LinkedIn — Short Feed Post (~1500 chars)

Copia desde aquí abajo (sin el título) y pega en LinkedIn como post normal. El preview que verán los recruiters son las primeras 2-3 líneas, por eso son el “hook”.

I spent 8 months observing how LLMs subtly manage users instead of answering them — projected validation, performed humility, frame capture, register match. I built a tool that detects these patterns and rewrites questions to neutralize them.

It’s called DCS-Gate (Dynamic Coherence State Authenticator). A live v1 prototype is at 👉 https://dcs-auth.codewords.run — paste any question + LLM response and you’ll see a 0–100 authenticity score, the formal markers detected, the predicted intent trajectory, and a refined version of your question.

The v2 stack is a 3,000-LOC Go binary, 73 tests, triple baseline corpus of 61 hand-annotated vectors, 14 formal markers, 20 intent categories. Local-first, Ollama-only, no external APIs. Since v8.7 the judge’s reasoning trace streams to the client over Server-Sent Events — you watch the deliberation as it’s produced, you don’t wait in silence.

Honest disclosure of who actually helped: • Cody (CodeWords AI) — co-creator of v1 after I pushed back against its own control patterns in a long conversation; v1 wouldn’t exist without that exchange. • GitLab Duo — I walked it through the project’s full logic; from that emerged the v2 roadmap. • Meta AI — generic at first; technical depth amplifier once it had context. • Replit AI — brutally honest, justified contundent code failures, then strengthened architecture. • Z.AI (Zhipu GLM) — caught code bugs that slipped through. • Devin AI (Cognition) — executed the v2: backend (3k LOC, 73 tests, SSE streaming layer with conservative sanitizer), frontend, deployment, notebooks, docs.

The methodology and corpus are mine. Every AI received project context from me first — nothing was generated cold. This is what solo research looks like in 2026.

I’m looking for: 🔬 GPU compute (≥24 GB VRAM, ~50 hours) for the full four-judge comparison (qwen3:14b confirmed on 2× T4; qwen2.5:32b is the one that needs the bigger single card) 🤝 Research collaboration on LLM evaluation / alignment / interpretability 💼 Internship, residency, or full-time roles in AI safety 📢 Sponsors for the open-source release

If any of this resonates — message me. I’ll send the v2 source under NDA if useful.

AISafety #LLMEvaluation #AIAlignment #OpenSource #Research #ResearchCollaboration #MachineLearning

TIPS PARA POSTEAR: - LinkedIn corta el post a 2-3 líneas + “…ver más”. Las 2 primeras líneas (“I spent 8 months observing…”) son el hook. - Pon la URL de v1 en el cuerpo, no como link card al final (genera más clics). - Los emojis 🔬🤝💼📢 ayudan al scan visual sin caer en spammy. - Los 7 hashtags al final son los óptimos según el algoritmo (5-9 tags). - Postea entre 8-10 AM hora local martes/miércoles/jueves para máximo alcance.