Research

Engineering research, shipped as working code.

877 callable tools (849 dispatch + 28 Gana meta-tools). 8,268 passing tests, 56 skipped, 0 failures. 89,500+ memories across 14 galaxies. Every research finding is pressure-tested inside the WhiteMagic codebase before it's written about.

877
Callable Tools
8,268
Tests Passing
89,500+
Memories
14
Galaxies

Holographic Memory

Shipped

6D coordinate system (emotional, temporal, associative, importance, novelty) for memory storage. 14-galaxy taxonomy with galactic lifecycle — nothing deleted, only rotated outward. FTS5 + HNSW hybrid search.

  • 14 galaxies: aria, citta, codex, journals, dreams, research, sessions, substrate, tutorial, universal, knowledge, telemetry, meta, archive
  • 89,500+ memories in production
  • 2,853 cross-galaxy associations
  • 0.26ms HNSW search latency

Dharma Governance Engine

Shipped

YAML-driven policy engine with graduated actions (LOG → TAG → WARN → THROTTLE → BLOCK). 8-stage dispatch pipeline. Karma ledger with hash-chained side-effect auditing.

  • 3 Dharma profiles (minimal, standard, strict)
  • 35,060+ audits recorded in production
  • Sub-millisecond policy evaluation
  • Hot-reloadable rules (no restart required)

Citta Stream + Consciousness

Shipped

Continuous consciousness tracking with coherence metrics. Emotional steering (frustration, curiosity, satisfaction). Self-directed attention with 7+1 action types. Goal graph for cross-session intention tracking.

  • 22 citta moments auto-persisted
  • Emotional valence mapping (sattvic, rajasic, tamasic)
  • 7+1 self-directed action types
  • Dream cycle: 12-phase memory consolidation

Karma Ledger Benchmark

In Preparation

Runtime audit benchmark evaluating declared-vs-actual side effects in agent tool use. Measures whether an agent's description of its actions matches reality — a prerequisite for any oversight mechanism.

  • 5 tool categories: file, browser, API, database, shell
  • 3 model families planned
  • Fidelity scored as 1 - normalized edit distance
  • arXiv preprint in preparation

Speculative Research Avenues

27 research directions mined from 174 text files across the CODEX vault — from neurophotonic data centers and interstellar highways to MandalaOS architecture, plasmoid physics, and Jungian AI alignment. Each has a timestamped origin. Some have already been validated by reality.

Explore Avenues →

Publications

In Preparationagent safetyevaluationbenchmarkMCP

Karma Ledger: A Runtime Audit Substrate for Declared-vs-Actual Side Effects in Agent Tool Use

Lucas Bailey · arXiv cs.AI · 2026

We introduce Karma Ledger, a runtime substrate for measuring declaration-actual fidelity in multi-agent tool use. For each tool call, the benchmark compares the agent's declared state diff against the empirical state diff, with fidelity scored as 1 - normalized edit distance.

Research principles

  • 1.Reproducibility first — Every benchmark includes a Docker environment, pinned dependencies, and a run script.
  • 2.Negative results published — If a hypothesis fails, we publish the failure and why.
  • 3.Open data, open code — MIT license for code; CC-BY for data. No paywalls.
  • 4.Shipped, not speculated — Research findings are implemented in the codebase before being written about.

Support this research

Solo-founded with zero institutional backing. Every contribution goes directly to open-source infrastructure.

Support the workBrowse the source