Snake Cult / Myth lab Study 001
Where do myth
trees begin?
A tree connects stories. Its root proposes a beginning. Change the assumptions. See what survives.
Enter the experimentLoading the research trees…
THE TREE
Scroll sideways to see the labels
Branch spacing is schematic. It does not show age, distance, or confidence.
READ THE RESULT
The interpretation
A better fit.
A harder question.
Moving a root changes which branches appear early. It does not change who is connected to whom. Under a stationary, reversible model, those different beginnings have the same likelihood.
Here, a directional model allows the distribution of traits at the root to differ from the long-run distribution. That gives the data a way to prefer a beginning. The result depends on that assumption—and on which traits were recorded.
The best-fitting sampled branch is a model result. It is not a date, an ancient population, or a demonstrated birthplace.
The calibration 2,900 datasets / 4,600 fits
Can the method recover
a known beginning?
Plant a root in a simulated history. Let the model search for it. On a Cosmic Hunt–derived tree with 93 traits, even a stronger directional signal rarely leads back to the exact root.
01 / A BEGINNING WE KNOW
Exact root recovered
Each row tests 600 simulated histories. The same histories are fitted on the known tree and on a tree rebuilt from the traits.
Weaker signal s = -0.8
Stronger signal s = -1.5
Point estimate Wilson 95% interval
“Exact” means the planted branch is the sole best fit. Low recovery persists even when the model is given the true connections.
03 / THE OBSERVED COSMIC HUNT
Does its gain clear the null?
The directional model improves fit by about 3.277 log-likelihood units. Rootless simulations exceed that gain often enough that it falls below both calibrated 95% cutoffs.
Observed: 3.277 Cutoffs: 5.551 / 5.721
Plus-one tail probabilities: 0.3831 / 0.3881 (known / rebuilt). Each uses 200 null calibration datasets.
EXPLORATORY STRESS TEST / 50 PAIRED HISTORIES
Repeated traits can make the fit look stronger.
In one terminal-root test on rebuilt trees, copying 12 underlying traits into 93 columns increased the gain while reducing exact recovery.
- Mean gain
- +24.4log-likelihood units
- Exact recovery
- −20percentage points
What this establishes. A test of this model on a Cosmic Hunt–derived tree and missingness pattern, using 93 binary traits. It neither proves an alternative homeland nor disproves an African origin. The stronger signal helps, but the selected root still needs more evidence than its rank.
Design, denominators & audited data
The primary panel plants roots at 12 preselected locations, with 50 histories per root at each signal. Known-tree fits receive the generating connections, then refit branch lengths and model parameters. Rebuilt-tree fits use neighbour joining (NJ); their recovery rates include cases where reconstruction loses the true split. The dots show pooled exact recovery, with Wilson intervals for simulation precision.
Signal is the difference between root and equilibrium presence frequencies on a log-odds scale. The two-state transition process is reversible; rooting information comes from a non-equilibrium root distribution. Under the stationary null those frequencies are equal, so there is no identifiable true root. Each connection mode has 200 calibration and 200 independent held-out null datasets. Held-out cutoff exceedances were 11/200 in each mode.
The accepted campaign completed all 2,900 datasets and 4,600 all-edge fits with zero failed or missing results. An independent audit reconciled every flattened fit to its raw result. Successful deterministic continuations resolved 37 optimizer termination flags without changing a selected edge. The frozen production scores remain the source of this display.
The correlated-copy result is an exploratory paired point estimate, not a general error-rate claim. An independent 206-trait comparison had positive recovery differences in all four tested cells, but every 95% interval crossed zero; a reliable larger-matrix improvement remains unresolved.
The summary preserves exact counts, intervals, model scope and SHA-256 provenance for the accepted source artifacts. It exports aggregate evidence, not the private research workspace.
Open the research notes
What the controls change
Each fitted comparison holds the connections in the tree fixed, then refits the root and model parameters. Fit differences compare roots within the selected model and matrix. They are not probabilities that a culture invented a story.
Where available, the trait-selection controls ask whether the conclusion changes when accounting for how motifs entered the matrix. A changed winner is a useful result: it shows which assumption carries the inference.
Method background: Felsenstein on reversible models and rooting; Lewis on correcting for character selection.
What the rows mean
The dragon and men–women matrices pool traits by region. Cosmic Hunt and Cinderella contain versions of a recognisable story. Neither kind of row is an ancient ancestor. Cultural borrowing and incomplete sampling can also make a tree an imperfect description.
Sources for the selected study
Source references are included in the downloadable dataset.