WikifitaGitHub live67e8de5
pesquisa · kaggle/pokemon_tcg_sampling_timezone_identity

Pokémon TCG — Sampling, Timezone and Entity Identity

Platform coverage, archive-window inference, ghost-team normalization and the limits these impose on Elo, ETL and curriculum claims.

Baixar raw

Pokémon TCG — Sampling, Timezone and Entity Identity

Three hidden variables in every result

The platform documents show that a tournament or training result depends on more than a model checkpoint:

  1. which official teams and replays are visible in the exported corpus;
  2. which archive window a replay belongs to under the platform's timezone behavior;
  3. whether a team name can be resolved to a stable entity across archives and submissions.

These are data and provenance variables. They should be joined to evaluation and curriculum records rather than treated as incidental cleanup.

Platform coverage

The platform analysis reports approximately 6,791 official teams but only about 1,059 teams represented in exported replays. Daily coverage was approximately 1.05–2.37% in July and reached about 5.57% on the highest August day. This creates blind spots around top-10, top-25 and top-100 comparisons.

Implications for this project:

  • --top-elo selects from observed daily coverage, not the full official population;
  • a strong local rating may be conditional on the sampled graph;
  • curriculum results can change when the visible opponent pool changes;
  • invariant Elo removes a common translation but not missing-edge bias.

Timezone inference

Episode IDs are monotonic, but archive ZIP boundaries and local reporting dates do not automatically share the same midnight. The timezone document interpolates a UTC cutoff near 00:20:42 from the observed ID ranges and notes a July 27 submission shift into the July 28 archive.

This is an inference from interpolation, not a hard platform contract. Any future daily curriculum or ETL query should store:

archive day
episode id range
local timestamp, if present
assumed timezone
inference method
confidence

Ghost-team normalization

The normalization heuristic uses two stages:

  • L1: the source report maps 142 of 144 ghost teams to leaderboard identities;
  • L2: the source report describes one additional identity approached through deck footprint or related evidence;
  • a residual Dieter identity has 1,068 matches requiring separate handling.

These are source-reported heuristic results, not a production identity table. The saved Antigravity probes use a wider episode window than the prose heuristic, so the exact method and confidence must remain visible in pokemon_tcg_entity_normalization.

These are useful heuristic mappings, not immutable identity facts. A mapping should retain the original string, normalized identity, rule used, confidence and unresolved alternatives.

Consequence for Wikifita

An Elo or ablation page that omits sample coverage, archive inference and identity resolution looks more precise than the underlying data. The correct editorial pattern is to put these analyses adjacent to the tournament ledger and ETL contract, not bury them as an implementation footnote.

Primary sources

  • docs/kaggle_platform_dynamics.md — official/exported coverage and blind spots.
  • docs/kaggle_timezone.md — archive-boundary interpolation.
  • docs/normalization_heuristics.md — ghost-team identity resolution.

Cross-references