Every claim ships with the source data, the method, and a way to reproduce. We win on some horizons. We lose on others. Both numbers are published. The methodology is what stays.
Eight methodology dimensions a sophisticated reader would ask about. KTC and FantasyCalc publish one number per player. FantasyPros publishes tiers. Our model publishes the math, the cohort, the curves, and the receipts. The cells below are specific rather than yes or no.
Dynasty General this product | KTC | FantasyCalc | FantasyPros | |
|---|---|---|---|---|
Public backtest Spearman vs actual production, multiple horizons | ✓Published | ✕Not published | ✕Not published | ✕Internal only |
Cohort calibration Real-roster percentile thresholds | ✓n=82, 5 leagues | ✕Single value | ✕Single value | ◐Tier breaks |
Per-position age curves Empirically refit peak bands | ✓4 positions, refit 05/12 | ◐Generic decay | ✕Implicit | ◐Tiered |
Lane identity model Multi-attribute roster classification | ✓11 lanes, 3 axes | ✕None | ✕None | ✕None |
Inflection scorecards Bimodal forecasts with signal scorecard + comparators | ✓Bimodal + signals | ✕None | ✕None | ✕None |
Tunable model User-controlled engine weights | ✓8 dials | ✕None | ◐Format toggles | ✕Expert blend |
Opponent dossiers Per-league trade fingerprints + pick patterns | ✓Per opponent | ✕None | ✕None | ✕None |
Reproducibility Source CSV + scripts published | ✓CSV + methodology | ✕Closed | ◐Public API | ✕Paid only |
We win convincingly on 1-year prediction. We trail FantasyPros ECR on 3-year. We publish both numbers and the loss function and the snapshot dates. Nothing cherry-picked.
Spearman rank correlation between each source's preseason ranking and actual cumulative fantasy points scored over the post-prediction window. Higher is better. Top-100 dynasty pool.
The honest read: Engine v1 wins convincingly on 1-year. We beat KeepTradeCut on every horizon we tested. We trail FantasyPros ECR and ADP on 2-year and 3-year. Both numbers are published; we do not cherry-pick. Full methodology + per-row reproduction at the scoreboard.
The 11 lane-identity thresholds (Win-Now Floor, Balanced, Future Stock, RB Bellcow, WR Anchor, WR Stable, QB Stable, TE-Premium Lock, Trade Capital, Sustained Contender, Zero-RB) are tuned against the percentile distribution of this cohort, not picked by feel. Baked 2026-05-08 from 5 dynasty and keeper leagues.
Your roster compared to the shapes of 82 real dynasty teams we benchmarked the model against. Each build below is a value-stack shape we found repeating across those real rosters. Your bar tells you how close your shape is to each pattern.
Your roster fits 1 build.
Strong fits: Future Stock. Partway into 1 more (Win-Now Floor).
What to do with this. The builds you fit name your trade leverage. Opponents who need pieces at those builds' signature positions are your best deal partners. Coach can route specific trade angles around these fits.
When a player is in a high-variance moment (aging cliff, rookie debut, post-major-injury return), the conditional distribution of outcomes is bimodal. The mean lands in the valley between the two modes, where no actual player ends up. KTC and FantasyCalc collapse this into one number anyway. We refuse to. Each inflection window publishes both stories with calibrated probabilities, a signal scorecard with confidence labels, and named historical comparators.
The model leans Story B (cliff) at 58% based on the 1,650 career carry load + team adding a young RB in the draft. YPC stability and pass-down share keep Story A alive at 42%. Single-point estimate would land in the valley.
| Signal | Observation | Confidence | Points to |
|---|---|---|---|
Career carries Cumulative regular-season + playoff carries | 1,650 (top decile for age 28 RBs) | validated | → Story B |
YPC trend Yards per carry, last 2 seasons vs prime | 4.6 → 4.4 (stable) | validated | → Story A |
Team RB draft Team drafted a young RB within last 18 months | Yes, pick 53 (round 2) | validated | → Story B |
Pass-down share Share of team passing-down RB snaps | 62% (locked in for now) | partial | → Story A |
OL continuity Offensive line returning starters | 4 of 5 returning | partial | → Story A |
Injury history Major-injury count past 3 seasons | data not in pipeline yet | weak | missing |
Per the calibration architecture: predicting a single point at a high-variance moment puts the estimate in the valley between modes, where no actual player ends up. The scorecard surfaces both stories, the signals that distinguish them with their confidence levels, and named historical comparators. The user reads the evidence and makes the call. The engine does not pretend to collapse a bimodal distribution into a point.
Most dynasty tools look at a manager's roster shape. We look at how they actually behave. The engine classifies each opponent's pick-trade history into one of four signature buckets (flipper, hoarder, seller, quiet) and feeds the framing to Coach so trade proposals match what the opponent actually does.
High two-way pick volume across 10 weeks. Sends and receives almost every week. Treats picks as currency rather than stockpiling them.
Lead with a pick swap when proposing a trade. Player-for-player offers from your side will sit. Asymmetric (you pick + you player for their pick + their player) tends to get accepted when the volume matches their pattern.
Four signature buckets the engine assigns: pick flipper (high two-way volume, net flat), pick hoarder (net pick receiver), pick seller (net pick sender), pick quiet (low volume, no strong fingerprint). Coach references the signature when proposing trades with that opponent so the offer shape matches what they actually do, not just what their roster looks like.
Move dials in the Rankings Lab; a doctrine line falls out the back. The line ("Aggressive Rebuilder · Soft Market · Confident") plus the dial-derived coefficient vector is what Coach references when it shapes a recommendation. Three preset doctrines side by side to show how the math actually changes.
Three preset doctrines from the Rankings Lab. The dial fingerprint + the synthesized doctrine line + the algorithm it implies, side by side. Tune your own at /rankings; the readout updates as you move dials.
Max future. Trade win-now for picks. Rookies and 2026-2028 upside, all of it.
Compete this season without torching future. Iceman default.
All chips in. Vets only, prime production, future is a rounding error.
Same engine, three different doctrines. The dial fingerprint produces a doctrine line ("Aggressive Rebuilder · Soft Market · Confident"), which produces a coefficient vector, which produces a ranking. Coach sees the doctrine line on every request and references it explicitly when the user has tuned dials off default.
A roster is not one archetype. It fits a vector across eleven builds, each with its own threshold calibrated against the cohort. Builds overlap by design; the composite builds (Sustained Contender, Zero-RB) are derived from the constituent base-build fits. Venn shows three builds for legibility.
One roster scored across three of the eleven builds. Builds overlap by design: a roster that fits Win-Now Floor AND RB Bellcow AND WR Anchor also fits the composite Sustained Contender build.
Fits when the top-K contributors' value sum exceeds the cohort threshold for proven win-now production. Calibrated from the 82-roster cohort.
Fits when the roster's RB room scores above the workhorse threshold. Specific to RB and weighted by role tier when the signal table fills out.
Fits when the roster's top WR clears the anchor threshold (high-value-70+ WR with surrounding production).
Derived from constituent builds. A roster that fits all three base builds above mathematically fits Sustained Contender. Architecturally distinct from a single-axis archetype assignment.
The Venn shows three builds for legibility; the full model scores all eleven (Win-Now Floor, Balanced, Future Stock, RB Bellcow, WR Anchor, WR Stable, QB Stable, TE-Premium Lock, Trade Capital, Sustained Contender, Zero-RB). A roster's full identity is a vector across all eleven builds, not a single archetype label.
Four position-specific curves derived from NFL starter production 2022 to 2025. Refits are dated and traceable. Most products use "RB cliff at 30" as a rule of thumb. We use a band that we update when the cohort moves.
Peak bands derived from 2022 to 2025 NFL starter production cohorts. The shaded region is the calibrated peak; the line is the engine's age multiplier from 20 to 36.
Narrowed 2026-05-12. Age 25 produced 0.78 of peak in starter cohort, not 1.0.
Widened lower bound to 23 on 2026-05-12. Year-1-3 starters (Daniels, Stroud, Caleb shape) post peak-band production.
Curves are refit when validation surfaces show the band edges no longer match data. Two refits on 2026-05-12: QB peak widened lower to 23 because year-1-3 starters now post peak-band production; TE narrowed to 26-30 because age 25 produces 0.78 of peak rather than 1.0.
Every Coach request ships the user's named players (name, position, age, team, KTC value), format rules (starter slots, K and DST presence, superflex, TE-premium), and a pricing block with KTC-anchored pick values. No "you need a TE" when Kincaid is on roster. Three-layer fix for every hallucination class: explicit context field, hard system-prompt rule referencing it, pre-LLM guard.
Available pool: 247 players Your roster: 23 players Question: should I trade for Bijan? (model has to guess at format, guess at what's on the roster, guess at the trade math)
format_rules:
is_superflex: true
qb_starters_max: 2
te_premium: true
has_k: false
has_dst: false
your_roster:
Patrick Mahomes (QB, age 30, KC, KTC 85)
Jahmyr Gibbs (RB, age 24, DET, KTC 100)
Travis Kincaid (TE, age 22, BUF, KTC 78)
[... 20 more]
pricing:
pick_values: { "2026.1.05": 92, ... }
player_values_present: true
question: should I trade for Bijan Robinson?Every fix ships with a test. Pre-deploy audits cover security, cost, legal, trade realism, and LLM context hallucinations. The lint rules are what catch a class of bug before it ships rather than after a user finds it.
Source CSV for the backtest: data/scoreboard/scoreboard_v1.csv. Cohort statistics: src/lib/strategy/lane-identity/cohort-stats.ts. Age-curve definitions: src/lib/strategy/lane-identity/lanes.ts. Every number on this page reproducible from the repository.