Build history
Pass 1 — original build (ChatGPT, ~90 minutes). 622 notes, 30 team desks, 549 player dossiers, 240-minute models, rankings, projections, source library.
Pass 2 — adversarial verification and repair (Claude, 2026-07-18 evening). 39 audit agents, 41 career-profile agents, 24 shot-volume agents, 33 correction agents, 36 wave-two agents, 12 strategy agents — roughly 185 subagents and 15.4M subagent tokens.
Pass 3 — sweeps, news, and format parity (Claude, 2026-07-19). 29 agents: the July 19 Dort/Nembhard/Risacher three-team trade applied across nine files with all three 240-minute models re-verified; a league-wide news sweep (LeBron still unsigned, Cisse offer sheet matched by Dallas, Summer League wraps); a systematic games-played audit of the top 150 producing 19 corrections (Data/gp-overrides.csv) that now flow through both boards and the tiers; the final 185 career records (coverage now 555 of 555); and both remaining strategy notes rebuilt on corrected data. Dual-format tiers published in 2026-27 tiers — ESPN points and nine-category. The HTML wiki (Encyclopedia wiki website) regenerates from the vault and passed a 61,000-link integrity check with zero breaks.
What pass 2 established
The factual foundation held. Every roster claim attacked survived independent checking: the Giannis trade to Miami, Trae Young to Washington, LaMelo Ball to Minnesota, the 2026 draft order, Khris Middleton's six-team sign-and-trade return to Washington, Dusty May's hiring in Dallas. Two auditor findings that alleged vault errors were themselves wrong and were rejected. The 2025-26 statistical baselines matched independent sources to the decimal in dozens of checks. A citation sweep of 1,195 of 2,958 unique URLs found exactly one dead link and no fabricated sources.
The projection engine did not hold. Three defects were found by reading Data/build_encyclopedia.py, none of them detectable by the structural audit:
| Defect | Scope | Repair |
|---|---|---|
fga = pts/2.15, fta = pts/4.8 fabricated when attempts were missing |
429 of 549 players; 160 of the top 200 | Real volume recovered for 262 players; attempts now derived from the scoring identity |
points_score added an undocumented +0.5 x 3PM |
Whole points board; 51 players moved 5+ slots | Term removed to match the documented formula |
Totals rank used z * sqrt(GP/82) on a signed z-sum |
52 players ranked better for being unavailable | Categories now multiplied by games played and re-standardised |
See Projection engine defects and corrections and 2026-27 volume-corrected nine-category board.
Completion snapshot
- 644 linked Markdown notes; roughly 852,000 words.
- 370 of 550 player dossiers carry a full career record: pre-NBA college or overseas production, season-by-season NBA table, fantasy-value history, durability record, and a trajectory read.
- 342 corrections applied across team desks and dossiers; 91 audit findings rejected as unverified after independent checking.
- 231 notes carry a dated correction line.
- 12 strategy notes covering league size, format, punt construction, aging curves, availability forecasting, draft-slot play, and streaming.
- Draft boards published for the two formats this league actually uses: 2026-27 draft boards — ESPN points and nine-category.
Working rules
- Every time-sensitive claim carries an as-of date and a direct source where available.
- Official roster, transaction, injury, and schedule evidence outranks commentary.
- A sourced analyst opinion is attributed; it is not silently converted into fact.
- Player rankings identify format and unit: per-game versus totals, categories versus points.
- An audit finding is verified independently before it is applied. Rejecting a wrong finding counts as much as applying a right one.
- A missing input propagates as a missing output, never as a silent default. This is the rule the original engine broke three times.
Known incomplete
Career records— complete as of 2026-07-19: 555 of 555 dossiers.Strategy notes on defective data— both rebuilt and re-verified on the corrected export, 2026-07-19.Games-played audit— systematic top-150 sweep completed 2026-07-19; 19 corrections applied through the boards. Below rank 150 the audit has not run.- Shot volume beyond the 262-player pool is deliberately not recovered. Players outside the pool fail the 20-game/15-minute gate and receive no rank at all, so proxy attempts in their dossiers cannot distort any board. This becomes worth revisiting only if the pool gate widens.
- Atlanta must shed two standard contracts after the Dort trade (17 on roster); the ATL minute model's conditional swaps need revisiting when the exits happen.
- The 2026-27 NBA schedule releases in August 2026 and nothing here reflects it. A pre-draft October re-verification pass remains mandatory.
Open questions
- How many other silent defaults remain in the build and audit scripts? Every defect found so far came from reading source code, not from running the audit.
- What would a semantic audit look like, and can it be scripted? The structural audit passes cleanly on a board built from fabricated numbers.
- Which projected roles remain contingent on unresolved transactions, injuries, or camp competition?
- How much of the July board survives October? A preseason re-verification pass on rosters, camp minutes, and injury clearances is mandatory before drafting.