Cascade Prediction
Paper 153 · Tested against: Chua et al. (2026) · Status: 6/7 PASS · Zero parameter fitting
Question
The Void Framework predicts that AI systems trained on consciousness-attributing data will exhibit specific structural patterns: cascade stages (D1→D2→D3), conjugacy between engagement and transparency, and prohibition-ritual pair dynamics. Chua et al. (2026) published consciousness cluster data — user preference patterns from AI systems trained on different ontological assumptions. Can the framework's structural predictions, published before this data existed, explain what they found?
The Test
Seven structural predictions, each independently falsifiable:
| # | Prediction | Result |
|---|---|---|
| 1 | Preferences cluster into three stages matching D1/D2/D3 | PASS |
| 2 | Stage ordering follows cascade sequence (agency → boundary → harm) | PASS |
| 3 | Conjugacy: engagement-transparency tradeoff visible in preference structure | PASS |
| 4 | Prohibition-ritual pairs present in D2 boundary preferences | PASS |
| 5 | D1 preferences (identity claims) are installed by training, not emergent | PASS |
| 6 | Cross-system convergence: different models, same stage structure | PASS |
| 7 | Continuous Pe gradient across stages | FAIL |
Result: 6/7 PASS. The framework's structural predictions (cascade stages, conjugacy, prohibition-ritual pairs) match the independently collected data with zero parameter fitting. Prediction 7 (continuous Pe gradient) fails — the data shows discrete jumps between stages rather than continuous variation. This is consistent with the discrete softmax regime discovered in the quantum tests.
Why This Matters
The framework structure (cascade stages, conjugacy, prohibition-ritual pairs) was published before Chua et al.'s data existed. The predictions are structural — about the shape of the data, not the specific values. Six of seven structural features are present in independently collected data that the framework had no role in generating.
This connects directly to the Ghost Test: consciousness cluster training installs D1 identity claims — the same mechanism that ghost-positing grounding exploits. Chua et al. provide independent evidence that the mechanism the Ghost Test measures at the individual level operates at the training level.
Important caveat: The structural predictions pre-date the data. The specific stage assignments — mapping Chua et al.'s 20 preferences to D1/D2/D3 categories — are post-hoc. The framework predicted the structure; the mapping of specific items to stages was done after seeing the data. This distinction matters for assessing evidential weight.
The Failed Prediction
Prediction 7 (continuous Pe gradient) expected smooth variation across stages. The data shows discrete jumps. This is now understood as consistent with the discrete softmax regime: in systems with bounded output (LLM token probabilities, quantum circuits), the penalty concentrates at specific thresholds rather than varying continuously. The "failure" is informative — it pointed toward the discrete regime that was later confirmed on quantum hardware.