40_draft_strategy_v3.md

40 — Draft Strategy v3 (specialist-corrected)

Date: 2026-06-02. Supersedes the strategy in 26_draft_strategy_preliminary.md (v1) and the KEY_FINDINGS resynthesis (v2). Trigger: David + Anurag pushed that the value engine "still favors generalists way too much," and that fading Kian and Vishnu is wrong. We red-teamed our OWN generalist bias (scripts/A_specialist_skeptic.py) and re-ran a forensic on the two fades. They were partly right. This file folds the correction into an actionable draft plan.

What did NOT change: the spine still holds — draft for projected production rate, coverage is cheap, team-building is additive (no synergy lever), exploit the field's mechanical raw-combine anchoring (F8, ~13 pts/slot). And the elite tier of the board is unchanged (corrected vs raw top-10 overlap 7/10): the bias does not change who you take in rounds 1–2. The correction bites the mid-board (rounds 3+).

1. The generalist bias is REAL — but not where we assumed

Claim. PPTF = net toss-up points / toss-ups faced, and the denominator counts all 6 cycling categories. A 2-category specialist is charged for ~4 categories of dead tossups their teammates would cover, mechanically deflating their rate vs a flat mediocre-everywhere generalist. Mechanism confirmed in the data: specialists' in-category PPTF is ~3.0× their all-category PPTF (generalists only 1.37×); and breadth correlates +0.56 with the all-category metric but −0.30 with in-category rate. So F15's "the elite tier is broad" is substantially a measurement artifact — the metric calls a player broad because the all-category denominator rewards scoring everywhere.

But the direction of the fix is the opposite of D8. Under a defensible coverage-marginal (set-cover) lens, the genuinely under-priced specialists are bio and ESS bodies (risers' strengths: bio 12/20, ESS 9/20). CS is NOT scarce at the top — 10 of 14 board elites are already CS-capable, so CS-strong players actually fall 12.5 spots under the corrected lens while non-CS rise 6.6. This refutes D8 ("reach for the scarce CS specialist"). Don't reach for CS; if you value scarcity, value bio/ESS depth.

Source: data/processed/specialist_adjusted_value.csv (88 nsba2/3 player-seasons), scripts/A_specialist_skeptic.py. Spearman(raw, corrected) = 0.787; median |rank shift| 10 spots; corr(n_strong, rank_shift) = −0.60 (specialists +4.8, generalists −17.3).
Confidence: Med on direction; Low on magnitudes (n tiny; naive in-cat PPTF over-credits low-volume one-trick players — trust the high-volume risers).
Caveat: elite tier stable; this is a mid-board re-ordering tool, not a round-1 lever.

2. The corrected value lens (how to actually use it)

For rounds 1–2: use the existing board (raw projected PPTF/VORP). The bias doesn't move it.

For rounds 3+: rank by coverage-marginal value — what a player adds to the categories your roster still needs, not their diluted all-category rate. A player elite in 1–2 subjects that you haven't covered is worth more than their PPTF rank says. Use the coverage_marginal_rank column once your early picks define your holes. Concretely: - A buried specialist with in_cat_pptf ≫ all_cat_pptf and a high dilution_ratio who fills an open category = a real value the field (and our old board) underrates. - A generalist whose strengths overlap categories you've already filled = redundant; fade relative to their raw rank.

3. Depth in your best subject > thin 6-coverage (the D14 lever we'd demoted)

Claim. Under the home-advantage rule (D14 — the home team strips the opponent's best subject), a team thin in its best subject loses 24.0% of production when stripped, vs only 9.4% for a team with a 2nd body there (corr(loss_frac, has_depth) = −0.73, n=38). The regular-season PPTF analysis is blind to this; it credits concentration for free.

Action. After you've cleared 6-subject coverage (the snake draft hands you that — F4), the next marginal pick should double up in your strongest category for playoff insurance, not spread into a 7th body of thin breadth. This is a playoff/elimination tiebreaker, not a regular-season win lever (depth-over-talent partial ≈ 0 in the regular season) — but NSBA is won in elimination, so it matters at the margin. - Source: D14 simulation in scripts/A_specialist_skeptic.py. Confidence: Low (static strip-and-substitute estimate, one rule, n=38).

4. Kian and Vishnu — the resolution (Anurag is half-right)

Forensic re-exam (full game logs, teammate context, contestation-adjusted, per-subject tape):

Vishnu M (oof7373, P0619) — FADE STANDS, high confidence. 27 games (most data on the whole board), consistent 0.37 PPTF (average-plus) behind a combine that climbed 14→24→34 (#4): textbook combine inflation. The stacked-team story fails — 2 of 3 seasons he was on weak teams and his best PPTF came on his weakest team; the teammate adjustment moves him down. The CS-scarcity rescue is actively refuted: CS is his worst in-game subject (10 buzzes → net −8 pts); his real value is math/ESS, the least scarce categories. He's a fine mid-board math/ESS body — do not pay the #4 sticker.
Kian D (pine2359, P0346) — DOWNGRADE FADE → NEUTRAL, medium confidence. Anurag is right that he's not a trap. In 9 nsba3 games he's a top-25 producer (0.533 PPTF, ~2× the field; win-shares 3.28; 47% point-share) and a genuine bio/chem specialist (in-subject bio 1.71). The fade only ever meant "not the #1 overall his combine #1 implies," and that's true — his combine is recall-heavy (bio/chem/phys) and climbed +8 (biking pattern). Correct treatment: draft him as a strong bio/chem specialist value in the late-1st / 2nd-round tier — not at the #1 sticker, but not a fade. (Bonus: bio/chem is exactly the thin-at-the-top cluster from §1, so his specialist value is real coverage, not redundant.)
Source: forensic on game_player_box, tossups_long, usage_bias_32, adjusted_ability, combine_player_category. Net: D7 holds — the tape isn't cheating either of them; one is correctly faded, one was over-faded.

5. The combine-absence opportunity (David's GM intel: Sohil R, William W, Boheng + others)

Returning players who skip / haven't taken the NSBA4 combine create a structural steal: per F8 the field drafts the visible raw combine almost mechanically — a player with no combine row is invisible to that anchor and will slide. If you have game tape on them, you buy at a discount the field literally cannot price. This compounds with §1 for buried specialists.

William Wang (P0629) — strong buy, fits the thesis perfectly. nsba3 chem specialist: all-category PPTF only 0.24 (raw rank 59, near-invisible) but in-category chem 1.44 — rises +18 spots under the corrected lens. A buried chem specialist (chem = a thin-at-top category per §1) with no combine = double-invisible. Exactly who you target rounds 3–5.
Sohil Rathi (P0559) — good player, but a fader on the corrected lens; price as solid, not a steal. nsba3: strong raw production (0.55 PPTF, raw rank 13, 0.94 win-shares) — but his strengths are cs/math/phys (the quant cluster the board elites already cover), so he falls −22 spots corrected. He'll "put up crazy numbers" the way a lone star on a weak team does (point-share 0.31 on a 3–8 team), but his categories are well-supplied. Take him if he slides to value (he will, with no combine) — just don't reach; he's redundant with a quant-heavy early roster.
Boheng — least information. Combine/draft history but no clean game tape in our tables (drafted, no matched games — possibly an availability question, see F10). Treat as pure GM-intel / eye-test until tape or a combine row appears. Confirm with David.
General rule: maintain a "no-combine returner" watchlist priced off game tape + GM intel only. These are the highest-leverage slides in a field that anchors on the combine.

6. The v3 draft rules (operational)

Rounds 1–2 — best projected producer, period. Board unchanged; the bias doesn't bite here. Take the broad elites (they genuinely win — F15 direction survives). Don't overthink.
Rounds 3–5 — switch to the coverage-marginal lens. Value in-category eliteness that fills your open categories; fade generalists redundant with what you've already drafted. This is where the ~13-pts/slot edge (F8) actually lives now.
Don't reach for CS. Refuted (§1). Secure one credible CS body opportunistically if it falls; the field already covers it. The scarce-at-top categories are bio and ESS.
After coverage, buy depth in your best subject, not a 7th thin body (§3, D14 playoff).
Exploit combine-absence (§5). No-combine returners with game tape slide — grab them. William Wang-type buried specialists are the model target.
Never trade R1 down for volume (F9, unchanged; your NSBA2 trade was −2.73 WS).
Fades/buys: Vishnu stays a fade (don't pay #4); Kian is neutral-to-value (don't pay #1, do take as a 2nd-round bio/chem specialist); Rohan G remains the cleanest game-backed steal (F12/F39 — self-fades "washed," tape says top-tier).

7. Honest limits

n is tiny (88 player-seasons, 38 team-seasons, 3 game seasons; nsba3 uses estimated TUH). Every magnitude is ±a lot — trust directions and rankings.
The corrected lens validates out-of-sample only marginally better than raw VORP (coverage-marginal↔win% +0.158 vs +0.120) — it's a better-aimed draft tool, not a proven predictive upgrade. Use it for mid-board re-ordering, don't bet rounds 1–2 on it.
The set-cover marginal assumes winner-takes-most per category, no synergy (consistent with the F14 additive null — but it's a model).
Blocked on David: real draft slot (~06-06), confirmed eligible/captain list (Open Q14/17), Boheng's status, and combine completion for the no-combine returners.

NSBA Draft Analytics · embargoed until after the SSB draft · ← hub