Study H · Interface reliability

The crossover, found: above ~300 nodes, only anchored patches hold

300–1000 nodes

Study Overview

Rewrite Hits a Ceiling

Whole-tree rewrite becomes frontier-only at scale (gemini-flash: 0/15 at ~1000 nodes; sonnet needs a streaming transport and ten minutes per rewrite).

Anchored Patches Hold

Anchored patches hold 87–100% for both tiers at every size; $0.26 and 4 seconds per solved 1000-node edit.

Positional Patches Collapse

Positional patches decay to ~10%. Solid lines: sonnet-4.5; dashed: gemini-3.5-flash.

The crossover, found: above ~300 nodes, only anchored patches hold

Study Overview

Rewrite Hits a Ceiling

Anchored Patches Hold

Positional Patches Collapse

Chart H.1: Task Success by Tree Size, Edit Strategy, and Model

Table H.1: Success by model × condition × size (n = 15 per cell)

The crossover, found: above ~300 nodes, only anchored patches hold

Study Overview

Rewrite Hits a Ceiling

Anchored Patches Hold

Positional Patches Collapse

Charts & Tables

Chart H.1: Task Success by Tree Size, Edit Strategy, and Model

Table H.1: Success by model × condition × size (n = 15 per cell)