Research

Research
atlas.

A structured map of the papers, architectures, experiments, and running systems. Start with a program, scan the cards, then open any work for the public reader page.

Programs

4

Works

25

Live

2

1147 corpus items

deep scan across papers, docs, PDFs, systems

Personal arXiv

standalone HTML papers to open and save

PDF archive

server-hosted papers, figures, diagrams

Paper schema

source, PDF, math, evidence, checkpoints

Consolidation

canonical, merge, archive, reject protocol

Exploration

connected ideas before proof

Absorption log

daily papers judged against the corpus

9 draft PDFs

rendered artifacts linked from paper pages

Program 01

Language as Infrastructure

12 works

Program 02

Agents That Account for Themselves

7 works

Program 03

Embodied Trajectory Systems

5 works

Program 04

The Absorption Loop

1 works

working paper: 17preprint: 1experiment: 1architecture: 2whitepaper: 1system: 2practice: 1

Canonical shelves

The work, consolidated

The archive is large by design. These shelves are the current canonical map: the smallest set of research lanes that explain the expertise without making people scroll through every draft.

Shelf 01canonical

Low-Resource Speech and Script Infrastructure

N'Ko, Manding speech, script-native ASR, phonemic evaluation, acoustic governance

Writing systems are not passive outputs. They define the representational and measurement surface of low-resource speech systems.

Strongest body of work. Offline script-native ASR and paper set exist; stable live iPhone recognition remains the boundary.

Primary papers

N'Ko as Computational Infrastructureworking paperThe Script That Machines Can't Read: Adapting Large Language Models for N'KopreprintLiving Speech: Script-Native Automatic Speech Recognition for N'Koworking paperFeatural Acoustic Codingworking paper

Frontier queue

N'Ko as an Extensible Phonemic Substrate for Governed Low-Resource SpeechMAOE: Mixture of Anticipatory Orthogonal ExpertsFAC Read-Speech Evaluation Harness

Next consolidation

Freeze the canonical paper set into flagship, script invisibility, script-native ASR, phonemic evaluation, FAC, and governed deployment lanes.

Shelf 02canonical

Governed Agents and Provenance

Graph Kernel, admissible context, trajectory reward, typed skills, autonomous agent accounting

Agents become trustworthy when their work is recorded, scored, typed, and grounded in admissible evidence.

Multiple running systems and paper drafts exist. The remaining consolidation task is to align Graph Kernel, KARL, TML, and typed skills into one provenance program.

Primary papers

Deterministic Provenance Engines for Autonomous Agent Systemsworking paperKARL: Trajectory Reward EnginesystemTrajectory Memory Ledgerworking paperA Typed Algebra for Agent Skillssystem

Frontier queue

Graph-Augmented Recursive Language Models for Personal Knowledge SystemsGeometric Motifs for Selecting and Routing Coding-Agent Training DataLive Knowledge Graphs: Runtime Graph Integration for Continuous Domain Adaptation in Language Agents

Next consolidation

Turn the agent papers into one chain: admissible context, trajectory accounting, reward selection, typed composition, and self-improvement.

Shelf 03canonical

Embodied Trajectory Systems

Anticipation Geometry, computational choreography, multimodal sensor fusion, motion generation, Lume

Motion, conversation, and graph traversal can be treated as governed trajectories through state space.

Architecture and implementation evidence are substantial; release claims need tighter evaluation snapshots and physical run references.

Primary papers

Anticipation Geometry: Domain-General Trajectory Characterization with Knowledge Graph-Grounded Rewardsworking paperComputational Choreography: Deterministic Motion-to-Audio Synthesis via Geometric Anticipation Signalsworking paperRecursive Polymodal Synthesisworking paperCC-MotionGen: Audio-Conditioned Latent Motion Diffusion with Validation-Based Candidate Selectionworking paper

Frontier queue

TrajectoryOS and RAG++

Next consolidation

Separate the theory paper, the Lume production architecture, and the physical-capture evaluation papers so MotionMix evidence maps cleanly into the research.

Shelf 04canonical

Research Operations and Absorption

Daily paper ingestion, paper schemas, public research archive, experiment routing, canonical consolidation

Staying current should be an inspectable process that updates papers, experiments, and release gates.

The public archive, PDF archive, corpus, schema page, and absorption log exist. Automation health now needs to be visible and the backlog needs Codex-run experiment packets.

Absorption log

public verdicts

Paper schema

checkpoints

Next consolidation

Convert every ABSORB or TEST verdict into a named experiment packet mapped to one canonical shelf and one current internal baseline.

Complete index

All works at a glance

Compact scan view first. Open a card for the public reader page, manuscript note, or live system page.

Language as Infrastructureworking paperPDF

N'Ko as Computational Infrastructure

2026 / Read

Language as InfrastructurepreprintPDF

The Script That Machines Can't Read: Adapting Large Language Models for N'Ko

2026 / Read

Language as Infrastructureworking paper

N'Ko as an Extensible Phonemic Substrate for Governed Low-Resource Speech

2026 / Read

Language as Infrastructureworking paperPDF

Dead Circuits: Activation Profiling and Script Invisibility in Large Language Models

2026 / Read

Language as Infrastructureworking paper

Script Invisibility Is Structural: Activation Profiling Across Three LLM Families

2026 / Read

Language as Infrastructureworking paperPDF

Living Speech: Script-Native Automatic Speech Recognition for N'Ko

2026 / Read

Language as Infrastructureworking paperPDF

Does Script Design Matter? Phonetic Transparency and CTC Decoding for N'Ko

2026 / Read

Language as Infrastructureworking paperPDF

Beyond Controlled Comparison: Deployment Properties of Script-Aware ASR for N'Ko

2026 / Read

Language as Infrastructureworking paperPDF

Featural Acoustic Coding

2026 / Read

Language as Infrastructureexperiment

FAC Read-Speech Evaluation Harness

2026 / Read

Language as Infrastructurearchitecture

MAOE: Mixture of Anticipatory Orthogonal Experts

2026 / Read

Language as Infrastructurewhitepaper

N'Ko Compute Network

2026 / Read

Agents That Account for Themselvesworking paperPDF

Deterministic Provenance Engines for Autonomous Agent Systems

2026 / Read

Agents That Account for Themselvesworking paperPDF

Graph-Augmented Recursive Language Models for Personal Knowledge Systems

2026 / Read

Agents That Account for Themselvessystem

KARL: Trajectory Reward Engine

2026 / Read

Agents That Account for Themselvesworking paper

Trajectory Memory Ledger

2026 / Read

Agents That Account for Themselvesworking paper

Geometric Motifs for Selecting and Routing Coding-Agent Training Data

2026 / Read

Agents That Account for Themselvesworking paper

Live Knowledge Graphs: Runtime Graph Integration for Continuous Domain Adaptation in Language Agents

2026 / Read

Agents That Account for Themselvessystem

A Typed Algebra for Agent Skills

2026 / Read

Embodied Trajectory Systemsworking paper

Anticipation Geometry: Domain-General Trajectory Characterization with Knowledge Graph-Grounded Rewards

2026 / Read

Embodied Trajectory Systemsworking paper

Computational Choreography: Deterministic Motion-to-Audio Synthesis via Geometric Anticipation Signals

2026 / Read

Embodied Trajectory Systemsworking paper

Recursive Polymodal Synthesis

2026 / Read

Embodied Trajectory Systemsworking paper

CC-MotionGen: Audio-Conditioned Latent Motion Diffusion with Validation-Based Candidate Selection

2025 / Read

Embodied Trajectory Systemsarchitecture

TrajectoryOS and RAG++

2026 / Read

The Absorption Looppractice

The Daily Absorption Log

2026— / Read

working paper2026

N'Ko as Computational Infrastructure

The flagship consolidation: script-native speech recognition for Manding reaching a meaningful error regime, a phonemically interpretable error metric, and admissible tone correction framed as text prior times acoustic evidence. Establishes the measurement thesis: the script you evaluate in changes what your numbers mean.

Status: 31 pages, builds clean. Companion papers in preparation.

Read

preprint2026

The Script That Machines Can't Read: Adapting Large Language Models for N'Ko

Script visibility in large language models: what happens inside a model when the writing system it is asked to process was effectively absent from its training distribution, and what adaptation actually recovers.

Status: Submission-ready.

Read

working paper2026

N'Ko as an Extensible Phonemic Substrate for Governed Low-Resource Speech

Representation and governance paper. Shows how N'Ko can cover Manding, French, and English phoneme inventories through documented and compositional extensions, while proving that representation coverage alone does not make ASR self-correction safe.

Status: Drafted; strong representation result, correction loop still requires acoustic evidence.

Read

working paper2026

Dead Circuits: Activation Profiling and Script Invisibility in Large Language Models

Pillar paper. Profiles what activates, and what stays dark, inside a language model presented with a script it never meaningfully saw in training.

Status: Drafted; part of the flagship's companion set.

Read

working paper2026

Script Invisibility Is Structural: Activation Profiling Across Three LLM Families

Pillar paper. The invisibility result replicated across three model families, arguing the effect is structural to training distributions rather than an artifact of any one architecture.

Status: Drafted.

Read

working paper2026

Living Speech: Script-Native Automatic Speech Recognition for N'Ko

Pillar paper. The core ASR result: recognizing Manding speech directly in N'Ko rather than through a Latin-script intermediary.

Status: Drafted; anchor result consolidated into the flagship.

Read

working paper2026

Does Script Design Matter? Phonetic Transparency and CTC Decoding for N'Ko

Pillar paper. Measures the script advantage directly: a phonetically transparent orthography changes what a CTC decoder can learn, with a measured gap against less transparent alternatives.

Status: Drafted.

Read

working paper2026

Beyond Controlled Comparison: Deployment Properties of Script-Aware ASR for N'Ko

Pillar paper. What changes when script-aware recognition leaves the benchmark and meets deployment: latency, on-device constraints, and failure modes that controlled comparisons hide.

Status: Drafted; venue-split set prepared (four-paper release plan).

Read

working paper2026

Featural Acoustic Coding

N'Ko as a designed sound carrier. The script natively encodes four to five descriptor axes of the acoustic signal: tone to pitch, length to duration, seven vowels to spectral color, consonant manner to onset. Reframes tone restoration as an acoustic problem rather than a text-only one, with a measured token-cost advantage: one glyph where descriptor systems spend two words.

Status: Drafted with runnable experiment harness; first read-speech evaluation pending.

Read

experiment2026

FAC Read-Speech Evaluation Harness

Read-speech experiment lane for Featural Acoustic Coding. It separates feature coverage, prompt audio, human review, and acoustic evidence so FAC cannot silently become a transcript claim.

Status: Experiment scaffolded; not yet enough reviewed live labels for a correctness claim.

Read

architecture2026

MAOE: Mixture of Anticipatory Orthogonal Experts

An expert-routing architecture where experts are separated by authority rather than just specialization, with an admissibility gate that constrains what each expert may assert. Designed for the N'Ko recognition stack; the governance idea generalizes.

Status: Architecture materialized; empirics pending.

Read

whitepaper2026

N'Ko Compute Network

Proof-of-Linguistic-Competence: a protocol where a worker's fluency in N'Ko is itself the compute being bought, settled on a Bitcoin L2. Language competence as a verifiable, payable resource.

Status: Drafted; contracts deployed to testnet.

Read

working paper2026

Deterministic Provenance Engines for Autonomous Agent Systems

The Graph Kernel: a single Rust binary that produces cryptographically signed, policy-governed context windows, admissible evidence bundles, for agent reasoning. Evaluated across 27 queries against keyword, BM25, and vector-RAG baselines; perfect relevance on multi-hop structural queries at sub-300ms.

Status: Paper drafted; system running.

Read

working paper2026

Graph-Augmented Recursive Language Models for Personal Knowledge Systems

The cognitive twin line: distilling one person's full interaction history into a queryable model of how they think, with a companion theorems document covering the synthesis math. Your AI interaction history is a cognitive fingerprint.

Status: Paper plus theorems document drafted; SFT pipeline operational.

Read

system2026

KARL: Trajectory Reward Engine

Every agent trajectory on my infrastructure is scored at emit time by a multi-signal composite reward: process quality, outcome integrity, motion, and domain-calibrated baselines over thousands of recorded trajectories. The scores feed supervised fine-tuning data selection, closing the loop between doing work and learning from it.

Status: Running; thousands of trajectories scored, retraining loop active.

Read

working paper2026

Trajectory Memory Ledger

Schema-normalized experience replay for self-improving coding agents. Records complete tool-use sessions, scores them with a six-signal reward engine, and exports advantage-weighted training examples.

Status: Paper drafted from the KARL deployment corpus; public abstract safe, raw traces private.

Read

working paper2026

Geometric Motifs for Selecting and Routing Coding-Agent Training Data

Behavioral motif paper for coding agents. Compresses sessions into symbolic inscriptions and geometric features, then uses those annotations to route, select, and evaluate training data.

Status: Paper drafted; motif routing results should be kept tied to exact evaluation snapshots.

Read

working paper2026

Live Knowledge Graphs: Runtime Graph Integration for Continuous Domain Adaptation in Language Agents

A runtime alternative to treating knowledge graphs as static training scaffolds. The graph is queried live during inference through provenance-tracked context slicing, admissibility tokens, and production graph traversal rather than being discarded after fine-tuning.

Status: Paper drafted; cc-graph-kernel production service is the implementation substrate.

Read

system2026

A Typed Algebra for Agent Skills

Nearly three hundred operational skills typed under a six-category composition algebra: generators, transformers, reducers, distributors, effectors, auditors, with a linter that rejects ill-typed pipelines and a two-tier router that selects skills by type compatibility. Skill libraries become checkable, not just searchable.

Status: Running; 99.7% of the library typed, linter enforced.

Read

working paper2026

Anticipation Geometry: Domain-General Trajectory Characterization with Knowledge Graph-Grounded Rewards

Defines seven scalar signals, commitment, uncertainty, transition pressure, recovery margin, phase stiffness, novelty, and stability, over arbitrary state-space trajectories. Evaluated across motion, conversation, and knowledge-graph traversal as a shared geometry of convergence.

Status: Paper drafted; Rust implementation exists in Comp-Core; downstream task lift remains the hard proof gate.

Read

working paper2026

Computational Choreography: Deterministic Motion-to-Audio Synthesis via Geometric Anticipation Signals

Turns heterogeneous sensor streams into deterministic audio synthesis through anticipation packets. The core claim is that movement intent should pass through a geometric intermediary before becoming sound, rather than mapping raw axes directly to parameters.

Status: Paper drafted with verified build artifacts; physical sensor-data evaluation still pending.

Read

working paper2026

Recursive Polymodal Synthesis

A multi-modal sensor-fusion framework for computational choreography: kinematic, physiological, and rhythmic streams converge through proximal updates into a coherent embodied representation for real-time generative control.

Status: Research draft found in the documentation corpus; not yet curated for public release.

Read

working paper2025

CC-MotionGen: Audio-Conditioned Latent Motion Diffusion with Validation-Based Candidate Selection

A phrase-level diffusion system that generates latent motion from audio features, then rejects implausible samples and ranks surviving candidates by musicality. The important design move is validation-based candidate selection around a generative model.

Status: Implementation-grounded system paper; evaluation and leakage-safe dataset splits remain open.

Read

architecture2026

TrajectoryOS and RAG++

Personal trajectory modeled as a dynamical system with thrust, alignment, gravity, mass, and state-based retrieval over successful transitions. This is the personal-knowledge branch of the same trajectory thesis.

Status: Architecture and paper drafts exist; public release needs a privacy pass before details are exposed.

Read

practice2026—

The Daily Absorption Log

A budget-paced pipeline: fetch the day's curated papers, score them against ten research domains, deep-read what survives triage, and issue verdicts where every ABSORB names a target system, every TEST defines a head-to-head, and every RIVAL must cite evidence or be downgraded.

Status: Live. Updated as the loop runs.

Read