12 teams, head-to-head over 9 categories: FG%, FT%, 3-pointers made, points, rebounds, assists, steals, blocks and turnovers (fewer is better). Each Mon–Sun week you play one opponent and win, lose or tie each category separately; the season standings count category wins. 13 roster spots (10 starters a day, 3 bench) plus 3 injured-list slots and 6 free-agent adds per week. The top 6 make the playoffs in weeks 18–20 (Mar 1–21, 2027). You draft 1st; the order is a snake with a third-round reversal, so your picks are 1, 24, 36, 37, 60, 61, 84, 85, … — big gaps, then back-to-back pairs.

• Rates: per-minute stats from the last three seasons weighted 5/3/2 (2025-26, 24-25, 23-24) by minutes, shrunk toward the position-group (G/F/C) average with a 400-minute prior; an age curve adjusts production. Makes and attempts are projected separately; points = 2·FGM + 3PM + FTM.
• Minutes: recency-weighted, with growth for young players and decline after 33.
• Games played: a fitted regression on the last three seasons' games and age (with "ironman" terms), trained on 2014–2024 and tested out of sample on 2025 and 2026: RMSE about 2 games better than the old rule, which was ~8 games too optimistic. Each player also gets a calibrated uncertainty gp_sd(15–24 games, larger for injury-prone players).
• Injuries & news: ESPN/CBS injury reports and manual overrides (trades, roles, returns from injury, rookie priors). A small playoff-availability multiplier (0.98–0.99) covers resting in March.
• Stats come from a GitHub mirror of Basketball-Reference (verified identical to the site); ADP and position eligibility from Yahoo.

Each player gets a per-category score standardised over the draftable pool (12 × 13 players). Counting stats use the expected per-game line (games played × production, with missed games partly back-filled by a streamed replacement player). FG% and FT% are computed from makes and attempts: the score is the player's "impact" makes − pool% × attempts, so a 90% shooter on 2 attempts matters less than an 85% shooter on 8.

Scores are G-scores (Rosenof 2024): the denominator adds week-to-week noise, so volatile categories (steals, blocks, 3PM, turnovers) count a little less — right for weekly H2H. The headline value is the sum over categories above replacement level. It is used for ranking and the fallback list; actual draft decisions come from the optimizer below.

A team's weekly total in each category is approximated as Normal (counting stats sum; percentages are total makes / total attempts with delta-method variance). The opponent is a league-average team, built by drafting the pool in ADP order, with the spread between league teams added as variance.

P(win category) = Φ(difference / combined sd); expected category wins = the sum of the nine probabilities; P(win matchup) treats categories as independent. This is the objective the optimizer maximises.

• For each candidate at the current pick, it imagines the rest of my draft: at each of my later picks it adds the expected pick — each ranked player weighted by the probability he is the best one still on the board — then scores the finished 13-man roster with the category-win model (optimizer-0.4.0, per the latest draft simulation run).
• Availability (default since 2026-10-19): from the current board, 300 simulated rooms of 3 "sharp" drafters (best available by their own noisy view of player value) on random seats plus 8 drafters following a noisy ADP order; P(still there at pick p) is the share of rooms where he lasts. That is the "X% chance he's still there" on the Draft tab. The older analytic model (draft position ~ Normal(ADP, sd), sd = 2.5 + 0.18·ADP) treated the sampled position as the pick number and ignored drafters who don't follow ADP, which made the optimizer reach too early (docs/research/draft_policy.md).
• Value search (default since 2026-10-19): the optimizer's top 6 candidates are re-scored by a learned season-outcome model (P(title) / P(playoffs) over 48 paired simulated drafts each); its best candidate replaces the optimizer's #1 only if the gain is at least one standard error. In simulated drafts against that room this policy gained +1.9 pp P(playoffs) and +1.4 pp P(title) over the previous default (200 paired drafts, slot 1; simulation evidence, not a guarantee). Each saved recommendation records the policy that produced it (recommend_policy); if the value search fails, the optimizer's order is used with a warning.
• Third-round reversal and custom orders are built in: the rollout knows exactly when I pick next (e.g. 24 → 36/37), so it can pass on a player who will probably last and take one who won't.
• It also checks positional feasibility and scans punt strategies (deliberately giving up a category), keeping the best full 9-category team; so far "balanced" wins almost always.
• In 200 paired simulations it beat best-player-available by +0.074 ± 0.006 expected category wins per week (in-sample); margins below ~0.03 between two candidates are noise.

Many full drafts are simulated from my slot. The other 11 teams pick by Yahoo ADP plus noise; my picks come from the optimizer and, on the same simulated rooms, from a best-player-available baseline. The paired difference gives the edge with a proper standard error (sd/√n). It also yields who I tend to take at each pick, example rosters, and the spread of my category profile.

A Monte Carlo of the whole season (season-sim-0.3.0): every team's roster, the real 2026-27 NBA schedule (games per team per day), week-by-week H2H matchups, a daily 10-starter cap, streaming adds, standings and a re-seeded 6-team playoff bracket.

• Season-level uncertainty: each simulated player draws a volume factor and per-stat factors (log-sd 0.08), plus shifts in FG% (±1.5 pts) and FT% (±3 pts).
• Injuries: games played drawn from the fitted model; 75% of missed games come as one or two injury spells, 25% as single rest days.
• Before the draft, fantasy sim preseason drafts several simulated leagues (me = optimizer, others = noisy ADP), plays each one out and grades it with our own projections. That is the status-quo scenario, the optimistic end (96.5% playoffs / 54.8% title for slot 1). After the draft it runs on the real rosters.
• The headline outlook is a scenario range from the optimism decomposition: the same draft seeds replayed with 0, 3, 6 or 11 opponents drafting with our optimizer, each league graded under three "truths" (our projections, the market's ADP-implied view, a 50/50 blend). Central estimate (3 sharp opponents, blend): 83%–89% playoffs, 28%–35% title; pessimistic (11 sharp market-informed opponents, market right) 67% / 12%; slot 1 alone (all teams draft like us) 61% / 13%; average team 50% / 8.3%. Research note.
• Reproducible: the same seed gives the same answer, and roster changes are compared with common random numbers.

• In-sample: the same projections pick the players and score the results, and simulated opponents draft by ADP with no strategy. A team that "beats the market" by our own numbers looks dominant.
• Absolute levels are optimistic (e.g. ~6.3 of 9 categories, 90% matchup wins vs a league-average team, 96.5% playoffs in the status quo). The decomposition shows most of that comes from ADP-drafting opponents and from grading with our own projections, which is why the outlook is shown as a range. Use the tools for relative comparisons: candidate A vs B, add vs drop, which categories to chase.
• Game-to-game variance is fitted (var = a·m + b·m² per stat, on 2010–24 box scores) and beats the old constant-CV model out of sample on every stat; weekly variance is still under-stated by up to 46% for FGA, PTS, FGM and REB at 24–30 mpg (docs/research/sim_calibration.md, section 8). Uncalibrated priors: ADP spread, season-level error, injury-spell structure and streaming recovery are informed guesses until backtesting (T-010) and calibration (T-016). P(title) moves ±4–7 pts with them; P(playoffs) is robust.
• Weekly variance ignores schedule variance and category correlation, so single-week win probabilities are overconfident. Week dates other than the playoffs (weeks 18–20) are derived until Yahoo's schedule is loaded.
• Nobody manages in-season in the simulations: no trades, waiver breakouts or lineup tactics.