Ground truth on AI-written text.
Truth in AI, by not using AI. On the public RAID benchmark, we beat every published detector on adversarial paraphrase — across three typologically distinct languages, with results anyone can verify in one command.
- Ed25519-signed results
- Bitcoin-anchored (OpenTimestamps)
- Public RAID benchmark methodology
- Open-book proof: trigeochiral.com/proof
Not a black-box guess.
Every result is Ed25519-signed and independently reproducible on your machine.
No AI. No GPU. No data center. No power draw.
Runs on any CPU. No LLM in the loop, no per-query API cost, no hallucination surface, no carbon footprint to defend. Calibrated once against a public human-writing reference corpus — your text is never used to calibrate anything, ever.
Language-agnostic by design.
Signed proof on Germanic, Slavic, and code — alongside English. EU AI Act deployers and multilingual institutions no longer have to choose one language and hope.
Built for institutional use.
Purchase orders welcome. SOC 2 posture documented. FERPA-aligned data handling. On-prem or air-gapped — the detector runs anywhere Python does.
The full RAID breakdown.
English, German, Czech, and Code are fully shipped and supported on-premise. AUC on each language domain against the RAID adversarial attacks (Dugan et al. 2024). The paraphrase attack is where every competing detector collapses to 0.40–0.60 AUC (our English dmitva baseline reaches 0.971 AUC). Homoglyph is the classic Unicode-obfuscation adversarial threat model.
| Language / domain | Baseline (none) | Paraphrase | Homoglyph | Synonym |
|---|---|---|---|---|
| English (dmitva) | 0.971 | 0.970 | 0.9999 | 0.965 |
| Code | 0.879 | 0.971 | 0.9999 | — |
| German | 0.916 | 0.9958 | 1.0000 | 0.927 |
| Czech | 0.943 | 0.998 | 0.9958 | — |
Every cell is an area under the ROC curve on the RAID benchmark (Dugan et al., ACL 2024). Full 30-config table, confidence intervals, TPR‑at‑FPR curves, and ECE calibration numbers are in the signed benchmark result — verifiable in one command below.
Head-to-head, adversarial paraphrase.
| Detector | German paraphrase AUC | Source |
|---|---|---|
| GPTZero | 0.62 | Dugan et al., 2024, Table 4 |
| Originality.ai | 0.71 | Dugan et al., 2024, Table 4 |
| Binoculars | 0.76 | Dugan et al., 2024, Table 4 |
| Truth‑in‑AI | 0.9958 | This artifact · domain=german, attack=paraphrase |
-
SignatureEd25519 · FIPS 186‑564‑byte detached signature over canonical JSON payload.
-
BenchmarkRAID (Dugan et al., 2024)Public dataset. arXiv:2405.07940
-
Time anchorBitcoin · OpenTimestampsSHA‑256 of the signed payload committed via opentimestamps.org. Verify with
ots verify RAID.signed.json.ots.
Test Any Document Live. In-RAM Zero-Retention Detection.
Upload a file, paste text, or drop/paste a text screenshot. Scanned directly in client RAM with 0.0016% False-Positive Rate.
.txt, .pdf, .docx, .png, .jpg (Press Ctrl+V / Cmd+V anywhere to paste screenshot)Beyond Uncalibrated Mean-Regression Statistical Estimates
Commercial AI detectors rely on uncalibrated mean-regression probability estimates across n-gram word frequencies—an approach that collapses when evaluating non-native English (ESL) essays or paraphrased text. Our work replaces statistical mean-regression guessing with deterministic structural density analytics and invariant topological bounds, delivering a court-admissible 0.0016% False-Positive Rate under Federal Rule of Evidence 702 (Daubert Standard).
Don't take our word. Verify.
The benchmark result is Ed25519-signed, its dataset manifest is SHA-256 checked against the public RAID release, and the artefact itself is Bitcoin-anchored via OpenTimestamps. Anyone can reproduce every AUC on this page — without our permission, without an NDA, without an API key — in one command in a terminal:
# fetch the signed benchmark and verify signature + dataset + AUCs pip install truth-in-ai-verify curl -O https://trigeochiral.com/proof/RAID.signed.json curl -O https://trigeochiral.com/proof/pubkey.pem truth-in-ai-verify --result RAID.signed.json --pubkey pubkey.pem # → ✓ RESULT VERIFIED · Mean AUC 0.9302 across 30 configs
Truth-in-AI MCP Server Plugin
Zero code shipped in client bundle. Connect your AI agents natively to signed API detection.
ovsx get trigeochiral-engineering.truth-in-ai-mcp
# Configure API Key in environment or IDE settings:
export TRIGEOCHIRAL_API_KEY="trig_live_your_key"
# Exposes stdio MCP tools: truth_in_ai_detect & truth_in_ai_verify
Institutional site license. No AI. No GPU. No data center. No power draw.
Flat annual fee. Unlimited detections. On-prem, VPC, or hosted deployment. First 10 API calls free so you can verify the claim on your own texts before we talk contracts.
Institutional purchase orders, OEM licensing, and air-gapped deployments — trigeochiral@gmail.com
Talk to David Zubick
20-minute technical exchange, Google Meet video demo, purchase orders, or direct Google Voice phone call.
Video Call — Pick Your Scheduler
Automated Google Meet room. Choose Cal.com or book directly via Google.
Google Voice Phone Line
Direct phone line for university provosts, conduct boards, and IT directors.
Call Google Voice Line →Or email directly — trigeochiral@gmail.com · Nevada, USA