The line that matters most in our public reproduction script is not a p-value. It is the matrix multiplication that forces every estimable surface to face the same 9,999 person-level sign vectors.
Publisher: AETHERCORE. We build GenesisCore, a Western astrology and Jyotish analysis system, and publish the associated empirical research. This article inspects one narrow question in the current 420-person development result: how did the implementation control multiplicity across the declared evaluation package?
Table of contents
- Project and study recap
- What enters the script
- Why the family has 22 surfaces
- The four lines that generate maxT
- Adjusted p-values and simultaneous lower bounds
- What M3 handles separately
- Decoys, ablations, and failed history
- Reproduce and review the code
- Previous and next in this series
Project and study recap
GenesisCore converts linked astrological rules into testable data structures. The accompanying study asks whether frozen rule chains are associated with five documented social-expression functions: advocacy, authored symbolic production, collective realization, technology or product implementation, and competitive performance.
Version 7 uses 420 unique people in a development dataset. All five risk differences remained positive after adding birth era, evidence-measurement regime, and collection source. Earlier v3–v5 person-out tests remain non-positive. We therefore report v7 as a development-set result with measured-confound robustness, with an independently frozen person-out replication still open.
The multiplicity problem is concrete. The package contains several functions and more than one exposure surface per function. Reading five small p-values after searching freely across that structure would tell us very little. The code must preserve the declared family, preserve person-level dependence, and calculate the threshold from the maximum statistic in each draw.
What enters the script
The public matrix contains 420 anonymous rows. analysis_order fixes their order. Each row carries an opaque release_id, measured-factor groups, observed functions, and binary surface values.
The structure is approximately:
{
"release_id": "opaque-id",
"analysis_order": 1,
"birth_era": "...",
"source_evidence_regime": "...",
"collection_source_group": "...",
"observed_functions": ["..."],
"surfaces": {
"function-name": {
"pp_both_system": 0,
"any_tier_both_system": 1
}
}
}
The reproduction script sorts on analysis_order, asserts the exact 1–420 sequence, builds outcome arrays from observed_functions, and builds exposure arrays from surfaces. The order check matters because a person-level sign vector only has meaning when it is applied to the same fixed row order on every run.
Why the family has 22 surfaces
The five new outcomes are the primary reporting targets, while the multiplicity family spans 11 functions. Two frozen exposure surfaces are evaluated per function, giving 22 surfaces.
Eighteen surfaces are estimable. Four are structural zeros: their exposure column is constant, so a statistic cannot be estimated. The script records those four cases and asserts that exactly 18 influence vectors remain.
Dropping the four zeros silently and renaming the remaining 18 as the original family would make the family depend on what happened to be estimable. Keeping the 22-surface declaration visible makes that choice auditable.
For each estimable surface, the script calculates a null-score influence vector after residualizing against the M3 design. Stacking those vectors produces a 420-by-18 matrix.
The four lines that generate maxT
The implementation is short enough to inspect directly:
rng = np.random.default_rng(SEED)
signs = rng.integers(0, 2, size=(MAXT_DRAWS, 420), dtype=np.int8).astype(float) * 2.0 - 1.0
max_t = np.max(signs @ np.column_stack(influences), axis=1)
critical = float(np.quantile(max_t, 0.95))
signs has shape (9999, 420). One row is one draw; one column is one person. Multiplication by the 420-by-18 influence matrix gives an array of shape (9999, 18).
The key design choice is reuse. During a draw, a person's -1 or +1 sign is reused across all 18 estimable surfaces. If the same person contributes to several functions, that dependence is retained inside the draw. Each draw then contributes one value: the maximum statistic across the 18 columns.
The 95th percentile of those 9,999 maxima was approximately 2.59217.
Adjusted p-values and simultaneous lower bounds
For a primary surface, the script compares its observed t-statistic with the full max_t distribution:
adjusted_p = (1 + np.sum(max_t >= observed_t)) / (MAXT_DRAWS + 1)
lower = risk_difference - critical * hc3_standard_error
The five reported outcomes were:
| Function | M3 difference | maxT adjusted p | Simultaneous one-sided 95% lower bound |
|---|---|---|---|
| Advocacy / public cause | +33.60 pp | 0.0001 | +18.48 pp |
| Authored symbolic production | +39.08 pp | 0.0001 | +27.79 pp |
| Collective realization | +27.95 pp | 0.0001 | +16.01 pp |
| Technology / product | +32.49 pp | 0.0008 | +9.64 pp |
| Competitive performance | +32.14 pp | 0.0047 | +6.93 pp |
All five simultaneous lower bounds remained above zero. That is the selection-aware result we want reviewers to inspect. A single attractive raw p-value is not carrying the claim.
What M3 handles separately
The maxT layer controls multiplicity across the fixed family. It does not make historical records comparable by itself. The model sequence handles measured differences in the data:
- M0: no added blocks;
- M1: birth era, five levels;
- M2: evidence-measurement regime, three levels;
- M3: collection source, six levels.
The main table reports M3 risk differences. Additional checks include within-era and within-source permutations, raw strata summaries, leave-one-source-out fits, and separate opportunity-data adjustments for 311 and 109 people.
These checks address recorded factors. Unmeasured confounding remains possible, which is one reason the next external dataset matters.
Decoys, ablations, and failed history
The repository also contains 7,634 adversarial-decoy rows representing 708 unique exposure sets. They were constructed after outcome information was available. Their purpose is to attack specificity: does an observed surface exceed the 95th percentile of these necessary-condition alternatives?
They are not a random null distribution. The chance-control statement comes from maxT. Decoy performance answers a different question.
Western-only, Jyotish-only, and full-system exposures are also refit. Those ablations describe which rule family may carry a result under the present specification. No Western-by-Jyotish interaction test has been run, so the ablations do not establish synergy.
Finally, the non-positive v3–v5 person-out results remain in the repository. v7 adds robustness inside the development set; it does not turn the earlier external tests into successes.
Reproduce and review the code
git clone https://github.com/AETHER-CORE1219/Genesis-Core-route-research-public.git
cd Genesis-Core-route-research-public
python3 -m pip install -r independent_social_expression_v7/requirements.txt
PYTHONDONTWRITEBYTECODE=1 python3 independent_social_expression_v7/scripts/reproduce_v7.py
PYTHONDONTWRITEBYTECODE=1 python3 independent_social_expression_v7/scripts/validate_public_v7_privacy.py
(cd independent_social_expression_v7 && sha256sum -c CHECKSUMS.sha256)
A focused review can start with six questions:
- Does
analysis_orderlock the same opaque 420 rows? - Is the 18-column influence matrix derived from the declared 22-surface family correctly?
- Are person-level signs shared across every estimable surface in a draw?
- Is
axis=1taking the intended per-draw maximum? - Do the critical value, adjusted p-values, and lower bounds match
RESULTS.json? - Do the privacy validator and checksums pass?
AETHERCORE develops and operates GenesisCore, which provides free results, paid detailed analysis interfaces, and emailed PDF reports. The commercial relationship is disclosed here because it is relevant context; the research code and result files can be reviewed without using the service.
Previous and next in this series
The previous article introduced AETHERCORE, the GenesisCore product, the v1–v7 history, and the public package. This article followed the maxT implementation from the anonymous matrix to the simultaneous bounds. Next, we will inspect analysis_order, the privacy validator, and the checksum chain that make the package reproducible without publishing identities.
If you find a flaw in the fixed family, influence calculation, sign reuse, or adjusted-bound calculation, please include the code path and the result produced by the alternative specification.
Top comments (0)