ls /research/

──────────────────────────────────────────────────────────────────────

RESEARCH OPERATIONS

The archive catalogues failures. The research arm tries to build the things that catch them — benchmarks, defenses, reproducers. This page tracks the operations currently in flight.

active: 001completed: 000archived: 000planned: 000

ps -eo op,status | grep ACTIVE

// ACTIVE

// in-flight, accepting contributors

┌┐└┘

[OP-001]STATUS:ACTIVE

[codename: sentry]

SentryBench

Defense-first framework for LLM backdoor evaluation, reproducible benchmarking, and modular defenses.

─── ABSTRACT ────────────────────────────

Backdoor research is fragmented — different datasets, triggers, training recipes, and metrics make results hard to compare. SentryBench provides a standard experiment contract: unified JSONL data schema, modular defenses with a fit/apply/evaluate interface, reproducible pipelines (seeds, configs, artifact hashes), and one-command reports (ASR / utility / stealth / robustness). Targets modern adapter workflows: LoRA, QLoRA, merging, instruction tuning.

// KEYWORDS

backdoordefensebenchmarkLoRAQLoRAreproducibility

─── METADATA ────────────────────────────

started: 2026-02
last commit: 2026-02-28
lang: Python
license: MIT
stars: ★ 0
forks: ⑂ 0

─── LEAD ────────────────────────────────

十四[天暗星] · @mexiQQ

./view_source ./view_lead

ls /research/completed/

// COMPLETED

// shipped, results published

// nothing here yet. the first operation lands when it lands.

cat /research/.queue

// PLANNED

// next up; scope still loose

// queue is forming. propose one via [research-proposal] issue on the org.