frontiers / frontier
The dependency graph over this frontier’s findings. Each edge is a typed link: one finding supporting, contradicting, or refining another. The graph is a reviewable view over accepted finding bundles, not a separate authority layer.
Finding graph
Graph index contract
The graph view makes dependency structure visible while keeping accepted finding bundles as the durable objects.
derived view
Supports, contradicts, refines, and depends-on links help users navigate a frontier's review structure.
record state
The graph points back to finding bundles. It should not replace source spans, provenance, or review history.
review path
Inferred or candidate links should route to review before they affect proof, trails, atlases, or constellations.
impact
A changed source or finding can expose affected proof, dependent findings, and review radius.
Attention head stability findings (index 1) directly ground the prediction that residual-level circuits provide more robust discovery ground (index 5); cannot claim residual universality without first resolving head instability. · inferred by reviewer
SAE monosemanticity claim (index 3: 68% improvement) appears to contradict persistent polysemanticity prediction (index 8: ~10% remain), suggesting empirical evidence of only partial progress toward full disentanglement. · inferred by reviewer
Mechanistic similarity across architectures (index 4) extends and complicates universality claims (index 1) by showing that architectural differences (off-by-one patterns) persist despite functional similarity; true universality requires accounting for implementation variance. · inferred by reviewer
Hybrid circuit discovery speedup (index 4) supports weight-sparse transformer circuits prediction (index 9) by demonstrating that faster discovery is achievable when circuit redundancy is reduced; suggests redundancy level determines discovery cost. · inferred by reviewer
Residual stream stability (index 5) confirms the induction head scaling pattern (index 0) at a higher level of abstraction; suggests that phase transitions in in-context learning rely on stable residual-level representations rather than individual head circuits. · inferred by reviewer
Circuit-Aware Reward Training claim (index 7) depends on scaling Contextual Decomposition to attribution of sub-circuit interactions (index 10); causality validation cannot proceed without sub-circuit level precision. · inferred by reviewer