26_draft_strategy_preliminary.md

26 — DRAFT STRATEGY (DRAFT v2) — NSBA4

26 — DRAFT STRATEGY (DRAFT v2) — NSBA4

Status: v2 WORKING DRAFT. Written for David (GM) to read, 2026-05-30. This is a working draft, not the final board. It is built directly on the canonical KEY_FINDINGS.md / SYNTHESIS.md (A30 synthesis, red-teamed) and now folds in the second analysis wave (findings 27–31) and its skeptic audit (finding 92). Every number below is a small-n point estimate — trust the directions and rankings, distrust the decimals. All numbers were pulled live from outputs/ and data/processed/ with the project venv. Real names are used throughout so you can sanity-check against your NSBA1/2 memory.

The nsba4 board in §3 is incomplete and provisional — see the heavy caveats there.

0. WHAT CHANGED IN v2 (read this first)

Five new analyses (F27–F31) and an adversarial audit (F92) ran since the preliminary draft. Net effect: the spine is unchanged and slightly firmed up — none of the new work overturns "draft best individual value-rate." The new waves mostly close doors the v1 draft had left open. Concretely:

1. The bonus-value pass resolved — and it does NOT rescue slow-but-smart players. Bonuses are 55–61% of the scoreboard (the GM critique was right on arithmetic), but there is no exploitable per-player bonus signal: book-knowledge does not predict bonus conversion (3 independent methods, knowledge coef −0.013, ns). You win bonuses by winning the toss-up to earn them, which PPTF already captures. v1's open worry — that a bonus model would upweight knowledgeable-but-slow bodies — is dead. Do not value a slow player for hidden bonus upside. This removes the tension v1 flagged against the speed tilt: speed-for-toss-ups stands, with no bonus counter-weight. (See new §1.BONUS; the old "BONUS-VALUE in progress" note is replaced.)

2. Duo synergy is a clean (bounded) NULL → team-building is ADDITIVE. No archetype-pair or disjoint-subject synergy survives FDR (0/8). Draft the best individual value; do NOT chase "fit" or complementarity. (Skeptic verdict: TRUST.)

3. Archetypes are real but soft → target broad-elite producers; CS is not its own type. Five soft subject-types exist, but only the broad "elite multi-science" tier reliably wins. Labels are feature-dependent (leave-one-subject-out ARI 0.52) — use directions, not labels. Confirms (independently) that CS is not a standalone archetype; secure one CS body, don't build around CS.

4. The "new undervaluation edge" was CIRCULAR — demoted to a watchlist. F30's headline ρ=0.60 "validation" is ~90% a shared-PPTF artifact (combine's own independent link to the slot-residual is 0.007, p=0.95). There is no new edge here. The real, established edge is unchanged: ~13 toss-pts/slot from drafting value over raw-combine rank (F8). The forward nsba4 board stays as a soft watchlist for re-ordering within tiers only. (Skeptic verdict: DISCARD the validation; watchlist the board.)

5. Playoffs: lean broad elites (low confidence). ESS specialists fade hard in playoffs (retention 0.44) while broad chem/phys elites hold or rise (1.47) — but this is nsba3-only (n=5 elites, one season). A tiebreaker for playoff equity, not a re-ordering lever.

Reproducibility gaps to flag: the F31 (field-depth) script is missing from scripts/ (only the artifact playoff_robustness_31.csv exists — F31 is partially non-reproducible), and the bonus-value script referenced by F27 (analysis_bonus_value.py) is also not in scripts/. Both should be restored/committed before these are treated as hard inputs.

1. CORE STRATEGY

The spine survives every red-team: draft for projected production RATE (PPTF / VORP), treat coverage + one CS body as cheap insurance the snake draft hands you for free, and exploit the field's mechanical anchoring on the raw combine board. Concretely:

• Draft on projected game value, not combine rank. The field drafts the visible raw combine almost mechanically (|ρ| 0.67–0.75 every season). That combine predicts realized value only at ρ≈0.36–0.43. The gap is a recoverable ~13 toss-points per slot (~65 over a 5-round draft) — this remains the one real undervaluation edge. Blend combine θ with game-history PPTF wherever a player has tape; for unproven players use shrunk, de-biked θ. NEW (F30, skeptic-corrected): there is NO additional undervaluation edge. The v2 "value-over-field" metric's headline ρ=0.60 "validation" turned out to be ~90% a circular shared-PPTF artifact (combine's own independent link to the slot-residual is 0.007, p=0.95). So do not stack a second value-over-field number on top of F8's ~13-pt/slot edge — they are the same edge, and only F8's is honestly estimated. The forward nsba4 "value-over-field" board (§3e) survives only as a soft watchlist that predicts which names the field will misrank within a tier — not as a new source of points.
• Concentrate the edge in rounds 3–5. Round 1 is roughly efficient (the field's board and the value board agree near the top). The anchoring leak — and therefore your edge — is largest in the middle rounds, where game-backed mid-combine players slide.
• Never trade Round 1 down for volume. The pick-value curve is steeply convex (R1 retains only 46% of its value into R2; tradeable chart pick-1=100 → R1 100 / R2 42 / R3 22 / R4 15 / R5 12 / R6 11). Paying linear prices undervalues early picks. Late picks are lottery tickets (R1 bust 15% vs R6 73%). See the NSBA2 back-test in §2 — this is the single most expensive mistake we have a real data point on.
• Coverage is table-stakes, not a lever — and team-building is ADDITIVE. Drafted teams reach 95–100% six-subject coverage automatically. The "don't punt" penalty is null when measured from roster design (corr −0.11, p=0.56). NEW (F29): there is no duo-synergy lever either — no archetype-pair or disjoint-subject combination over-performs the additive sum of its members (0/8 tests survive FDR; the clean team-level complementarity↔ synergy corr is +0.025, p=0.88). So: draft the best individual value-rate, and do NOT chase "fit," complementarity, or "this pair covers each other." A pair's production is its two solo baselines added. (Honest bound: a true synergy smaller than ±0.3 win-shares/pair would be invisible at n=38 — so "additive" means "no exploitable fit edge," not "proven exactly zero.") Rule: don't actively punt a subject — but never pay for breadth or fit. Wins are star/rate-driven; the drafted 3rd man is ~0 with win%.
• Target the broad-elite producer; treat single-subject types (incl. CS) as interchangeable coverage. NEW (F28): five soft subject-archetypes exist, but only the broad "elite multi-science" tier (chem/phys/math-deep AND above-average everywhere) reliably wins — it carries ~4× the win-shares and essentially all the VORP of any specialist type. Pure bio / ESS / math-phys specialists are middling, roughly interchangeable coverage (~0.3–0.4 PPTF). Use the direction (broad-and-deep wins), not the labels — clusters are soft (silhouette 0.23) and feature-dependent (leave-one-subject-out ARI 0.52), so an individual player's archetype tag can flip on a different feature set. This is a re-statement of the spine, not a new lever: chase the broad-elite value-rate, not an archetype.
• CS: secure ONE good body, then stop — and CS is NOT its own archetype. The old "3.5×" is dead — CS is ~2× raw, collapsing to 1.2–1.4× under SOS/team controls (CI spans a factor of several). NEW (F28): high-CS players cluster with the broad elites, not in a standalone CS pod — there is no "pure CS specialist" type to build around. Use theta_cs (the honest r=0.60 signal) to identify ONE credible CS body; a 2nd/3rd is redundant. Do not reach, and do not treat CS as a roster pillar.
• Discount for availability. 37% of nsba3 drafted players never played a clean game (vs 0% in nsba2). Multiply projected contribution by max(reliability_prior, flag_prior); for unknowns use the round base rate (R1≈0.85 … R6≈0.23). The discord availability_flag tracks this monotonically.
• Buy game-backed sandbaggers, fade combine-inflated "bikers." One reversion+growth model only — never stack a growth bump on a reversion shrink, and never credit a brand-new entrant with returner "development."

1.NEW — Exploit the packet-submission difficulty shift (subject-deficiency + speed)

NSBA3/4 questions are team packet-submissions and are measurably easier than the alumni-written NSBA1/2 packets (Finding 20, clean 4,318-question logit): nsba3 conversion 0.819 vs nsba1 0.783 / nsba2 0.738 — i.e. nsba1/nsba2 are ~9–10% harder per buzz. nsba4 is the same packet-submission era as nsba3, so the easier-question regime carries over. That changes what wins, in two compounding ways:

(a) Target subject deficiencies — the subjects packets write easiest. Per-subject nsba3 conversion (the nsba4-comparable reference):

The exploit: in Bio and ESS, almost every question is gettable, so the marginal knowledge edge is small and the question is essentially "who buzzes first." These are the subjects where a fast, broad answerer converts nearly everything and a slow specialist adds little over the field. Conversely CS and Math stay genuinely hard — that is where specialist knowledge still separates teams, and where a missing answerer actually costs you live points. So:

• In easy subjects (Bio, ESS, and the easy half of Physics): prioritize SPEED — fast, confident generalists who win the buzzer race. Easier questions reward whoever gets there first, not who knows the most. You do not need a Bio "specialist"; you need a fast trigger who covers Bio.
• In hard subjects (CS, Math): this is where you actually spend a roster slot on depth/knowledge — but per the CS finding, exactly one good body, identified by theta_cs, no reaching.

Why speed, mechanistically: in the natural-roster baseline, corr(win%, avg celerity) = +0.78, on par with conversion (+0.79) and just under raw points. Celerity isn't directly logged for NSBA games, so it transfers as a prior, not a measured quantity — but the direction is strong and it lines up exactly with the easy-packet regime: when conversion ceilings are high (Bio 92%), the buzzer race is the game.

Caveat on 1.NEW: the per-subject factors are nsba3-only baselines (155–264 q each), conversion conflates question difficulty with field strength, and "speed" is a transferred prior from a different (school, Energy-not-CS) dataset. Treat this as a tilt to apply at the margin between similar players, not a hard rule that overrides the VORP board.

v2 update — the speed tilt no longer has a bonus counter-weight. v1 worried this speed bias might be wrong because a forthcoming bonus model would rescue slow-but-knowledgeable players. It does not (F27, below): book-knowledge does not predict bonus conversion, so there is no hidden bonus value in slow-but-smart bodies. Speed-for-toss-ups therefore stands unopposed at the margin. Do not hold a "slot of doubt" for proven deep-knowledge players on bonus grounds — that door is closed.

1.BONUS — why production RATE is the right currency (and bonuses don't change the board)

The GM critique was: bonuses are 10 pts vs 4 for toss-ups and are where games are won, so PPTF (toss-ups only) must badly undervalue a knowledgeable-but-slow player who lifts team bonus conversion. The bonus pass (F27, 181 clean games, 3 methods) confirms the arithmetic and refutes the player-valuation implication:

• Scoreboard truth: bonuses ARE 55–61% of all points (net-toss-up basis 0.61, gross 0.55; per season 0.51–0.57). They decide a real minority of games — ~15% are won despite losing the toss-up battle, on bonuses alone. So the GM is right that bonuses matter on the board.
• But conversion is knowledge-agnostic and not per-player attributable. Bonus conversion is a real, varying team skill (~49%, std 0.20) — yet it does not load on book-knowledge: player-conversion ↔ combine theta corr +0.016; ON/OFF lineup effect −0.034; roster knowledge → conversion coef −0.013 (p=0.37) and −0.008 (p=0.57). Three independent estimators, all non-positive. The eight actual "knowledgeable-but-slow" players convert bonuses at 0.542 vs league 0.489 on n=8 — noise, not a hidden edge.
• The mechanism: you earn bonuses by winning the toss-up — which PPTF already captures. A player who can't buzz never brings the bonus to the table. Crediting full bonus value to the board re-ranks it by ≤1 spot once you strip out toss-up volume (Spearman vs PPTF rank 0.945 full-credit / 0.996 knowledge-premium-only). The movers are high-volume toss-up scorers, not slow-smart bodies.

Implication for the board — production RATE is the right currency precisely because bonuses amplify toss-up wins. Every toss-up you win is worth its 4 points plus a ~0.49 shot at a 10-point bonus, and that amplification is roughly constant across players. So the player who wins more toss-ups (high PPTF) also harvests more bonus points — automatically. Do NOT add a per-player bonus term, and do NOT value a slow-but-smart body for "hidden bonus value" — there is none. The only honest bonus-aware adjustment is team-level (a ~±0.2 roster-trait swing, low-confidence), never per-player.

One team-level caveat, flagged honestly: roster subject-breadth correlates with win% at +0.53 at the team level. Do not bank on this. It is almost certainly talent- confounded — the same confound that killed the realized-coverage "don't punt" penalty (it collapsed from −0.54 to a null −0.11 once rebuilt from roster design). Breadth here is a proxy for "this is a good team," not an independent breadth lever. Treat it as table-stakes (don't actively punt), not a reason to pay for breadth or for slow deep-knowledge bodies.

2. NSBA2 RETRO-OPTIMAL DRAFT (validation — you know these players)

Using realized win-shares (the exact decomposition: player shares sum to team wins), here is the "if we'd drafted on value" board for nsba2_2023, with each player's actual draft slot. 60 picks, 12 GMs, snake.

Top realized value vs where they actually went

The biggest STEALS (drafted far later than they earned)

• Mihir K — pick 60 (the literal last pick), 1.27 WS. Best value-per-slot in the draft.
• Rohan G — pick 48 (R4), 2.05 WS — a top-11 player taken in the 4th round.
• thedoge — pick 16 (R2), 3.50 WS — the #3 realized player, a full round late.
• lolmao (pick 17) 2.69 WS and minitarrasque (pick 15) 2.14 WS — top-10 value at R2.

The biggest BUSTS (early pick, little/no production)

• Sean — pick 9 (R1), 0.50 WS — an R1 pick that returned replacement-level value.
• owen fei (cityblock) — pick 12 (R1), ~0.19 WS — last pick of R1, near-zero.
• pyrrolysine — pick 13 (R2), 0.32 WS; lavar ball — pick 21 (R2), 0.31 WS — early-R2 picks that produced like late-round fliers.

The pattern that validates the thesis: R1 was mostly efficient (8 of the top-9 WS players went in R1 or early R2) — consistent with "the top of the board is hard to beat." The recoverable edge lived in R2–R5, exactly where the field's combine anchoring leaks: Mihir (60), Rohan (48), thedoge (16), lolmao (17) were all gettable far below their value.

The ideal NSBA2 roster a value-GM could have built from a mid slot

A value-drafting GM picking around the middle (the snake hands ~slots 6–7 in R1) and never trading R1 down could plausibly have assembled, in realistic reach order: Joshua Wang (5, 2.87) → thedoge (16, 3.50) → lolmao (17, 2.69) → Rohan G (48, 2.05) → Mihir K (60, 1.27) — roughly 12+ WS of toss-up production from five picks, versus a typical drafted core of 4–6 WS. The steals alone (thedoge + lolmao + Rohan + Mihir) total 9.5 WS that the field left on the board in rounds 2–5.

Your actual NSBA2 team ("David" / Xpoes's X-riskers)

Your game roster scored 4.0 win-shares — best players Stephen Chen (trivial) 0.99, Akul Saxena 0.88, cymbidium 0.71, Adhitya Chandra (Cyaniphor) 0.66. You finished 4–6, .400, 9th of 13. A balanced, coverage-complete team with no star — the exact profile the data says doesn't win (wins are star/rate-driven; depth is ~0 with win%).

The trade: 1st-for-two-2nds — it cost you a star (−2.73 WS)

You held 0 R1 picks and 3 R2 picks (slots 19, 22, 23) vs the normal 1+1. Your natural R1 seat was slot 6; Connor carried three R1s (4, 6, 11). So you shipped your R1 (slot 6) to Connor for two extra R2s. The realized ledger:

Net = 0.50 − 3.23 = −2.73 WS. The pick chart priced this near break-even (R1 slot-6 ≈ 72 pts ≈ two R2s at 34+28). Reality was a blowout because the chart prices the average R1 pick — the actual player there was a league star, and the convex curve's whole point is that a realized R1 ceiling dwarfs two R2 floors. To rub it in: Connor, holding your gifted R1 → Daniel Sun, finished 7–3, 2nd, built around exactly that pick. This is n=1 and Daniel Sun could have busted — but it lands precisely where theory said it would, and it is why the "never trade R1 down for volume" rule is the firmest line in this document.

3. NSBA4 PRELIMINARY BIG BOARD

READ THIS FIRST — HEAVY CAVEATS. - The nsba4 combine is INCOMPLETE: only 52 players have a combine row so far, and only 3 of the top-25 historical-value players are in it (Akhil Batchu, Rohan G, Kian Dhawan). The deep value pool (Sanjay, Yufei, thedoge, Joshua, Anurag, Daniel Sun, lolmao…) has not registered / combined yet as of this pull. The board below is what the current partial pool supports, not the final draft. - Eligibility and captains are UNCONFIRMED. Several listed names may be captains (uncbuyable), staff, or ineligible. Ziang Z is the open question — captain or draftable? Resolve before acting. - Draft slot is UNKNOWN. "#9" is a placeholder; VONA/trade plans can't be finalized without it. - Santhosh V is EXCLUDED — confirmed cheat (biking + leaking; his combine is inflated). He appears high on the raw theta_cs list; do not draft.

(a) Returning known-value players in the nsba4 pool — by projected value (VORP)

These are the players with real game tape who currently hold an nsba4 combine row. This is the trustworthy core of the board.

Everyone below Chris Wang projects below replacement on current data (Mihir K, Aldric B, Nihar Bhave, Roshan A, etc.) — they are role/coverage bodies, not value picks.

(b) Combine-only newcomers — by de-biked θ (HIGH UNCERTAINTY)

No game history; price off shrunk, de-biked θ only (no growth bump — they have no prior to revert). se_overall ≈ 0.45–0.51 on every one of these, so the ranking is soft.

Treat all of (b) as fliers. The combine→value link is r≈0.6 and optimistic (no-shows excluded), so an unproven high-θ newcomer is a round-appropriate dart, not a core pick.

(c) CS targets — by theta_cs (the honest r=0.60 signal)

Secure ONE. Ranked, excluding the cheat:

1. Vishnu M — theta_cs 1.31 (also has the most in-game CS reps; the safe pick if his availability/age is acceptable).
2. ~~Santhosh V (1.06)~~ — EXCLUDED, cheat.
3. Harry G — theta_cs 1.00 (newcomer; doubles as your best raw-θ flier — efficient if one slot buys both CS and upside).
4. Travis D (0.82), Daniel Lu (0.78), Edward C (0.67) — fallback CS bodies.
5. Rohan G (0.57) — already a top-3 core pick who also covers CS adequately → may let you skip a dedicated CS slot entirely (this is the cheapest path: get CS for free off a value pick you wanted anyway).

Recommendation: if you land Rohan G, you likely don't need a separate CS reach at all. Otherwise Vishnu (proven) or Harry G (proven-on-combine + upside) is the one body. Do not take a second.

(d) Buy-low / fade list (actionable subset — entries in the nsba4 pool)

BUY-LOW (game tape or strong intel beats their combine — field will let them slide): - Mihir K (modernnewton) — mispricing +1.21, conf 0.71; game-backed, returning, rusty bio-only but a known producer (1.27 WS in nsba2 as the last pick). - Rohan G — +1.00, conf 0.79; already in your core, and the field underrates him. - Ziang Z (deasert_willow) — +1.90, conf 0.40 (INTEL ONLY): NSBA staff/writer, scored a low 7 on combine, called "genuinely sandbagging." High upside IF draftable — confirm he isn't a captain/staff-ineligible first. - Riyan N (nocombomomento) — +2.02, conf 0.40 (intel): returning, self-IDs CS as a best subject and "CS is unbikable" — a credible cheap CS angle. - Akhil B (+0.41) — already core; the field will let him slide off a "wing it" combine.

FADE (combine-inflated; the field will overdraft them — let someone else): - Vishnu M (oof7373) — mispricing −1.20, conf 1.00 (the most-data fade on the board). High combine, declining game value, aging out. Exception: if you specifically need the proven CS body, he is still the safest CS pick — fade him as a star, not as a CS role. - Kian Dhawan (pine2359) — combine rank #1 overall, but NOT #1 in games (−0.76). The textbook combine-inflated early name the field will reach for. Let them. - Suzuko O, Varyan J, Shannon I, Roshan A — combine ≫ realized; coverage bodies at best.

(e) Value-over-field watchlist (F30 — soft, NOT a new edge)

The v2 "value-over-field" board (outputs/undervaluation_board.csv, forward nsba4 pool) is kept only as a soft watchlist for re-ordering within a tier — it is not a new source of points. Its headline "validation" (ρ=0.60) was shown to be ~90% circular (shared-PPTF artifact; combine's own independent link to the slot-residual is 0.007, p=0.95), so the real edge it points at is just F8's established ~13 toss-pts/slot from value-over-raw-combine. Use it the way you'd use the buy-low board: as a flag for which names the field will misrank (it anchors on raw combine), concentrated in the middle rounds, never to invent a tier and never as significance (BH-FDR: 0 of 94 names individually significant). It inherits soft archetype labels (F28) and the combine selection-bias caveat. Among its combine-only forward flags, Sanjay O is the most actionable buy (only elite-multi-science archetype in the top buys); Santhosh V independently re-lands on the FADE list — consistent with the confirmed- cheat exclusion (do not draft him regardless).

(f) Playoff-robustness tiebreaker (F31 — directional, LOW confidence)

If two players grade near-equal on the value board, lean to the broader chem/phys-flavored producer for playoff equity. In the one clean playoff season (nsba3, 12 games, n=5 elites), within-player retention split sharply: broad chem+phys elites rose to ~147% of regular output while single-subject ESS specialists collapsed to ~44% (their one subject is the one that hardens most under playoff difficulty). Both flagged rows survive leave-one-out within the season, but this is nsba3-only, n=1 season, no cross-season replication — a tiebreaker, not a re-ordering lever. The combine does not detect playoff-robustness (it predicts regular-season PPTF better than playoff PPTF; retention corr with theta ~0.1, ns) — so do not pay a combine premium expecting playoff upside.

Reproducibility flag: the F31 script is missing from scripts/ (only the artifact playoff_robustness_31.csv exists) — restore it before hardening this into anything.

EXECUTIVE SUMMARY (10 lines)

1. Draft for projected production RATE (PPTF/VORP), not combine rank — the field anchors on raw combine (ρ≈0.7); that leaves ~13 toss-pts/slot of recoverable edge.
2. Never trade R1 down for volume. The pick curve is steeply convex (R1 = 100, R2 = 42).
3. Concentrate your edge in rounds 3–5, where combine-anchoring leaks most; R1 is efficient.
4. Coverage + one CS body are free insurance; team-building is ADDITIVE — don't actively punt, but never pay for breadth or "fit." No duo synergy survives FDR (0/8); draft best individual value, not complementary pairs (a real synergy <±0.3 WS would be invisible).
5. CS ≈ 1.2–1.4× (not 3.5×) and is NOT its own archetype: secure ONE body via theta_cs, then stop. Target the broad-elite producer (the only type that reliably wins); use archetype directions, not labels (clusters are soft, ARI 0.52 leave-one-subject-out).
6. NEW — exploit easier packet-era questions: in easy subjects (Bio, ESS) prioritize SPEED/fast buzzers (the buzzer race is the game at 90%+ conversion); spend real knowledge slots only in the hard subjects (CS, Math).
7. Discount for availability (37% of nsba3 picks never played); buy game-backed sandbaggers, fade combine-inflated bikers.
8. NSBA2 RETRO HEADLINE: drafting on value, R1 was efficient but the steals lived in R2–R5 — Mihir K (last pick, 1.27 WS), Rohan G (R4, 2.05), thedoge (R2, 3.50). Your 1st-for-two-2nds trade cost a star: you gave up Daniel Sun (3.23 WS) for 0.50 → −2.73 WS, and Connor rode that pick to a 2nd-place finish.
9. NSBA4 PRELIM TOP NAMES (proven core): Akhil Batchu, Rohan G, Kian Dhawan (the only 3 of the top-25 currently combined); CS = Vishnu M / Harry G (or free off Rohan G); newcomer fliers = Harry G, Ryan K; FADE Kian-as-#1 and Vishnu-as-star.
10. BONUS RESOLVED: bonuses are 55–61% of the scoreboard but carry no exploitable per-player signal (knowledge → conversion coef −0.013, ns) — production RATE already captures them, because winning the toss-up is what earns the bonus. Do not value slow-but-smart bodies for hidden bonus value; do not add a per-player bonus term.
11. UNDERVALUATION = no new edge: the v2 "value-over-field" ρ=0.60 was circular (combine's own link to slot-residual 0.007, ns). The only real edge stays F8's ~13 toss-pts/slot (value-over-combine); the forward board is a soft watchlist, not points.
12. PLAYOFFS (low conf): lean broad chem/phys elites — ESS specialists fade (retention 0.44) while elites hold/rise (1.47), nsba3-only, n=5. A tiebreaker, not a lever.
13. BIG CAVEAT: the nsba4 combine is thin (3 of top-25), eligibility/captains unconfirmed, slot ("#9") is a placeholder, and the F31 + bonus-value scripts are missing from scripts/ (reproducibility gap). This is a working draft — trust directions, not decimals.

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