loading the cybergraph into an LLM context window — selecting the most valuable pages to fit a token budget

the full graph with subgraphs is ~6.5 MB (~1.4M tokens). a 1M context window holds ~900K tokens. context packing selects which pages enter the window and which stay outside

method

each page receives a score derived from graph structure:

$$\text{score} = \gamma_{\text{eff}}^2 \times (1 + \delta) \times \log_2(s)$$

symbol meaning
$\gamma_{\text{eff}}$ effective gravity — inbound link count + reflected gravity from outbound targets
$\delta$ density — outbound wiki-links per KB
$s$ substance — content size in bytes

reflected gravity: if a page links to high-gravity pages, it inherits a fraction of their gravity. one step of diffusion with coefficient $\alpha = 0.05$. this ensures subgraphs pages that reference core concepts receive nonzero score even with zero inbound links

pages with a stake: field receive a 1.5× bonus

the packer sorts by score and greedily fills the token budget — largest score first, skip if exceeds remaining capacity

usage

# graph only, 900K token budget (default)
nu analizer/context.nu ~/git/cyber

# full graph + all subgraphs
nu analizer/context.nu ~/git/cyber --subgraphs

# custom budget
nu analizer/context.nu ~/git/cyber --subgraphs --budget 500

# show ranking table without packing
nu analizer/context.nu ~/git/cyber --subgraphs --stats

output goes to /tmp/cyber-context-{budget}k.md by default, or specify -o path

results on current graph

metric value
total pages scanned ~2700
pages packed (900K budget) ~1400 (50% coverage)
zero-gravity pages excluded ~450
subgraphs pages included proportional to their cross-graph gravity

connection to cyberank

context packing is a simplified cyberank. gravity is the degree centrality analog. reflected gravity is one iteration of the diffusion operator. the scoring formula approximates what the tri-kernel computes in full: which particles deserve the most focus

the key difference: cyberank runs to convergence on the full cybergraph in consensus. context packing runs one step offline as a build tool

see analizer/context.nu for implementation. see concat.nu for the simpler alternative that packs everything without ranking

Local Graph