Skip to content
View jigonyoo's full-sized avatar

Block or report jigonyoo

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
jigonyoo/README.md

Jigon Yoo

I build defect-detection tools and publish the catch rate rather than assert it. Every figure below sits in a public MIT repository beside the script that produced it.

What it catches Measured Repository
Side effects that execute twice 40 concurrent OS processes racing one refund — 20/20 deterministic detections of two reclaim-path defects that a statistical race test passed once-guard
Prompt injection, tested against itself 26/26 on the corpus written for it — 5/42 on the corpus written to break it llm-guardrails
Airflow DAGs, parsed instead of imported 12/12 planted defects caught, 0 false positives on the clean DAG dag-guard
Loads that exit zero and are still wrong a sabotaged batch fails 12 contract tests; the clean batch passes with 0 — catches a load that overstates revenue by 1,139% warehouse-quality-gate

The 5/42 is the number I lead with. A filter tested only against the corpus written for it has not been tested. Both corpora are published, so the bad number is reproducible in one command rather than quoted.

Documentation is a review surface too. I numbered all 158 behavioural claims in one README and tried to produce the command proving each: 7 were wrong, 18 had nothing behind them. I deleted the latency figures I could not reproduce and published the correction rather than dropping them quietly.

Written communication only — no calls required at any stage.

Korean · English · Chinese. RWS TrainAI language assessment: Korean 35/35, English 34/35, both Proficient User.

Gyeonggi-do, South Korea · jigonyoo.com · dev.to/jigonyoo

Pinned Loading

  1. dag-guard dag-guard Public

    Reads your Airflow DAG files and tells you what a green DAG in the UI will not. It parses the source with ast instead of importing it, catching 12 of 12 planted defects and 0 on the clean DAG.

    Python

  2. guardstack-docs guardstack-docs Public

    Documentation, limits and evidence for GuardStack — five safety gates for LLM agents on one hash-chained audit log. Leads with 5/42, the number that is bad.

  3. llm-guardrails llm-guardrails Public

    Input/output guardrails for LLM apps — and the 42-attack corpus that scores my own filter 5/42. MIT, zero deps.

    Python

  4. mcp-permission-server mcp-permission-server Public

    A permission layer in front of MCP tools, and a log that can prove why a call was allowed. Naive: 7 calls execute with no grant, 16 decisions cannot be reconstructed. Guarded: 0 and 0. Offline, zer…

    Python

  5. once-guard once-guard Public

    One-file idempotency guard for side effects that must not run twice. No daemon, no dependencies, and a demo that prices every failure.

    Python

  6. warehouse-quality-gate warehouse-quality-gate Public

    Stops a bad batch before it becomes a wrong number by running a dbt contract over every load, blocking a sabotaged batch on 12 failed tests while the clean batch passes with zero.

    Python