Evidence

The reproducible SAE-1.4 validation package — read it here, or download the full code and data.

This is the artifact behind the validation status on the equation page: the reference implementation, the 5,000-trial structural harness, the synthetic-data generator, and the full empirical suite — including the openly-published null result. Everything is reproducible end to end with fixed seeds.

↓ Download SAE14_analysis (code + data, .zip)

Includes sae14.py, test_structural.py, gen_data.py, analyze.py, the synthetic dataset, and the canonical spec.

Validation summary

Condensed from Awareness Engineering v0.7. The failed sub-claim is shown openly — structural passing is not the same as real-human validation.

TestResult
Structural invariants (5,000 trials)all PASS
Out-of-sample fit (synthetic)R² ≈ 0.896 (components) vs ≈ 0.156 baseline; 0.584 fixed-form
Directionality (cross-lag)Blockers(t)→Awareness(t+1): β = −6.19, p ≈ 3e−11; reverse null
Falsification (shown openly)“Purity(t)→fewer future Blockers” did NOT hold (β ≈ 0, p ≈ 0.91)
Fairness (matched subgroups)high- vs low-constraint differ by only +2.70 / 100

README

SAE14_analysis/README.md
# SAE-1.4 Validation Package

A rigorous structural + empirical test of the Sanatan Awareness Equation v1.4,
run against honestly-generated synthetic HOPE-style logs.

## Files
- `sae14.py` — canonical SAE-1.4 reference implementation (all clamps, coherence gate, stage gate, overdrive metric, display transform)
- `test_structural.py` — 5000-trial randomized invariant property tests
- `gen_data.py` — synthetic data generator (200 users x 120 days = 24,000 rows)
- `analyze.py` — full empirical suite (model comparison, stratification, causality, fairness)
- `hope_synth.csv` — the generated dataset
- `results_output.txt` — captured run output

## Critical methodological choice
The ground-truth synthetic world uses a DIFFERENT functional form than SAE-1.4
(saturating tanh in purity, multiplicative blockers, smooth belief gate, diminishing
shakti returns, heavy-tailed additive grace). SAE-1.4's additive bracket is only an
APPROXIMATION of it. This avoids circularity: SAE-1.4 is not being tested against data
it generated. It must compete to recover an independent causal structure out-of-sample.

Crucially, the ground truth was built so that PHYSICAL CONSTRAINT does NOT reduce inner
awareness (only an expression channel), to test the spec's fairness requirement.

## Findings
1. STRUCTURAL: all invariants PASS over 5000 trials (monotonicity in blockers/purity/
   constraints, range bounds, stage gate).
2. PREDICTIVE: SAE-1.4 components model reaches R^2 = 0.896 out-of-sample (walk-forward,
   train day<80 / test day>=80) vs best baseline 0.156. SAE earns its keep on data it
   did not generate.
3. STRATIFICATION CONFIRMED: corr(Purity, Awareness) = +0.489 in positive-belief regime
   vs +0.403 pooled. Belief regime must be conditioned on, exactly as the spec warned.
4. DIRECTIONALITY: Blockers(t) -> Awareness(t+1) strong (beta=-6.19, p~3e-11); reverse
   null (p=0.93). Arrow points as the model claims.
5. FALSIFICATION (important): "Purity(t) reduces future Blockers(t+1)" did NOT hold
   (beta~0, p=0.91) — because that pathway was not built into the ground truth. This is
   the model being genuinely falsifiable. If real HOPE data shows the same null, that
   sub-claim must be dropped or rebuilt.
6. FAIRNESS: after matching on Purity+Belief, high- vs low-constraint users differ by
   only +2.70 on a 0-100 scale. Small residual leak (mental-constraint/blocker confound),
   within fair range, but flags the constraint split for careful design.

## Honest scope
This validates that SAE-1.4 is STRUCTURALLY COHERENT, FALSIFIABLE, and BEHAVES CORRECTLY,
and that it can recover non-trivial structure better than baselines. It does NOT validate
that SAE-1.4 describes real consciousness — the data is synthetic. Real validation
requires real (even anonymized) HOPE logs of >=60-120 days with a logged ObservedAwareness
proxy (e.g., a daily 0-100 Witness-access/clarity/steadiness rating).

## Next step with real data
Provide anonymized HOPE log columns matching the schema in `gen_data.py`. Then:
compute SAE outputs, fit weights under the spec's definitional constraints, run
walk-forward validation against baselines, rank blocker/purity dimensions by predictive
leverage, and characterize the empirical grace (epsilon) distribution.

Captured run output

SAE14_analysis/results_output.txt
======================================================================
B. MODEL COMPARISON  (walk-forward: train day<80, test day>=80)
======================================================================
model                                 MAE   R2(oos)
B0 constant                        11.932    -0.001
B1 purity-only                     10.129     0.156
B2 blockers-only                   11.841     0.010
B3 wellness(phys+mental)           11.865     0.012
SAE-1.4 raw composite               7.831     0.584
SAE-1.4 components                  4.191     0.896
SAE-1.4 + interactions              5.217     0.842

======================================================================
C. BELIEF-REGIME STRATIFICATION  (the key insight)
======================================================================
POOLED (all)        corr(Purity,Aw)=+0.403   corr(Blockers,Aw)=-0.108   n=24000
POSITIVE belief     corr(Purity,Aw)=+0.489   corr(Blockers,Aw)=-0.132   n=16473
NEGATIVE belief     corr(Purity,Aw)=+0.244   corr(Blockers,Aw)=-0.078   n=7527
Note: pooling positive+negative regimes can flip/wash signs — must stratify.

======================================================================
D. CROSS-LAG / DIRECTIONALITY  (does Blockers(t) lead Awareness(t+1)?)
======================================================================
Blockers(t) -> Awareness(t+1): beta=-6.189 (p=3.4e-11)
Awareness(t) -> Blockers(t+1): beta=-0.000 (p=9.3e-01)
Purity(t) -> Blockers(t+1):    beta=+0.000 (p=9.1e-01)  (expect negative: purity reduces future blockers)

======================================================================
E. FAIRNESS CHECK  (does high constraint unfairly imply low awareness?)
======================================================================
High-constraint users, matched mean Awareness: 23.00
Low-constraint users,  matched mean Awareness: 25.69
Difference (lo-hi) after matching on Purity+Belief: +2.70
Ground truth was built so inner awareness does NOT depend on physical constraint.
If diff is small, SAE measurement layer is fair; if large, constraints leak into awareness.

DONE.
Structural invariant tests over 5000 trials
  blockers_monotone     : PASS
  purity_monotone       : PASS
  constraint_monotone   : PASS
  stage_gate            : PASS
  ranges                : PASS
OVERALL: PASS