Pokemon TCG - Research Backlog

Boundary

This page lists research directions, not implemented capabilities. Alefita keeps scientific decision authority. The page records option space and technical dependencies so future work does not get confused with the current system.

Current Implemented Base

The current architecture has recurrent MLX behavioral-cloning training, would-KO metadata and auxiliary heads integrated into the Parquet-era primary pipeline, deterministic action coverage, and strict-FP32 PyTorch submission inference. The Prospective V2 sidecar and lateral planner are historical design/runtime artifacts removed at a942373; their group-relative objective is a research antecedent, not a current sidecar runtime. The FP16 submission artifact belongs to the historical delivery record. See pokemon_tcg_prospective_v2, pokemon_tcg_training_overhaul_2026_07_29 and pokemon_tcg_current_state_reconciliation.

That is not strict on-policy GRPO and not a population self-play system.

Deferred Directions

Direction	Role	Status
AlphaFita / FitaStar	Project-level names for stronger search/self-play research lines	Backlog
AlphaStar-like population self-play	Maintain a population of policies, opponents, exploiters, and evaluation snapshots	Backlog
Strict on-policy GRPO	Sample branches from the current policy, retain behavior log-probs, compute group-relative advantages on-policy	Backlog
Fita GRPO continuation	Extend the historical group-relative prospective supervision framing toward a stricter policy-gradient path	Backlog
Task embeddings amortized across tasks	Condition policy/planner on task descriptors so repeated subtasks share representation	Backlog
Learned task weighting	Learn or adapt loss/task weights instead of manually fixed coefficients	Backlog
Deck selection	Optimize a deck outer loop from local self-play and matchup evidence	Backlog
Kimi K3 / AoR / KDA ideas	External research concepts to evaluate on their own terms before importing	Backlog
Adapter libraries	Opponent/archetype/date-specific low-rank modules with explicit leakage controls	Backlog
Population or PSRO-style deck-policy ecology	Model non-transitive matchups across decks and policies	Backlog

Strict GRPO Difference

The historical Prospective V2 objective can use group-relative advantages across counterfactual branches and may include behavior/reference log-prob terms when available. However, a strict on-policy GRPO claim requires a stronger contract:

branches are sampled from the current policy being updated;
behavior log-probs are stored for the sampled actions;
policy updates compare against those stored behavior probabilities;
data collection and update cadence define an on-policy or near-on-policy window;
reward normalization and clipping are tied to that sampled group.

The historical sidecar was offline and replay-derived. Its design can support future GRPO research, but it is not present in the current runtime and was not proof of strict on-policy GRPO.

Research Hygiene

Backlog items should carry:

date and source of the idea;
evidence needed before implementation;
illegal-information risks;
expected artifact changes;
tests that would distinguish real progress from data leakage or opponent-set overfitting.