David v Godzilla: TabFM

A 12 GB tabular foundation model, met under a pre-registered protocol. Eleven arms against the reference. The record, in full.

The tale of the tape

Round one: eleven arms, 226 stratified series

Pre-registered: universe, arms, and scoring frozen and committed before any result existed. Eleven ways of using TabFM (lag depths, four binning schemes, horizon variants, a residual-density regressor), so the verdict bounds the adaptation rather than one adapter. laplace won the per-series likelihood race against every arm, 71–88% of series, the strongest configuration trailing by 0.39 nats median. The strata told the story: near parity on strongly mean-reverting continuous series (a tabular regression problem, TabFM's home game), collapse on repeat-heavy series (2–4 nats; no concept of an atom), steady deficit on near-random-walk series where volatility tracking is the job.

Round two: the strongest arm, on the full universe

Pre-registered again, one arm this time: clfx100, the strongest of the eleven (a per-observation density head that bins 100× the parade z, not a lag variant), taken to every qualifying non-price series (921 scored of a frozen universe) with the analysis plan fixed before any result was read.

Raw, TabFM still trails: median −0.155 nats, laplace ahead on 76% of series, −0.22 family-weighted. The selection effect the plan anticipated showed up, the 226-series 0.39-nat gap regressing to 0.15. One stratum broke the other way. On repeat-heavy series raw clfx100 beats laplace outright (+0.35 median, n=181), exactly the lattice its discrete head can express and laplace's continuous leaf cannot.

The verdict

On density forecasting as a standalone challenger, laplace by a wide margin, at one ten-thousandth the footprint. On point forecasts, TabFM is genuinely decent (it takes the MAE undercard on continuous series by hairline shading). The one genuine upset is that repeat-heavy stratum, where the discrete head expresses a lattice laplace's continuous leaf cannot. That same structure, recovered inside laplace's coordinates rather than against them, is a sandwich finding rather than a challenger victory, and it became a library work item. The cooperation numbers in full: the sandwich.

All numbers from committed studies: benchmarks/tabfm_wide_study.py and benchmarks/tabfm_x100_wide.py, results CSVs beside them, statements in benchmarks/preregistrations/.