Research transparency · SAE-1.4 archive and SAE-1.5
Results, failed tests and reproducibility
The project has two distinct evidence records: corrected synthetic regression results for the old SAE-1.4 forms, and boundedness checks for the new SAE-1.5 candidate. They must never be combined into a claim that SAE-1.5 predicts human awareness.
1. Corrected SAE-1.4 regression
The earlier constructed data used 200 hypothetical subjects across 120 days (24,000 synthetic subject-day records). Temporal holdout fit used days 0–79 for training and 80–119 for evaluation. These R² values belong to historical SAE-1.4 experiments, not SAE-1.5.
Temporal held-out results
| Model | R² | Interpretation |
|---|---|---|
| Purity-only baseline | 0.158 | Simple baseline |
| Persistence baseline | 0.153 | Simple baseline |
| Train-calibrated composite | 0.259 | Fitted variant |
| Train-fitted components | 0.308 | Fitted variant |
| SAE-1.4 fixed formula | -3.283 | Worse than held-out mean baseline |
R²=0.308 means a fitted-components variant explained roughly 30.8% of the synthetic test-set variance under that constructed design. The fixed displayed SAE-1.4 formula scored −3.283; it performed worse than predicting the constant training mean. Fitted variants are different from the fixed formula.
Holdout by unseen synthetic subject
| Model | R² | Interpretation |
|---|---|---|
| Calibrated composite | 0.220 | Fitted on training split |
| Fitted components | 0.264 | Fitted on training split |
| Fixed SAE-1.4 | -3.371 | Worse than held-out mean baseline |
The synthetic grouped test separated subjects entirely between training and evaluation; again, this is not an out-of-sample human study.
2. Correction and negative findings
Superseded R² = 0.896 and 0.584
Those earlier figures inadvertently used grace_mag, a generator-only hidden variable, as if a predictor could know it. They are invalid for prospective prediction and are retained only as a correction history.
The proposed temporal relationship that greater Purity(t) would predict fewer Blockers(t+1) did not hold in the corrected independent synthetic world: the reported 95% interval included zero. This is a meaningful negative finding, not a result to conceal.
See the historical test inventory
- Evidence scope: Synthetic data only (200 constructed subjects × 120 days). No human validation and no measurement of consciousness, Ātman, sākṣī, grace, or liberation.
- Temporal holdout — simple baselines: Purity-only R² 0.158; persistence R² 0.153 on held-out future days.
- Temporal holdout — train-fitted forms: Calibrated composite R² 0.259; fitted components R² 0.308. These fitted forms modestly outperform the simple baselines inside the synthetic world.
- Temporal holdout — fixed equation: The uncalibrated fixed SAE formula performs poorly: R² −3.283.
- Grouped-subject holdout: Calibrated composite R² 0.220; fitted components R² 0.264; fixed equation R² −3.371 on subjects excluded from training.
- Historical result — superseded: Earlier R² 0.896 and 0.584 figures used the generator-only grace_mag value as an oracle predictor. They are not valid predictive results and are retained only as a correction record.
- Synthetic falsification check: The proposed Purity(t) → fewer Blockers(t+1) relation remained null in the corrected independent world (95% CI includes zero).
3. SAE-1.5 — Structural checks, not regression proof
Latest CI mathematical cases
50,000
Seeded checks of the current Option B formula and bounded K dynamics
Bound violations in those checks
0
Checks of algebra, not evidence that the assumptions describe people
SAE-1.5 human predictive R²
Not measured
No human-labeled independent target or validated Awareness scale exists
Earlier one-trigger V04 and background-dynamics V05 experiments reported 100,000 randomized consistency checks each, before the final removal of the direct-B gate. They are historical structural exercises, not an independent replication of the current full model. The present CI source test rechecks boundedness and rate updates but does not calibrate coefficients.
The currently checked candidate
A = C × (1 − K_effective)
K_effective = K_base + (1 − K_base) × R
dK_base/dt = u(1−K_base) − v K_base
v = β D max(q,0) + H, H≥0
For the detailed event closure and rate definitions, open the SAE-1.5 equation specification.
4. Reproduce or challenge the results
The project repository is private, so external reviewers may not be able to open GitHub source links. Public reproducibility requires a separately reviewed public package and stable provenance; a GitHub link alone is insufficient. Do not claim a complete externally reproduced SAE-1.5 study.
5. Tests that could falsify the theory
- Test whether Belief predicts later independently assessed cultivated Purity, compared with practice-only and secular-commitment baselines.
- Measure P and K independently from the eventual outcome; never derive K from the A equation and use it to claim validation.
- Compare with stress, psychological flexibility, autoregression and basic attention/regulation baselines.
- Test null, negative and reversed associations rather than only scenarios designed for success.
- Specify and validate overlapping triggers, multi-domain Shakti and longitudinal changes in P, S and B before fitting a complete dynamic simulator.
No scientific inference from constructed cases alone establishes a spiritual mechanism. The seven-day personal diary remains separate and unstarted.