Novelty Engines
The novelty engines are the core differentiator. Each thinks from a completely different cognitive frame. They produce hypotheses that no human reviewer would generate, then score them for novelty and tractability.
Contrarian
"What if the field is wrong about one fundamental assumption?"
Reads the literature, extracts every "well-established" claim, inverts it across 6 axes: polarity, direction, scope, relevance, existence, priority. Outputs 10 counter-hypotheses ranked by impact if true × plausibility.
Cross-Pollinator
"What does a completely unrelated field know about this problem?"
Imports solutions from 15 unrelated fields: astrodynamics, epidemiology, portfolio theory, thermodynamics, linguistics, ecology, economics, neuroscience, materials science, music theory, urban planning, immunology, geology, philosophy of science. Extracts the mechanism, not the metaphor.
Assumption Excavator
"What unstated assumptions does every paper make?"
Extracts assumptions at 3 levels: explicit (stated), implicit (never stated but present), foundational (so basic the field doesn't state them). Builds assumption trees and identifies the critical 1-3 that would cause catastrophic failure if false.
Counterfactual Generator
"What if the key paper had never been published?"
Builds the field's history as a DAG, identifies branching points, rewrites history. Example: "What if the Transformer had failed on WMT 2014?" → No BERT, no GPT-1, LSTM research continues for years. Outputs 5 counterfactual histories with traced consequences.
Paradox Sifter
"What contradictions does everyone notice but nobody resolves?"
Extracts every "Limitations" and "Future Work" sentence from the entire literature. Cross-references to find contradictions, elephants in the room, and paradoxes classified as: direct contradiction, mutual ignorance, hidden dependency, or escalation.
Heretic (Crown Jewel)
"What if I read only the title and abstract, then generated 50 wild guesses?"
Reads ONLY the title and abstract. Generates 50 wild hypotheses in 5 categories (methodological, theoretical, empirical, foundational, wild cards). Then reads the full paper and scores each hypothesis. The Haunting Idea is the one the authors should have explored but didn't – novelty gap ≥ 8, tractability ≥ 5, impact ≥ 7, surprise ≥ 8.
Example: Heretic Engine Output
Input paper: “Attention Is All You Need” (Vaswani et al., 2017)
Haunting idea found: “What if the attention mechanism’s success is not due to attention at all, but due to the residual pathway allowing gradients to flow through 100+ layers? Any architecture with the same residual structure might perform equivalently.”
| Metric | Score |
|---|---|
| Novelty Gap | 9/10 |
| Surprise Factor | 9/10 |
| Tractability | 7/10 |
| Potential Impact | 8/10 |
| Overall | 92.6 |
Minimum experiment: Compare a Transformer against a version with attention replaced by learned linear mixing (no pairwise interactions), keeping the residual structure identical. If performance gap < 15%, the hypothesis is confirmed.