TW3Cast achieved the 3rd place out of 130 entries on the GIFT‑Eval benchmark (mean MASE rank) as of 2026‑09‑14. Unlike the two higher‑ranked multi‑step agentic systems, TW3Cast runs without any agent or language model; it relies on a routing table computed once on the training split and then frozen, while its experts are public foundation models lightly fine‑tuned on that same split.\ \ For each of the 97 dataset‑frequency‑horizon configurations, the table offers four modes: a specialist mode (LoRA or full fine‑tune of Chronos‑2, TiRex or Toto whose training data were cleaned and enriched by explicit rules), a quantile blend that incorporates a specialist, a blend of base models, and a selection tournament played on a backtest carved from the training split.\ \ Every decision in the table was taken on that backtest. A specialist is admitted only when it beats the tournament, so a candidate costs a few megabytes and minutes of GPU time; a failed candidate leaves the routing unchanged. Three guarded mechanisms protect the selection from its own biases: a dual accuracy‑and‑calibration criterion, an asymmetric margin against candidates that saw the series during training, and conservative per‑window gates. The selection rules themselves were chosen inside a temporal meta‑backtest.\ \ In terms of performance, the best single base model reaches a mean MASE rank of 33.8, the tournament applied to every configuration reaches 38.0, and the full router reaches 19.4. The routing table, expert index, pinned base‑model revisions, submitted score file, and dated snapshot of the public scores are released, and every leaderboard number in this paper can be regenerated by a single script.\ \ Review