# Myth Lab calibration notes

This is a curated public summary of the accepted `root_recovery_2026` campaign.
The accompanying `calibration-summary.json` contains exact counts, intervals and SHA-256
provenance for the source artifacts. It is a summary export, not the complete numerical archive.

## What was tested

The experiment simulates binary traits on a Cosmic Hunt–derived 47-tip tree, then asks a fitted
model to recover a planted root. Its primary size is 93 traits, with the exact five-cell source
missingness mask. A two-state reversible transition process allows the presence frequency at
the root to differ from the equilibrium frequency. That non-equilibrium distribution supplies
directional information; the transition process is not a general nonreversible model.

The controlled panel uses 12 preselected roots with 50 histories per root at each of two signals.
Signals −0.8 and −1.5 are root-minus-equilibrium log-odds contrasts. Each root is equally weighted
in the pooled summaries. “Known tree” supplies the generating connections, but branch lengths
and model parameters are refitted. “Rebuilt tree” reconstructs connections using neighbour
joining (NJ). Exact recovery means the generating split is the sole winning edge within 1e-6
log likelihood; unconditional NJ recovery includes cases where reconstruction loses that split.

## Primary results

| Signal | Known tree | Rebuilt tree (NJ) |
| --- | --- | --- |
| Weaker, −0.8 | 39/600 = 6.50% | 32/600 = 5.33% |
| Stronger, −1.5 | 86/600 = 14.33% | 67/600 = 11.17% |

Wilson 95% intervals appear in the JSON and chart. They describe Monte Carlo precision under
this design, not uncertainty about a historical homeland. Stronger signal improves recovery,
but exact-root recoverability remains low even with the correct connections supplied.

Under the stationary null, root and equilibrium frequencies are equal. There is no identifiable
true root and therefore no meaningful root-recovery accuracy. Nonetheless, 772/800 fits (96.5%)
select one unique best edge: 389/400 on known trees and 383/400 on rebuilt trees. The 800 fits
come from 400 simulated datasets, each fitted in both modes.

The observed Cosmic Hunt directional gain is 3.277441746 on the known tree and 3.277441653 after
NJ reconstruction. The corresponding conservative 95% null cutoffs are 5.550961557 and
5.720708555, fitted from 200 calibration datasets per mode. The observed gain is below both.
Calibration exceedances are 76/200 and 77/200; plus-one Monte Carlo tail probabilities are
(76+1)/(200+1) = 0.3831 and (77+1)/(200+1) = 0.3881. These are model-conditional tail comparisons,
not probabilities that an origin claim is false. Independent held-out exceedances are 11/200
in each mode, with Wilson 95% interval 3.10%–9.58%.

## Secondary stress evidence

Correlated-copy tests repeat 12 latent traits into 93 columns. In the terminal-root, weaker-signal
NJ comparison, the mean gain increases by 24.392 log-likelihood units while exact recovery falls
by 20 percentage points. This is an exploratory paired point estimate from 50 histories, not a
precise general error-rate claim. Internal and rootless comparisons also inflate gain; repeating
with 5% independent flips produces still larger gain inflation. Selected exact stress contrasts
and their denominators are preserved in the JSON.

An independent 206-trait comparison produces positive recovery differences in all four tested
cells, but every Newcombe 95% interval crosses zero. Each size has 50 independent histories per
cell. These are unpaired samples and do not resolve a reliable larger-matrix improvement.

## Acceptance and provenance

The campaign completed all 2,900 planned datasets and 4,600 all-edge fits, with no failed, missing,
malformed or unknown results. An independent audit reconciled every flattened fit with its raw
likelihoods and recovery metrics, reproduced the null thresholds, all 20 controlled recovery
cells, all 21 paired stress contrasts and all four unpaired site-count contrasts.

Thirty-seven selected-edge fits saved an unsuccessful optimizer termination flag despite valid
scalar scores. Deterministic continuations found successful solutions for all 37 and retained
every selected edge. Frozen production scores remain the source of the summary. This numerical
sensitivity check does not certify a global optimum.

The JSON records hashes for the accepted aggregate, per-fit table, independent audit, final
production acceptance and methods. `scripts/build-myth-calibration-summary.py` exports these
bounded results from the local accepted research directory; it does not run experiments.
The site repository contains the public summary, not the private source working files.

## Scope

These results concern recovery under this specified model, Cosmic-derived tree, finite root panel,
93-trait primary matrix and missingness pattern. They do not validate an ancient geographical
origin or a historical diffusion process. They neither prove an alternative homeland nor
disprove an African origin. A selected branch is a split of sampled versions, not an ancient
population or a birthplace.