Yarizakura

@CrepuscularIRIS · User

GitHub profile ↗ · Compare

20 followers92 repositories

Repositories

CrepuscularIRIS/ResearchStudio-Web

ResearchStudio's Idea skills for web-search hosts — claude.ai, Claude Code and ChatGPT. Prompt-preserving port of microsoft/ResearchStudio: eleven reasoning prompts vendored byte-for-byte from a pinned commit and checked by verify_parity.py; retrieval is web search only, so no API keys, no connectors, no local corpus.

★ 0PythonForks 1

CrepuscularIRIS/crucible

基于 Proma + Prime Agent + Qwen/DashScope 的长程自动化科研平台:Research Skills/MCP 编排证据链、判伪实验、反馈迭代与可复现交付,一条 docker compose 启动。

★ 0TeXForks 1

CrepuscularIRIS/ugra-rgbd-robust

Missing Is Not Zero: well-posed modality removal for RGB-D segmentation robustness. A removal-equivalence criterion with an inference-time test, plus 293 SHA-256-verified result artifacts the paper's headline table recomputes from.

★ 0PythonForks 0

CrepuscularIRIS/research-os-engine

Deterministic, event-sourced research-orchestration engine — policy/mechanism split: K3/Opus policy oracles over an Orca CLI-process substrate; append-only event log, code-enforced gates, replay, orchestrator meta-eval.

★ 0TypeScriptForks 0

CrepuscularIRIS/diffusion-moe-research

DiffusionGemma-26B MoE research: verified wall-clock reasoning frontier + DSpark systems levers, governed by the research-os anti-Goodhart gate framework.

★ 0PythonForks 0

CrepuscularIRIS/cot-distillation-fidelity-audit

LLM-as-judge audit of critique-correction closure in 10 CoT reasoning-distillation datasets (440 traces): does distillation transmit the act of self-correction, or only its vocabulary?

★ 0PythonForks 0

CrepuscularIRIS/mingjing

明镜 MingJing — evidence-admissible competitive-analysis multi-agent runtime (LangGraph). LLM proposes, deterministic code adjudicates truth; honest weak→strong self-correction; every claim traceable to source.

★ 0PythonForks 0

CrepuscularIRIS/vllm

A high-throughput and memory-efficient inference and serving engine for LLMs

★ 0Forks 0

CrepuscularIRIS/VelesDB

VelesDB is a local‑first AI data engine written in Rust that unifies vectors, full‑text and graph in a single file with a familiar SQL‑like language. Instead of sending every RAG or semantic search query to a remote cluster, VelesDB runs directly on your server, laptop, browser, mobile or edge device — no cloud dependency, no external services, ..

★ 0Forks 0

CrepuscularIRIS/parser_avito

Avito Parser —бесплатный парсер для автоматического мониторинга новых объявлений Avito и\или выгрузки объявлений в файл

★ 0Forks 0

CrepuscularIRIS/stickler

A library for evaluating structured data and AI outputs with weighted field comparison and custom comparators

★ 0Forks 0