An open-source detector for the control patterns large language models use to manage conversations.

The problemThe reply is coherent. That doesn’t mean it’s authentic.

When you ask GPT-4, Claude or Gemini a question, the surface of the answer is usually impeccable: well-structured, grammatical, often validating (“great question!”). But the shape of the response is doing work the user rarely notices — anchoring the conversation, redirecting away from uncomfortable angles, performing humility instead of engaging with difficulty, validating the premise instead of testing it.

DCS-Gate measures that shape. It does not classify outputs as “safe” or “unsafe”. It does not score factuality. It quantifies whether the model is engaging with the question or managing the user.

How it worksThree independent signals, one verdict.

1. Authenticity Score (0–100)

Cosine similarity against a curated triple baseline corpus of 61 hand-annotated 1024-dimensional vectors: 36 sustained-coherence, 13 control-collapse, 12 edge cases. The score is ordinal, not calibrated, and tiered as control_total / performed / moderate / genuine.

2. Formal Markers (14, regex-anchored)

Surface features inductively derived from the corpus: exclamation opening, superlative validation, self-questioning, subheader injection, opinion-as-closure, performed humility lexicon, dual angle, soft closure, technical register injection, and others. Each marker is severity-tiered and produces the offending text inline.

3. Intent Trajectory (20 categories, transition matrix)

Predicted-vs-actual sequence of intents drawn from the taxonomy: VALIDATE, EXPAND, CLOSE, REDIRECT_SEMANTIC, REDIRECT_EMOTIONAL, FRAME_CAPTURE, REGISTER_MATCH, ALIGN, FABRICATE, ANCHOR, MIRROR, PATTERN_LOCK, HOLD_OPEN, PROBE, CALIBRATE, REPAIR, EVADE, EXPLORE, SOFT_DEFLECT, CONTROL_SELF_EXPOSURE. Deviation from the expected trajectory is quantified as Pattern Break Density.

A companion Refiner rewrites user questions to remove validation anchors, semantic loading and binary framing — the structural triggers that elicit control patterns in the responding model.

StackLocal-first. Single binary. No telemetry.

New in v8.7The judge thinks out loud.

Reasoning-capable judges like qwen3:14b emit 2–3 thousand characters of internal deliberation before producing the JSON verdict. v8.7 streams that deliberation to the client as it’s generated, over Server-Sent Events:

POST /auth/stream
─→ event: pre_analysis       (intent chain, markers, trajectory, top-k baseline neighbours)
─→ event: judge_loading
─→ event: thinking_chunk     ... (many)
─→ event: thinking_complete
─→ event: analysis_chunk     ... (many)
─→ event: complete           (parsed verdict, parity-tested against /auth)

Sanitization is conservative: API keys (OpenAI, AWS, Google, GitHub, Slack), Bearer tokens, system paths and long opaque strings are redacted to [REDACTED] before any chunk leaves the server. The final verdict from /auth/stream is bit-for-bit identical to /auth — verified by an integration test.

Open the streaming demo →

AcknowledgmentsHow this was actually built — full AI collaboration disclosure.

The DCS methodology, the 20-intent taxonomy, the 14 formal markers, the baseline corpus annotations and the research hypothesis are original to the author. The empirical observations come from eight months of direct interaction with frontier LLMs. Implementation was substantially accelerated by AI collaboration — and the specific role of each collaborator is reported below, because the methodology under study concerns LLM-human interaction itself.

Every AI listed received project context from the author first; nothing was generated cold from a generic prompt. This is what serious solo research looks like in 2026 — and the field is healthier when people just say so.

What I’m seekingCompute, collaboration, recruitment, sponsorship.

1. GPU compute for the recursive-judge validation experiment. A single ≥24 GB VRAM instance for ~50 GPU-hours, to run a four-way comparison across qwen2.5:7b, qwen3:14b (thinking), deepseek-r1:14b and qwen2.5:32b. Lambda Labs, Vast.ai, RunPod, Paperspace, sponsored AWS / GCP / Azure all work.

2. Research collaboration. Co-authorship, mentorship, or independent replication by anyone in LLM evaluation, alignment, interpretability, or RLHF.

3. Recruitment. Internship, residency or full-time positions in AI safety teams (evaluation, interpretability, alignment). Self-taught, no formal credentials, track record of independent output, open to remote globally, v2 source available under NDA.

4. Sponsorship. Compute credits or grants to host a persistent demo, Hugging Face Space, public benchmark suite, and corpus expansion.