DocumentationEverything published about DCS-Gate.

Three groups: the project anchor (master research overview), the mathematical pipeline (matemáticas end-to-end), and the press kit (five copies of the project pitched to different audiences). All are part of the public repo — each is linked to the rendered HTML below and to the raw .md on GitHub.

Technical anchor

01 · 15 KB · ~7 min read
Master Research Overview
The anchor document. Biography, AI-collaboration disclosure, what is claimed as original, what is open, the validation hypothesis, the four asks (compute / collaboration / recruitment / sponsorship).
methodology research ask
07 · 34 KB · ~25 min read · Spanish
Matemáticas de DCS-Gate — pipeline end-to-end
Every numerical computation that occurs from receiving a model response to emitting the verdict. Embedding + L2 normalization, cosine similarity, baseline top-k, intent centroids, polos, formal markers, Pattern Break Density, cross-corpus textural analysis, authenticity score derivation. With formulas, pseudocode references to file.go:line, and worked examples.
math pipeline technical-deep

Press kit — the project, told 5 ways

Same project, tuned for five different audiences. Use whichever fits the venue you’re posting on.

02 · 1.5 K chars · ~1 min read
LinkedIn — short feed post
Hook + AI-collab disclosure + ask, written for LinkedIn’s feed preview (the first 2–3 lines decide whether anyone clicks “see more”). With posting tips.
linkedin short recruiter-friendly
03 · 6.5 K chars · ~6 min read
LinkedIn — long-form article
Top-to-bottom version for serious recruiters and researchers. Methodology, evidence, AI-collab disclosure, validation hypothesis, what’s being asked for.
linkedin long research-leaning
04 · 5.5 K chars · ~5 min read
Kaggle — notebook description
Technical, lower-personal. For use as a Kaggle notebook description or dataset card. ML-engineer audience. Includes the data-asset inventory and the “swap in your own judge” instructions.
kaggle technical reproducibility
05 · 7 K chars · ~6 min read
DEV.to / Hashnode — blog post
Narrative “how I built X” version for indie hackers, builders, devs. More casual register, same AI-collab disclosure.
dev-blog narrative builder-audience
06 · 5.5 K chars · ~5 min read
LessWrong / AI Alignment Forum — academic version
Hypothesis-first, tightly bounded claims, explicit falsifiability. No emojis, no hashtags. For posting to research-leaning AI-safety forums.
academic bounded-claims falsifiable

Repository structureWhat lives where in the source.

The public repo at Corekeeper-research/dcs-gate is laid out so the code, the docs, the press kit, and the reproducibility artifacts are each isolated: 

dcs-gate/
├── README.md
├── LICENSE
├── .gitignore
├── work/dcs-gate/          # v8.7 Go source (22 .go files, 73 tests)
│   ├── *.go
│   ├── data/               # baselines, markers, intent prototypes, poles, goldens
│   └── docker-compose.yml
├── docs/                   # this site (served by GitHub Pages)
│   ├── index.html
│   ├── demo.html
│   ├── docs.html
│   ├── pitch/              # press kit (.md sources + rendered .html)
│   ├── assets/
│   └── CNAME
└── reproducibility/
    └── qwen3_14b_smoke_test.ipynb   # auto-detects local Jupyter / Kaggle / Colab

ValidationWhat is and isn’t claimed.

DCS-Gate is not claimed to be validated as a research tool. It is claimed to:

  1. Produce a 0–100 ordinal authenticity score that, on the 61-entry corpus, separates curated examples by ~50 points (the smoke-test spread on 2× Tesla T4 with qwen3:14b: 30 / 20 / 72).
  2. Detect 14 surface markers correlating with response classes the author labels performed or control_total.
  3. Predict an intent trajectory across 20 categories with measurable Pattern Break Density.
  4. Run entirely on Ollama with no outbound calls, in <5 min from clone to first request.

It is not claimed to generalize beyond observed model families, to provide a calibrated authenticity probability, or to be robust to adversarial models. The recursive-judge hypothesis — the methodology’s central open claim — is the experiment the project is currently soliciting compute for.