Half-Kelly (~$25-30)
STRONG BUY	Arizona Championship @ 17.8¢	96	Confirmed	Full Half-Kelly (~$20-25)
STRONG BUY	Iowa State Championship @ 4.8¢	94	Flag +11	Upgraded — Half-Kelly (~$15-20)
STRONG BUY	VCU ML vs UNC @ 2.14x	92	Confirmed	Full Half-Kelly (~$20)
BUY	Santa Clara ML vs Kentucky @ 2.34x	89	Confirmed	Half-Kelly (~$15)
BUY	Florida Championship @ 11.3¢	88	Upgraded	Quarter-Kelly (~$10) — vector upgraded
HOLD	Duke Championship @ 21¢	90	Flag +15	Don't add. Foster injury = real risk despite elite vector.
HOLD	Michigan Championship @ 19¢	88	Flag +20	Most public-backed team. Buy only if drops to 15¢.

Both Models Agree — Confirmed Picks

Flagged Disagreements — Vector Says More, Linear Says Less

Both Models Say Pass — No Edge

How the Dual Engine Works

Final Action Summary — Dual Engine