Debunked: stacked per-coin specialists

1 · The appeal

Per-coin specialisation is intuitively attractive: each coin gets a model fitted to its own dynamics, weak coins can be excluded entirely, and the portfolio becomes a curated roster of proven single-name edges. Our Phase 1.7 program built exactly this — a large grid of per-coin cells, a promotion gate, and stacking experiments to blend the survivors into a book.

2 · The three failure mechanisms

1. Sample starvation. One coin's history is a few hundred thousand hourly bars covering a handful of regimes. A model fitted to it learns that coin's recent regime, not its structure — which is why our promoted specialists' subsequent performance clustered around the moment their training regime ended (Paper № 13 documents the live version of this failure).
2. Selection multiplicity. Choosing the best per-coin cells out of a large grid is a maximum over noisy draws. The June 2026 audit found the promoted set's backtest Sharpes consistent with selection on noise — the “curated roster” was substantially a leaderboard of luck.
3. Correlated drawdowns. The specialists' PnLs were not independent bets. Most single-coin models load on the same underlying factor — recent alt-market beta — so the stack concentrated exactly when diversification was needed. Market-neutral construction at the book level cannot fix correlation baked in at the signal level.

3 · The head-to-head resolution

The clean comparison came when both shapes were run through the same promotion funnel. The pooled cross-sectional ranker — one model trained on every (coin, bar) cell of a 50-coin panel, making one promotion decision — passed every pre-registered gate including 15/15 seeds positive on the causal walk-forward (Paper № 12). The per-coin stack never produced a configuration that could clear the same bar without gate-shopping. Two years of panel data supervising one model beats fifty slices of it supervising fifty models.

A nuance worth preserving: the per-coin work was not worthless. A small number of individual specialists showed genuinely bidirectional, persistent signal in paper trading, and single-name diagnostics remain the best microscope for understanding what the pooled model trades. The debunked claim is specifically the stack as a portfolio architecture — per-coin models as the unit of capital allocation.