Pokemon TCG - Deck Strategy Boundary
Deck selection and local self-play strategy under the immutable submission deck.csv contract.
Pokemon TCG - Deck Strategy Boundary
Competition Contract
The submitted artifact contains main.py and deck.csv at archive root.
deck.csv is the deck the Kaggle harness gives the agent for that submission.
Once a submission archive is built and submitted, that deck composition is an
immutable part of that submission artifact.
Runtime deck switching inside a submitted match is not part of the current
legal contract. The agent reads deck.csv; it does not receive an alternate
deck-selection channel or hidden state that would allow choosing a deck after
seeing private opponent information.
Local Outer Loop
Deck optimization belongs outside the submitted match:
candidate decks
-> local arena / self-play / matchup evidence
-> select one deck
-> write agent/deck.csv locally
-> build submission.tar.gz
The current dashboard supports this local outer loop. The deck builder can save
exact compositions, export CSV, compare a builder deck to agent/deck.csv, and
write agent/deck.csv with a .bak backup. The arena tab ranks local decks and
can replace agent/deck.csv with the locally best-ranked deck.
That is a local preparation mechanism, not a hidden runtime input.
Future suggested_deck.csv
suggested_deck.csv is a plausible future artifact for separating "what the
local optimizer recommends" from "what the current submission actually ships."
It is not implemented in the inspected code. Until it exists, agent/deck.csv
is both the local default deck and the submission deck source.
If suggested_deck.csv is added later, it should remain an outer-loop artifact:
the build process or dashboard may use it to update agent/deck.csv before a
submission is packaged, but the submitted policy still must not receive illegal
post-matchmaking deck choice or hidden opponent information.
Evidence Requirements
A candidate deck record should include:
- exact composition fingerprint;
- source and authoring path;
- local tournament or self-play evidence;
- opponent set and date;
- model/checkpoint identity used during evaluation;
- whether it was only suggested, locally active, or actually submitted.
This prevents deck evidence from being conflated with model evidence.