37_identity_fix.md

37 — Identity Reconciliation Fix (draft → game link rate)

37 — Identity Reconciliation Fix (draft → game link rate)

Date: 2026-05-30 Scripts: scripts/reconcile_identities_v2.py (merge engine) and scripts/build_reconcile_worklist.py (worklist builder) Outputs: data/processed/draft_picks_reconciled.csv, identity_merges.csv, canonical_id_remap.csv, identity_ambiguous.csv, reconcile_worklist.csv

TL;DR

The low draft→game link rate was a reconciliation bug, not reality. The original pipeline (scripts/reconcile_identities.py) gave every source its own canonical_id and only bridged across sources via discord stems and exact real names. For NSBA1 (nickname-only games) and the abbreviated-name NSBA3 game box, the same human ended up with two different canonical_ids — one carrying the draft real name, one carrying the in-game nickname — and they never linked.

After adding five new bridges and merging 43 canonical_id pairs, the link rates improved:

(The "12/72 = 16%" David quoted counts game-box records only; counting win_shares or player_season_master gives 14/72 as the pre-fix baseline.)

The remaining NSBA1 gap is not mostly a bug: the unlinked NSBA1 picks largely played under an unrelated in-game nickname (no algorithmic bridge exists) or did not play. Those, plus a handful of genuine ambiguities, are in the worklist for David to confirm by hand.

Root cause (per season)

• NSBA1 (nickname-only): game box uses in-game nicknames (arolakiv, cryo, VulcanForge, anonymouscapybara…). Draft + combine use real names. The combine workbook tab names are sometimes the nickname (Rosetta stone) but for ~25 entrants the tab is just the player's first/real name, so no nickname↔realname bridge exists in any data we hold. The NSBA1 discord export does not exist (only NSBA4 is in data/raw/discord/).
• NSBA2: game player_raw carries the full name#dddd discord tag → already linked at 100%. No work needed.
• NSBA3: game box uses real names, but abbreviated (Ishaan A, Anurag S) or bare first names (evan, daniel, Ryan). Draft/combine use full real names + discord. The original first-name-only fuzzy matcher refused to merge anything ambiguous, leaving 29 picks unlinked.

A second, subtler bug surfaced for NSBA3: several bare-first-name game ids are shared by 2+ humans (one evan id spans teams limitless and no like blue; one daniel spans okc and the other team; one Ryan spans lebron and usmic). These are NOT safe to merge — they need a split, which only David can do. They are flagged as ambiguous_collision in the worklist.

Bridges added (and merges made)

Plus shared-discord-tag (discord_tag) and exact multiword real-name (exact_realname) bridges that the original engine already had (re-applied here for completeness; they produced 0 new cross-id merges beyond what existed).

Total: 43 canonical_id pairs merged → 646 ids collapse to 603. Full list in data/processed/identity_merges.csv.

Conservatism guards (no false merges)

• A bare-first-name game id that appears on >1 team is treated as multiple humans and is never auto-merged (→ ambiguous_collision worklist).
• NSBA3 team-name merges require exactly one team-local draft candidate.
• NSBA1 team-fuzzy uses greedy 1-to-1 assignment with a score floor so two drafted players can't both grab the same game record.
• Multi-candidate first+last-initial cases (two Daniels, two Ishaans) go to the worklist rather than guessing.

How to apply downstream

canonical_id_remap.csv (old_id → new_id) is the single source of truth. Any downstream table keyed by canonical_id should be remapped through it. The draft join should use draft_picks_reconciled.csv (adds canonical_id re-resolved + linked_game). The merges have NOT been folded back into canonical_players.csv to avoid silently rewriting every dependent artifact; remap explicitly when you rebuild.

Worklist for David — data/processed/reconcile_worklist.csv

52 rows, sorted by importance (early draft picks & high win-share players first). Three kinds:

1. ambiguous_collision (4) — DO THESE FIRST (data-quality risk)

One game canonical_id is shared by multiple drafted humans; confirm the split.

2. draft_unlinked (38) — early picks with no game record

Each has best-guess game candidate(s) + a confidence. Highest-value:

Many other NSBA1 early picks (Gid, Ben, Yucc, Jonathan Huang, Caleb, Srijan Oduru, Atharva, Saheb Gulati, William Huang…) have no candidate — they either played under an unrelated nickname or did not play. Confirm/deny each.

3. game_high_ws_unlinked (10) — high-WS NSBA1 nicknames with no identity

These scored a lot but we can't name them. Tell us who they are:

arolakiv (WS 5.10), reverse (2.69), TechnologicalError (1.47), somath (1.38), 4435 (1.31), Fez_Keyreb (1.28), abcisosm5 (1.26), cryo (1.22), Fumarole (1.10), Fishface (1.00).

Confirmation format

Open reconcile_worklist.csv, add a verdict column with y/n (and the chosen candidate where there are several). For ambiguous_collision rows, note which game record belongs to which person so we can split the id. Re-run reconcile_identities_v2.py after wiring confirmed pairs into a small override table and the link rate climbs further automatically.

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