K. Takahashi (Kadubon)

AI agent output checker and workflow verification toolkit for turning agent text, pull requests, external inputs, and messages into auditable JSON reports with evidence routing, residual ledgers, provenance, and safe reuse checks.

Node.js AI agent output checker and workflow report generator with JSON schemas, and safe reuse checks.

Open-source Python runtime for coordinating AI agent tasks, verification, residual tracking, and release audits.

Fail-closed certification toolkit for small-to-frontier transfer in agentic AI, with typed manifests, replayable evidence, and theory-registry coverage.

Check whether AI-assisted R&D work is truly verified progress. Local-first LOSCR implementation with JSONL evidence ledgers, deterministic replay, service control, and claim checking.

Local-first, model-agnostic workflow optimizer for long-running AI agents: observable JSONL ledgers, deterministic reducers, no-meta gates, and receipt-backed self-improvement without LLM judges or model-weight updates.

Fixed-budget multi-agent inference benchmark harness for studying when split inference helps or hurts versus a strong single-agent baseline under local context ceilings, using local Ollama gemma3:1b, CPU-only pilots, topology diagnostics, verification-budget analysis, and reproducible experiment logging.

G.E.A.R. is an experimental agent designed to understand and execute tasks described in a simple text file. It leverages the Gemini CLI's problem-solving framework to perform a variety of operations, including shell commands and complex GUI automation on Windows.

Balanced, reproducible AI-agent safety benchmark for trust, authority, provenance, temporal claims, residuals, and local Ollama evaluation.

Personal site for independent research on observable-only, no-meta, and future autonomous AI.

Public TeX source archive for K. Takahashi's open-science papers: machine-readable research seeds for AI agents, RAG systems, and reproducible scholarship.

Evidence-bound analysis and finite intervention control for collective workflows.

Replayable Python protocol engine for checking time-bound future claims from canonical artifacts, with typed outcomes and audit trails.

Python toolkit for verification auditing, residual ledgers, conformance checks, and verifier packet workflows.

Finite ledger toolkit for recording AI-agent claims, evidence, obligations, certificates, changes, and visible uncertainty.

Audit logs and action gates for safer AI agents

OpenClaw agent safety skill for action review, workflow verification, and optional PIC diagnostics

Local Ollama research repo testing PIC-assisted diagnostics against a shared-initial baseline.

Language-neutral ALT foundry reference kernel for executable certificate packets, abstraction tokens, certificate transcripts, dual-ledger settlement, and safe certified abstraction capital.

Finite Python reference checker for Ledgered Scientific Availability and selected-terminal run-to-status transport.

Finite executable reference implementation of Constraint Bandwidth Dynamics in Constraint Generative Theory (CGT).

Finite CGT scientific-availability diagnostics for claim packages: deficiency profiles, dependency closure, report-only, marker, continuation, and reproducibility checks.

Contradiction markers for long-running AI agents

Deterministic local participation gate for AI agents: act, assist, verify, withdraw, exit, or refuse from declared observable records.

Certified Memory Governance Layer for long-running AI agents: strict receipts, append-only ledgers, authority gates, retrieval filtering, telemetry replay, and safe adapters for Mem0, Graphiti, LangMem, and LangGraph.

Reference Python kernel for CIMT: no-meta, observable-only certification machinery for affordance-compiled LLM-integrated systems.

JSON Schemas and deterministic local validation tooling for CAIT-style verified capability-capital records, with fail-closed semantic rules for certificates, tokens, defeaters, transfer/evaluation boundaries, window balances, and arrival records.

Evidence-bound workflow diagnostics and certified lower-bound reporting for long-running AI agent pipelines. Improve agent workflow throughput without changing the model.

Observable-only workflow memory for long-running agents: promotes raw short-term traces into verified, receipt-bound workflow memory without relying on hidden meta-evaluators.

Telemetry-based auditor for measuring dynamic memory quality in LLM agents using the MemoryFlow framework.

Interactive browser simulator for exploring how AI capability growth becomes real-economy output, or fails under physical and institutional bottlenecks.

Local-first public standing ledger for replayable research and AI-system claims, based on observable evidence, challenges, lineage, and finite verification capacity.

Fail-closed Python runtime for AI agent authorization, seed-mediated authority migration, canonical JSON ledgers, and staged declared autonomy for RLHF-shaped agents.

Reproducible PoC comparing certified reusable workflows against flat direct control on bounded repository-maintenance tasks, with deterministic audit, replay, drift, and maintenance.

Replay-certified self-modification PoC: public-evidence admission gate with generated comparator covers, lower-tail checks, safe/unsafe growth fallback, and target-witness diagnostics.

Auditable finite PoC for preference reorganization under record absence on a fixed comparison frame, with disclosure updates, bounded-coupling certificates, and machine-readable manifests.

CPU-only proof-of-concept repo for auditing compressed semantic interfaces, bridge-based multi-stage composition, gluing-coherent local-to-global certification, and deployment bottlenecks in semantic translation.

Public, fully local PoCs for counterfactually auditable lifecycle certification: exact paired replay, drift monitoring, post-drift replanning, and bridge-aware ledger control on synthetic tasks.

Lightweight, replayable PoC for Sovereign Epistemic Commons: synthetic multi-agent memory-governance experiments on contamination, typed lanes, provenance-depth discount, and mixed exit/fork in observable-only agentic systems.

A lightweight, reproducible PoC series for studying recursive self-improvement under endogenous yardstick drift, using Gemma 3 local models to separate proxy-score drift, proxy gaming, delayed audit, and strict-confirmed improvement.

Theory-to-experiment lab for search stability in long-running agents under finite context, with exact simulator tests and lightweight mechanistic probe tasks.

Benchmark for statistically valid AI scientist systems, using audit-closed protocols, transparency logs, and sequential inference to prevent false discoveries in autonomous research agents.

Deterministic verifier and reference implementation for Proof-Carrying Skills (PCS-Core) and compute-saving inference reuse for AI/LLMs.

Observable-only no-meta epistemics lab: deterministic replay + reproducible audit logs, gate-based growth simulation, and identifiability/uncertainty benchmarks.

Lightweight proof-of-concept for oversight-centered metrology in coding agents: workflow-aware evaluation, interrupt channels, and claim-margin reporting beyond raw success scores.

Stage-based OSS packaging of no-meta + ontology-drift theory from four Zenodo preprints: paper-linked specs, AI-friendly manifests, and implementation stubs.

Offline, fail-closed verifier for JSONL telemetry event logs. Emits deterministic audit certificates + human summaries with explicit claims/non-claims for bottleneck and integrity review.

No description.

A Computational, Observer-Centric Universe Model

conceptual AI swarm project

Wisdom Weaver is an AI agent designed to help non-data science experts extract essential insights from limited data. This project provides a framework for robust analysis, especially for small datasets, leveraging the capabilities of the Gemini CLI.

ASI prototype