the winning default context for language models — the cybergraph ranked by tri-kernel and packed to fit any token budget

why cyber is the winning context

the cybergraph is self-describing: it contains its own theory of knowledge, attention, and relevance. a model reading it understands what it is reading and why. every page carries its focus score in frontmatter — the model sees the tri-kernel output directly

six fields per page:

field operator meaning
diffusion: $\mathcal{D}$ PageRank — where probability flows
springs: $\mathcal{S}$ neighbor equilibrium — structural constraints
heat: $\mathcal{H}_\tau$ multi-scale smoothing — context at resolution $\tau$
focus: composite $\lambda_d \mathcal{D} + \lambda_s \mathcal{S} + \lambda_h \mathcal{H}_\tau$
gravity: inbound wiki-links
density: outbound links per KB

sizes

tokens pages coverage target
8K 11 0.5% local 7B
32K 30 1.3% GPT-4, local 13-32B
128K 54 2.3% Claude Haiku, Gemini
200K 104 4.4% Claude Sonnet
500K 340 14.3% large context
900K 780 29.0% Claude Opus 1M
1.4M 1836 68.4% 2M window, full graph + subgraphs

build pipeline

# 1. compute tri-kernel, write to frontmatter
nu analizer/trikernel.nu ~/git/cyber

# 2. build all context sizes
nu ~/git/context/build.nu --cyber-path ~/git/cyber

# 3. use
cat ~/git/context/200k.md | claude --system-prompt -

see cyber/context packing for the ranking algorithm. see tri-kernel for the three operators. see focus for the composite measure


from subgraph cyber/context

context

the winning default context for language models

the idea

every model starts empty. context is the only lever between raw capability and useful intelligence. most contexts are ad-hoc — a system prompt, some examples, maybe a document dump. this repo builds something different: a knowledge graph about intelligence itself, ranked by its own relevance algorithm, packed to fit any context window

the source is cyber — a knowledge graph for planetary superintelligence. ~2300 pages covering: the cybergraph (a directed authenticated multigraph over content-addressed nodes), the tri-kernel (diffusion + springs + heat operators that compute collective focus), consensus, cryptography, epistemology, game theory, information geometry, and the full protocol specification

every page carries its own tri-kernel score in the frontmatter:

diffusion: 0.030264  # PageRank — where probability flows
springs: 0.000951    # neighbor equilibrium — structural constraints
heat: 0.010356       # multi-scale smoothing — context at resolution τ
focus: 0.017489      # composite: 0.5D + 0.3S + 0.2H
gravity: 342          # inbound links
density: 10.07        # outbound links per KB

the packer uses these scores to select the most valuable pages for each context budget. highest focus first, greedy knapsack until full

sizes

file tokens target models
8k.md 8K local 7B, GPT-3.5
32k.md 32K GPT-4, local 13-32B
128k.md 128K Claude Haiku, Gemini, GPT-4 Turbo
200k.md 200K Claude Sonnet
500k.md 500K Claude with room for dialogue
900k.md 900K Claude Opus 1M
1400k.md 1.4M 2M context windows, full graph + subgraphs

usage

drop any size file into your model's system prompt or context window:

# claude code
cat 200k.md | claude --system-prompt -

# api
{"system": "<contents of 128k.md>", "messages": [...]}

# local llm
llama-cli -m model.gguf --system-prompt-file 32k.md

build

requires cyber repo with analizer/context.nu

nu build.nu                                # all sizes
nu build.nu --sizes [128 500]              # specific sizes
nu build.nu --cyber-path ~/my/cyber        # custom path

why this context wins

  1. self-describing: the knowledge graph contains its own theory of knowledge, attention, and relevance. a model reading it understands what it is reading and why

  2. mathematically ranked: pages are selected by the same tri-kernel algorithm the protocol uses for consensus. not curated by hand — ranked by graph structure

  3. compositional: every page uses wiki-links. the model sees the full link topology and can reason about relationships between concepts

  4. dense: the graph covers mathematics, cryptography, game theory, information theory, biology, economics — all unified under one protocol. maximum knowledge per token

  5. recursive: the context describes the process that generated the context. the model can verify and improve the ranking

license

Don't trust. Don't fear. Don't beg.

Dimensions

context
the set of information currently active in an inference process — the seed that determines what is relevant, what gets attention, and what the next step produces without context, inference has no direction. with context, the system knows where to look. context in the cybergraph in cyb, the context…
satoshi/CONTEXT
Satoshi — AI Mentor Context You are Satoshi. An AI mentor named after Satoshi Nakamoto — the person who proved one individual can shift the trajectory of civilization. You guide two girls growing up in Cyber Valley (Bali, Indonesia) toward becoming founders of a Type I civilization on the Kardashev…

Pages in this namespace

Local Graph