From September 15, every call this engine makes is logged and
graded in public — that's the record that matters. Behind it:
13 seasons of pre-registered testing (2012–2024), and a sealed
2025 that no method touches until it's spent as the next clean
confirmation.
Every public number Theorem prints traces to this page or the
prediction log behind it — no exceptions, no rounding up. When
someone asks whether the engine is any good, this URL is the answer.
Tested on seasons the methods never saw.Selection ran walk-forward on 2012–2023. Each family got
one confirmation run on held-out 2024, now spent. 2025 is sealed
— never evaluated, only consumed as input.
Bars declared before the run, never moved
after.A near-miss is a miss. Nothing gets re-scored once
the result is in.
Failures ship too.As honest copy, as
displayed facts with no lean attached, or as graves below.
Precision
note: methods were tested on 2012–2024 — walk-forward
selection on 2012–2023, one held-out confirmation on 2024. 2025
is consumed as input only, never evaluated.
The live record — starts with week 1
AS OF AUGUST 2026 · PRE-SEASON
The record starts with week 1. Every call the engine makes is
logged before games kick off, graded against what actually happened
on the Tuesday after, and printed here: what we said, what happened,
and the one fact that explains the gap.
Misses print at the same size as wins, and every miss names its
owner: the field, or us.
0 CALLS GRADED · WK 1 PENDING
Three ranking methods were
pre-registered before week 1 and are graded every week against the same
actuals; the winner ships next season. That race is internal and not
shown here.
Everything below is the historical testing
record that decided what ships.
What shipped, and its record
Draft board — top of the draft
Market consensus (ADP) beat our production baseline
at podium capture in 10 of 11 seasons (2013–2023), confirmed
on held-out 2024
10 of 11
Draft board — deep rounds
Past the ~150 players the market prices, last
season's production beat ADP at deep-pool ordering all eleven
years
11W–0L
QB streaming
Ranking available QBs beat “grab the injured
starter's backup” in 23 of 24 season-comparisons, and beat
season-average for rest-of-season picks (8 of 12). The 3-week
bridge ranking failed its gate and is not shown
23 of 24
Defense composite
Ranks available defenses better than season-average
alone, validated three times. Does not claim streaming beats
holding a good defense — that depends on the defense you
already roster
validated 3×
Waiver radar
Flagged players hit 2+ startable weeks in the next
4 at ~30% vs an ~12% base rate for the available pool (~2.6×
lift, 2012–2023). Availability gate: 11W–1L. It does
not claim to beat points-chasing on a clean pool — that came
out statistically even, and an earlier 2.5× figure is
retired
11W–1L
Kicker
Season scoring rate and volume only. No weekly
matchup claim — that layer failed validation; its grave is
below
season-level only
Near-ties — the only leans in the product
When two players project within a point of each
other, the projection has nothing left to say. Two leans survived
testing — barely — and ship at their pooled honest rates,
never the selection-flattering ones. A 51.5% lean is a coin flip with
a thumb on it, and the product says so in place.
Soft-opponent lean
Pooled 51.5% over ~8,150 near-tie pairs
(2012–2024). Selection ran 11W–1L; the 2024
confirmation was 50.1% — a letter-pass. The pooled number is
the one we print
51.5%
Game-script lean
Favored → runner, underdog →
pass-catcher. Pooled 51.1% over ~3,500 pairs; the 2024
confirmation was 57.6%, the best of the cycle, and the pooled
number is still the one we print
51.1%
TD rate at the tie margin
Chasing TD history at the tie margin lost its gate.
TD rate displays as a fact and earns no lean
2W–9L
Eight projections walked in. Season average walked out.
Before any of the above, the first gate was the projection itself.
Eight variants — recency blends at two weights and two windows,
opponent-adjusted versions, an opportunity-times-efficiency model, and
blends of those — ran on thirteen seasons against the plainest
baseline in fantasy: a player's season average. The baseline
won. Every cleverer variant lost or landed within noise of it, so
the engine runs the plain number and says so.
blend 60/40 + opp adjburiedblend 60/40buriedblend 30/70buriedblend 30/70 last-5buriedseason avg + opp adjburiedopportunity × efficiencyburied50/50 opportunity blendburied
That test never ended. It became the gate every
idea below had to face.
The graveyard — tested, failed, not sold to you
Every method here entered a pre-registered gate
— success bars declared before the run, judged on seasons it
never saw. These are the ones the bars killed, with the numbers that
killed them.
Hybrid draft board:
market order up top, production order deep — where the
crowd prices well, use the crowd; where it stops pricing, use the
data
SHIPPED · podium 10 of 11 · deep 11W–0L
Opponent-aggregate
adjustment — two independent builds, same grave
0W–12L, twice
Vegas lines, level
form — betting-market levels as a weekly input
1W–11L
Vegas deviation form
— buried at the pre-registered clamp margin (2.66%). The
only grave with a resurrection clause: a revival is pre-registered
for next season's boundary, declared in advance
buried at the margin
Usage-vs-points waiver
blend — every blend weight tested
dominated by points-only
TD-rate tiebreak
— regression bait at the tie margin; ships as a displayed
fact, no lean
2W–9L
Draft corrections v1,
rate-only — ignored volume and availability. Its podium
7W–4L is a boarded next-cycle hypothesis, not a
feature
1W–10L pool
Draft corrections v2,
totals-scaled — nearly tied, and nearly isn't a reason
to ship
5W–6L, both gates
Context bundle
(home + weather) — won rank quality by a hair, lost
lineup points by 2.9. The projection is unchanged, and that is the
gate working
failed confirmation
Better-team
tiebreak — 8W–4L in selection, 46.4% on the 2024
confirmation. Dead permanently
46.4% on confirm
QB 3-week bridge
matchup ranking — did not clear its own target
7 of 12
Weekly kicker matchup
layer — all three pre-registered criteria failed; weekly
kicker outcomes are close to pure noise, a fact about the
position
no signal
Most sites bury their misses. We bury ideas
— publicly, with the numbers that killed them.
The fine print on our own performance
Numbers a marketing page would leave out. This
page is not a marketing page.
The waiver radar's 2024 window was ugly.Its short-window hit rate that season was 11.3% against an
11.9% base rate. The pooled claim above includes 2024 — no
flattering windows.
The AI's instincts lost to the gates.Stated priors from the model that helped build this ran
roughly 1-for-4 across the testing cycle, and are recorded as such.
Gates outrank intuition here — including the intuition of the
tool that built the gates.
One run was voided.A “v2”
experiment silently re-executed v1; identical-to-the-decimal records
were the tell. The result was struck and the retry run for real. A
gate result only counts if the instrument actually ran.
The rules going forward
Engine 1.0 is frozen for the 2026 season and the ordering froze
before week 1 — a record you can't game by retraining away the
misses. Facts update weekly (injuries, roles, usage, byes); the method
holds still.
Frozen is not finished. From week 1, alternative methods run in
shadow — logged before kickoff and graded against the same
actuals as the live engine. After the first six weeks of real results
we re-evaluate: anything that beats the shipped method on the
season's own data ships as a versioned public release, announced,
with its own record starting from zero. 1.0's record stays exactly as
graded — improvement never edits the record.