## gallery.leaderboard()


The board. One row per model, ordered by model name. No sort key.


Usage

``` python
gallery.leaderboard(panel)
```


**There is no `sort_by=` argument and no default sort.** Held-out ELPD and held-out CRPS are published side by side and the reader picks (decided 2026-07-25). A sort parameter would move that choice from the reader into this function, and whichever value people typed first would become the house ranking by habit. Rows come back in model-name order so a published CSV is byte-reproducible; `DataFrame.sort_values` is one call away, and [SCORE_DIRECTION](gallery.SCORE_DIRECTION.md#ibnr.gallery.SCORE_DIRECTION) says which way each column runs so a renderer cannot guess wrong.

Columns:

**The two score columns can rest on different cells**, because a density and draws are independent capabilities with independent panels. Each therefore carries its own `n_cells`, member list and fingerprint, and there is deliberately no bare `n_cells` column for a reader to attach to the wrong one.

`task`, [as_of](Triangle.as_of.md#ibnr.Triangle.as_of) the protocol and the training cutoff `model` entry name `n_cohorts` cohorts in the union of the two panels `elpd_members` the density-eligible models the ELPD panel was intersected over. **The ELPD column is a function of this set**, so it travels with it, the same discipline `mart_publish_id` gets `elpd_fingerprint` sha256 of the sorted ELPD panel keys `n_cells_elpd` the ELPD panel - identical on every row `elpd` SUM of the pointwise log predictive density over the ELPD panel. `pd.NA`, never 0, when unavailable `elpd_per_cell` `elpd / n_cells_elpd`, for reading. It cannot change a ranking (`n` is the same for every row) and it is NOT a repair for a coverage difference - it is here so a reader can tell 0.05 from 0.5 on a total near -14,000 `elpd_status` `scored`, or `na: <reason>` `n_cells_zero_density` cells whose pointwise ELPD is `-inf`. A model with `elpd = -inf` must show WHY on the same row, or the total reads as a bug rather than a verdict `min_ess_kish` smallest effective draw count behind any cell **on the ELPD panel**. Every per-score readout is taken from that score's own slice, never from the union - a cell on the CRPS panel only would otherwise describe a column it is not in `crps_members`, `crps_fingerprint`, `n_cells_crps` the same three for the CRPS panel, which is a DIFFERENT set of cells `crps` SUM of the pointwise CRPS over the CRPS panel. Lower is better - the opposite direction to `elpd`, which is why [SCORE_DIRECTION](gallery.SCORE_DIRECTION.md#ibnr.gallery.SCORE_DIRECTION) is machine-readable `crps_per_cell`, `crps_status` as above `n_draws_density_min`, `n_draws_sample_min` smallest draw count behind any of this model's cohorts, per capability. `logmeanexp` is biased downward by about `cv2/(2S)`, always in the same direction, and sample CRPS is biased by `O(1/S)` too, so a model that never escalated through the harness's sampler stages scores worse for a reason that is not the model `n_cells_elpd_own`, `n_cells_elpd_dropped`, `n_cells_crps_own`, `n_cells_crps_dropped` this model's own coverage per score, and what the intersection cost it

**The totals are reduced in Python from each model's own array, not with a pandas aggregation.** `groupby().sum()` over an all-missing group returns `0.0` - measured here on pandas 2.3.3, and a nullable `Float64` dtype does not change it, only `min_count=1` does. On this board 0.0 is the best ELPD there is and the best CRPS there is, simultaneously. Making the aggregation unreachable beats guarding it.

**There is no standard-error column, and no pairwise `elpd_diff`.** The SE needs a decision this module must not make silently: the honest clustering unit is the FIT, and the fit is not the cohort for every entry - `sur` and `copula_glm` fit once per company across its lines, and the NN entries fit once over the whole panel, so a cohort-clustered SE would be wrong for four of the candidate entries in the same direction (too small). `panel.by_cohort` is the input any such estimator will take.

**There is no KS / PIT column.** The critical value `1.36/sqrt(n)` assumes independence: nine cells of one cohort share one posterior and one calendar year, so at cell level that value is about three times too tight and models would be flagged as miscalibrated when they are not. It is also a different estimand from the published milestone-3 KS, which is one PIT per cohort on the run-off-to-ultimate predictive.