docs: DotCheck Hub card + license (no weights)

CITATION.cff

@@ -1,6 +1,6 @@
 cff-version: 1.2.0
-message: "If you use DotCheck engines or cite measured gates, please cite as below."
-title: "DotCheck — Vermeer, Valla, and Muybridge AI-likeness engines"
+message: "If you cite DotCheck engines or measured holdout gates, use the citation below."
+title: "DotCheck closed-weight AI-likeness engines (inhouse@5, inhouse-text@9, inhouse-video@2)"
 authors:
   - name: "DotCheck"
 url: "https://dotcheck.ai"
@@ -8,14 +8,15 @@ repository-code: "https://huggingface.co/DotCheck"
 license: "other"
 date-released: "2026-07-26"
 abstract: >-
-  DotCheck scores AI-likeness for images (Vermeer), text (Valla), and video
-  frames (Muybridge) via closed in-house heads on frozen commercial-clean
-  backbones. Public Hub pages document engines and gate metrics; weights are
-  not redistributed. Product scoring system: Leviathan.
+  DotCheck serves binary AI-likeness scores for images (wire inhouse@5 /
+  Vermeer), text (inhouse-text@9 / Valla, seven languages), and video frame
+  bags (inhouse-video@2 / Muybridge) using closed heads on frozen commercial-
+  clean backbones. Public Hub repos document architecture and holdout metrics;
+  .npz head weights are not redistributed.
 keywords:
   - ai-detection
-  - image-forensics
-  - text-detection
+  - image-classification
+  - text-classification
   - video
   - siglip2
   - closed-weights
@@ -23,6 +24,6 @@ preferred-citation:
   type: soft
   authors:
     - name: "DotCheck"
-  title: "DotCheck engines (Vermeer, Valla, Muybridge)"
+  title: "DotCheck engines (inhouse@5, inhouse-text@9, inhouse-video@2)"
   url: "https://dotcheck.ai/docs"
   year: 2026

README.md

@@ -16,7 +16,6 @@ tags:
   - text-classification
   - ai-detection
   - multilingual
-  - roberta
   - closed-weights
   - dotcheck
   - valla
@@ -29,118 +28,130 @@ model-index:
     results:
       - task:
           type: text-classification
-          name: AI-likeness (text, English)
+          name: binary AI-likeness (text, en)
         dataset:
-          name: DotCheck text gates EN (human vs holdout AI)
+          name: DotCheck text holdout EN
           type: other
           split: holdout
         metrics:
-          - name: Human mean P(AI)
+          - name: mean_P_AI_human
             type: mean_score_human
             value: 0.045
-          - name: AI mean P(AI)
+          - name: mean_P_AI_ai
             type: mean_score_ai
             value: 0.909
-          - name: Balanced accuracy @ thr
+          - name: balanced_accuracy
             type: balanced_accuracy
             value: 0.980
         source:
-          name: DotCheck MODEL-CHOICE / Data.json
+          name: text_gates_v9 / Data.json
           url: https://dotcheck.ai/docs
 ---
 
-# Valla — DotCheck text engine (`inhouse-text@9`)
+# DotCheck/valla-text-v9
 
-**Valla** is DotCheck’s live **text** AI-likeness stack. It returns a probability ∈ [0, 1] (higher ≈ more AI-like) for a text segment, with an explicit `lang`.
+Closed-weight **text** AI-likeness stack for DotCheck serve. Hub repo = card + license only (no `.npz`).
 
-| | |
-|--|--|
-| **Product label** | Valla |
-| **Wire id (EN)** | `inhouse-text@9` |
-| **Public token** | Valla@9 |
-| **Live artifact (EN)** | `text_stack_head_v9.npz` |
-| **Bases** | [`Oxidane/tmr-ai-text-detector`](https://huggingface.co/Oxidane/tmr-ai-text-detector) (MIT) + [`fakespot-ai/roberta-base-ai-text-detection-v1`](https://huggingface.co/fakespot-ai/roberta-base-ai-text-detection-v1) (Apache-2.0) + DotCheck logistic head |
-| **Languages** | `en`, `es`, `pt`, `fr`, `it`, `de`, `nl` |
-| **Weights** | **Closed** — not published on this Hub repo |
-
-Part of **Leviathan** (engines + generator-surface prior + shared content-hash cache).
+| Field | Value |
+|-------|--------|
+| Hub id | `DotCheck/valla-text-v9` |
+| Wire id (EN) | `inhouse-text@9` |
+| Label | Valla |
+| EN artifact | `text_stack_head_v9.npz` |
+| Feature bases | [`Oxidane/tmr-ai-text-detector`](https://huggingface.co/Oxidane/tmr-ai-text-detector) (MIT), [`fakespot-ai/roberta-base-ai-text-detection-v1`](https://huggingface.co/fakespot-ai/roberta-base-ai-text-detection-v1) (Apache-2.0) |
+| Heads | EN + `es_v2` / `pt_BR_v1` / `fr_v2` / `it_v1` / `de_v1` / `nl_v1` (separate `.npz`) |
+| Output | `p ∈ [0,1]` — P(AI-like) |
+| Serve | CPU FastAPI `POST /v1/analyze-text` + required `lang` |
+| `lang` | `en\|es\|pt\|fr\|it\|de\|nl` (`pt` → `pt_BR` head) |
 
 ## Model description
 
-Two frozen text detectors (TMR + Fakespot) produce features; DotCheck logistic heads (one EN + six language heads) map to P(AI). Serve loads **one shared** base pair for all seven heads. Request language must be supported; unsupported languages fail closed (no silent fallback).
+Frozen TMR + Fakespot encoders → shared feature vector → language-specific logistic head. Serve loads **one** base pair for all seven heads. Unsupported `lang` → fail closed (`unsupported_language`; no silent EN fallback).
 
-This repository is **documentation only**. It does **not** contain `.npz` or checkpoint files.
+Text hygiene (TC1) applied before hash/score: strip wiki-style cites, collapse whitespace, drop `#category` tokens (mirrored in Express / extension).
 
-## Evaluation — English public claims
+**Files in this repo:** card + license artifacts only. Heads: private `DotCheck/heads-live-private`.
 
-Lockstep with `Data.json` / MODEL-CHOICE:
+## Architecture
 
-| Gate | Measured |
-|------|---------:|
-| Human mean P(AI) | **0.045** |
-| AI mean P(AI) | **0.909** |
-| Balanced accuracy @ threshold | **0.980** |
+```text
+raw text
+  → text_clean (TC1)
+  → TMR + Fakespot frozen forward (shared)
+  → concat / stack features
+  → lang-selected logistic head (npz)
+  → p_AI
+```
 
-Wiki-monitor mean is ~\<0.001 vs prior live (see MODEL-CHOICE). No live ChatGPT/Claude scrape holdout is published.
+## Inference
 
-## Evaluation — language heads (holdout gates)
+Do **not** `from_pretrained("DotCheck/valla-text-v9")` for DotCheck heads.
 
-Same absolute floors as EN (`humanMax` / `aiMin` / `balMin` in `Data.json`). Cite MODEL-CHOICE evidence JSONs:
+```bash
+curl -sS -X POST "https://dotcheck-server-c221c1f32c68.herokuapp.com/analyze-text" \
+  -H "Authorization: Bearer dc_YOUR_KEY" \
+  -H "Content-Type: application/json" \
+  -d "{\"text\":\"...\",\"lang\":\"en\"}"
+```
 
-| Lang | Engine | Human mean | AI mean | bal_acc |
-|------|--------|----------:|--------:|--------:|
-| es | `inhouse-text-es_v2` | **0.088** | **0.953** | **0.972** |
-| pt | `inhouse-text-pt_BR_v1` | **0.103** | **0.945** | **0.928** |
-| fr | `inhouse-text-fr_v2` | **0.082** | **0.976** | **0.975** |
-| it | `inhouse-text-it_v1` | **0.069** | **0.986** | **0.972** |
-| de | `inhouse-text-de_v1` | **0.064** | **0.968** | **0.958** |
-| nl | `inhouse-text-nl_v1` | **0.101** | **0.978** | **0.933** |
+UI: https://dotcheck.ai/check · API: https://dotcheck.ai/api · docs: https://dotcheck.ai/docs  
+Min length / FUP enforced at Express (product policy).
 
-Public marketing claims for text remain the **English** floats above unless a surface explicitly shows a language table.
+## Training data
 
-## Intended use
+| | EN | Other langs |
+|--|----|-------------|
+| Fit AI | Qwen2.5-7B + Mistral-7B (multi-temp) | open models, wiki-style prompts |
+| Holdout AI | **Qwen2.5-1.5B** | per-lang holdouts (~200/200) |
+| Human | hard-neg wiki / WikiText / Gutenberg mix | Wikipedia lead prose |
+| License | commercial-clean only (no NC banned sets) | same |
 
-- AI-likeness scoring for supported languages inside DotCheck product surfaces.
-- Citation of measured gates with attribution.
+No published holdout from live ChatGPT/Claude scrapes. Evidence: `text_gates_v9.json`, `text_gates_{lang}.json`.
 
-### Out of scope
+## Evaluation
 
-- `from_pretrained` on `DotCheck/valla-text-v9` for DotCheck head weights (not published).
-- Unsupported languages (no silent English fallback in product).
-- Authorship or plagiarism legal determinations.
+### English (public claim SSOT)
 
-## How to use
+| Metric | Target | Measured |
+|--------|--------|---------:|
+| mean P(AI) \| human | ≤ 0.12 | **0.045** |
+| mean P(AI) \| AI | ≥ 0.85 | **0.909** |
+| bal_acc @ thr | ≥ 0.90 | **0.980** |
 
-**Do not** call `from_pretrained` on DotCheck head repos.
+Also: beat prior `@6` holdout; wiki-monitor mean ~\<0.001; OOD / RAID-lite protocol OK (`TEXT_GATES_OK`).
 
-1. **Check:** [https://dotcheck.ai/check](https://dotcheck.ai/check) — Free/guest daily limits apply.
-2. **Pro API:** create a `dc_…` key in [Dashboard](https://dotcheck.ai/dashboard). Docs: [https://dotcheck.ai/api](https://dotcheck.ai/api)
+### Language heads (holdout; same absolute floors)
 
-```bash
-curl -sS -X POST "https://dotcheck-server-c221c1f32c68.herokuapp.com/analyze-text" \
-  -H "Authorization: Bearer dc_YOUR_KEY" \
-  -H "Content-Type: application/json" \
-  -d "{\"text\":\"Your paragraph here…\",\"lang\":\"en\"}"
-```
+| Lang | Wire | Human | AI | bal_acc |
+|------|------|------:|---:|--------:|
+| es | `inhouse-text-es_v2` | 0.088 | 0.953 | 0.972 |
+| pt | `inhouse-text-pt_BR_v1` | 0.103 | 0.945 | 0.928 |
+| fr | `inhouse-text-fr_v2` | 0.082 | 0.976 | 0.975 |
+| it | `inhouse-text-it_v1` | 0.069 | 0.986 | 0.972 |
+| de | `inhouse-text-de_v1` | 0.064 | 0.968 | 0.958 |
+| nl | `inhouse-text-nl_v1` | 0.101 | 0.978 | 0.933 |
 
-Keys are managed only in Dashboard (Pro). Confirm stays opt-in on Pro.
+## Intended use
 
-How: [https://dotcheck.ai/how](https://dotcheck.ai/how) · Docs: [https://dotcheck.ai/docs](https://dotcheck.ai/docs)
+- Supported-language AI-likeness scoring in DotCheck inference.
+- Citation of the tables above.
 
-## Training and evaluation data
+### Out of scope
 
-English fit AI: Qwen2.5-7B + Mistral-7B (multi-temp); holdout AI: Qwen2.5-1.5B. Humans: hard-negative wiki / WikiText / Gutenberg mix (see MODEL-CHOICE). Language heads: prose Wiki + wiki-style prompts; English absolute floors. Commercial-clean only.
+- Head weight download from this Hub id.
+- Languages outside `{en,es,pt,fr,it,de,nl}`.
+- Plagiarism / authorship adjudication.
 
 ## Limitations
 
-- Short / heavily edited / machine-translated text can be unstable.
-- Only seven languages are live.
-- Detector scores are not generator attribution.
+- Short, MT-heavy, or heavily edited text: higher variance.
+- Seven languages only.
+- Output is not generator attribution.
 
 ## License
 
-Custom **DotCheck Inference License (Closed Weights)** — [`LICENSE`](../LICENSE). Upstream TMR / Fakespot licenses in [`NOTICE`](../NOTICE).
+[`LICENSE`](LICENSE). Upstream TMR / Fakespot: [`NOTICE`](NOTICE).
 
 ## Citation
 
-[`CITATION.cff`](../CITATION.cff) · Docs [https://dotcheck.ai/docs](https://dotcheck.ai/docs) · token **Valla@9** / wire `inhouse-text@9`.
+[`CITATION.cff`](CITATION.cff) · wire `inhouse-text@9` / Valla@9 · https://dotcheck.ai/docs