Research Program
Every track shares the same geometric architecture. The language track proved it works. Each subsequent track extends it into the domain where CPU-native geometric AI matters most, and feeds findings back into the core.
Track Overview
27 experiments. 5 papers. Validated the architecture and identified the physical-versus-cultural ceiling. Every other track feeds findings back here. It does not close.
44k–97k tokens/sec · 52 bytes · ceiling resolution ongoing
4 experiments defined and underway. Validates the architecture in the cleanest domain and provides a calibration reference for the language ceiling. Operations discovered here transfer back into language.
Target: Gram < 0.08 · 70% on M3 and M5 · function transfer
Open source Python library and reference implementation. A developer with no GPU and no cloud account builds a working semantic agent in under an hour on a standard laptop.
No GPU · no cloud · standard laptop · under one hour
Geometric reasoning for code structure. Syntax is rule-governed, scope has spatial analogs, control flow has directional properties. The structural conditions are the same as mathematics. Informed by math track findings.
Experiment design phase · depends on Track 2 results
Tests whether physically grounded axes are truly universal across languages. A geometric architecture built on physical universals is structurally better positioned for multilingual use than any statistical model trained primarily on English.
Arabic · Mandarin · Spanish · cross-lingual geometry validation
Geometric transparency makes bias visible and correctable without retraining. You can inspect where a word sits on any semantic axis and correct a biased position directly. The people most harmed by AI bias are the same people currently most excluded from the tools to audit it.
CPU-native · full vocabulary · no retraining required
Active and Upcoming
Foundation · Active
Track 1 · Language
27 experiments validated the geometric architecture and produced five papers. Open questions: the abstract axis ceiling confirmed at zero top-1 accuracy across six experiments, empirically derived decay rates, route library construction at scale, and dual state tracking for user and agent trajectories. Every finding from every other track lands here.
Active
Track 2 · Mathematics
Experiment 1 builds the axis space from 200 mathematical concept pairs targeting a Gram mean below 0.08 versus 0.123 in the language baseline. Experiment 4 tests whether geometric operations transfer back into the language space, with orthogonalization as the candidate fix for the moral/truth entanglement. Level 3 success: plain English problem in, correct mathematical answer out, on any CPU.
Proposed
Experiment 28
Validates trajectory velocity as an importance signal for episodic memory storage. The L2 distance between SSM states before and after each sentence detects semantically significant moments. Geometric traceback then recovers the full conversational context that produced the shift. Both mechanisms require no additional training and emerge directly from the geometric structure already validated in Experiments 1 through 27.
Planned
Experiment 29
Implements parallel SSM states for user and agent, validates that the agent state correctly tracks the AI's output trajectory, and measures the coupling between states across multi-turn conversations. A complete conversation fits in approximately 104 bytes of active state, with full geometric memory of both participants, retrievable in microseconds.
Planned
Track 3 · Deployment
All 27 experiments, the benchmark suite, the route discovery algorithm, and the full KAIA architecture released as an open source Python library. Documentation written for developers without ML research backgrounds. The equity argument becomes a deployable tool rather than a validated claim.