Trust
Data sources and licences
This is a trust page. It lists every dataset, photograph, typeface and model that our demos and case studies use. For each one it says what we use it for, under which licence, where the original lies and how we credit it.
We checked each licence at its source between 26 and 28 October 2026, and we check again once a quarter and whenever a source changes. Millwright and Patrol use no external dataset: Millwright generates its orders and faults, and Patrol its terrain and contacts, both from a random seed. Do you see a mistake in a credit? Write to contact@tenzanlogic.com.
Data
Measurement and public data
| Source | Used for | Licence | Address | Credit |
|---|---|---|---|---|
| NREL PVDAQ, system 1204 | Helio, its case study, the Energy page and the sample report: inverter losses in MWh and PLN. | CC BY 4.0 (opens in a new tab) | doi.org/10.25984/1846021 (opens in a new tab) | Deline et al. (2021), PVDAQ, NREL/OEDI, doi:10.25984/1846021. The figures in the demo are our own calculations on these readings; NREL did not check or endorse them. |
| FHC, twin plants | The measurement of the fault detector in Helio: 4 strings, 25 days. | CC BY 4.0 (opens in a new tab) | doi.org/10.5281/zenodo.18979877 (opens in a new tab) | Bradl et al. (2026), Zenodo, doi:10.5281/zenodo.18979877. Change: readings every 30 seconds averaged to 5 minutes. |
| FLEURS, Polish test set | Voxmill: the word error measurement on 758 recordings and the 12-second transcript sample on the home page. | CC BY 4.0 (opens in a new tab) | huggingface.co/datasets/google/fleurs (opens in a new tab) | Conneau et al. (2022), FLEURS, Google, arXiv:2205.12446. Change: the sample is one recording encoded as mono, AAC 24 kb/s. |
| CelesTrak, satellite catalogue (GP) | AstrOS: the positions of about 16,700 real objects. | No separate licence; the usage policy applies (opens in a new tab) | celestrak.org (opens in a new tab) | Orbital data: CelesTrak. The catalogue is built into the demo when it is built (read on 16 October 2026). |
| KSeF Latarnia | KSeF Guard: the live KSeF status and the archive of failure announcements, in the user's browser. | No licence in the interface description; a public status feed (opens in a new tab) | github.com/CIRFMF/ksef-latarnia (opens in a new tab) | Source: KSeF Latarnia, the Ministry of Finance. The app never shows sample data as a status. |
Photographs
Photographs on this site
| Source | Used for | Licence | Address | Credit |
|---|---|---|---|---|
| Belén Montero, Pexels | Inside an old industrial hall. | Pexels License (opens in a new tab) | pexels.com (opens in a new tab) | Photo: Belén Montero, Pexels. |
| Do-yeon Kim, Pexels | A plant roof from above: a network of pipes and ducts. | Pexels License (opens in a new tab) | pexels.com (opens in a new tab) | Photo: Do-yeon Kim, Pexels. |
| Filipe França, Pexels | A gantry crane in a plant at sunset. | Pexels License (opens in a new tab) | pexels.com (opens in a new tab) | Photo: Filipe França, Pexels. |
| Brett Sayles, Pexels | An industrial plant beside railway tracks. | Pexels License (opens in a new tab) | pexels.com (opens in a new tab) | Photo: Brett Sayles, Pexels. |
| Szymon Shields, Pexels | A tall building seen from below. | Pexels License (opens in a new tab) | pexels.com (opens in a new tab) | Photo: Szymon Shields, Pexels. |
Typefaces
Typefaces
| Source | Used for | Licence | Address | Credit |
|---|---|---|---|---|
| Schibsted Grotesk | The typeface of the interface and the argument on this site. | SIL OFL 1.1 (opens in a new tab) | fontsource.org/fonts/schibsted-grotesk (opens in a new tab) | Schibsted Grotesk, SIL OFL 1.1 licence. |
| Literata | The typeface of the text the product produces or the buyer agrees to: the transcript sample and the legal texts. | SIL OFL 1.1 (opens in a new tab) | fontsource.org/fonts/literata (opens in a new tab) | Literata, SIL OFL 1.1 licence. |
| Geist, Geist Mono | The typeface of the TinyServe and Patrol demos. | SIL OFL 1.1 (opens in a new tab) | fontsource.org/fonts/geist (opens in a new tab) | Geist and Geist Mono, SIL OFL 1.1 licence. |
Models
Language models in the demos
| Source | Used for | Licence | Address | Credit |
|---|---|---|---|---|
| Llama 3.2 1B Instruct (Meta) | TinyServe: the model runs in the visitor's browser, which downloads the weights from huggingface.co. We do not host them. | Llama 3.2 Community License (opens in a new tab) | huggingface.co/onnx-community/Llama-3.2-1B-Instruct (opens in a new tab) | Built with Llama. Llama 3.2 is licensed under the Llama 3.2 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved. Acceptable Use Policy (Meta) (opens in a new tab) |
| LFM2.5 1.2B Thinking (Liquid AI) | TinyServe: as above. | LFM Open License v1.0 (opens in a new tab) | huggingface.co/LiquidAI/LFM2.5-1.2B-Thinking-ONNX (opens in a new tab) | LFM2.5 1.2B Thinking is licensed under the LFM Open License v1.0, by Liquid AI, Inc. The licence allows commercial use below 10,000,000 USD of annual revenue. Notice in the demo (opens in a new tab) |
| Qwen3 0.6B, Qwen3 1.7B (Qwen) | TinyServe: as above. | Apache License 2.0 (opens in a new tab) | huggingface.co/onnx-community/Qwen3-0.6B-ONNX (opens in a new tab) | Qwen3, Apache 2.0 licence, as the model card states. |
| Qwen3.5 0.8B (Qwen) | Patrol: an optional language model that summarises the narrator's reports in 1 to 3 sentences. The visitor downloads the weights from huggingface.co only after a click (489 MB), and the model runs on the graphics card of their computer. We do not host the weights. Nothing is sent anywhere. | Apache License 2.0 (opens in a new tab) | huggingface.co/onnx-community/Qwen3.5-0.8B-Text-ONNX (opens in a new tab) | Qwen3.5 0.8B, Qwen Team, Apache 2.0 licence as the base model card states; the ONNX weights card has no licence field of its own. ONNX weights q4f16 from onnx-community, revision 25cdf33f386b487a25f3d743e86442517bdb77ff. It runs through Transformers.js 4.3.1 (Apache-2.0) and ONNX Runtime Web 1.31.0-dev.20260914-8d85527a0 (MIT). Every answer carries the label “generated by a model, may be wrong” and makes no decisions. Base model card (opens in a new tab) |
| DeepSeek-R1-Distill-Qwen-1.5B (DeepSeek) | TinyServe: as above. | MIT (opens in a new tab) | huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B (opens in a new tab) | DeepSeek-R1-Distill-Qwen-1.5B, MIT licence, as the model card states. |
| distilbert-base-multilingual-cased-ner-hrl (Davlan) | Paste Shield: finding names, addresses and organisations in text, in the browser. We use the 8-bit ONNX conversion (Xenova), the weights unchanged. | Academic Free License 3.0 (AFL-3.0) (opens in a new tab) | huggingface.co/Davlan/distilbert-base-multilingual-cased-ner-hrl (opens in a new tab) | Model by Davlan, under the AFL-3.0 licence; ONNX conversion: Xenova. The model's author does not endorse Paste Shield. Notice in the demo (opens in a new tab) |
Own
3D models and parameter assumptions
| Source | Used for | Licence | Address | Credit |
|---|---|---|---|---|
| 3D models in Patrol | Patrol: the multirotor, the fixed wing, the vehicle, the mast and the figures. Every solid is built in the window's code at true scale, so there are no model files to download (0 B). | Our own, no third-party licence (opens in a new tab) | tenzanlogic.com/patrol/en/about/ (opens in a new tab) | Models: Tenzan Logic, not copied from real equipment. They are drawn by three.js 0.186.1 (MIT). |
| Asset class data cards | Patrol: the mass, the pack, the speeds, the accelerations, the orbit, the ceiling and the sensor ranges of the classes TL-Q4, TL-F12, TL-M20, TL-V4 and TL-SAT. | Our own assumptions, no outside licence (opens in a new tab) | tenzanlogic.com/patrol/en/about/ (opens in a new tab) | The class names are fictional. The parameters are assumptions within ranges from public literature: catalogues of civilian classes, scientific reviews and open documentation, with no manufacturer names. The exception is the satellite image delay (30 to 90 minutes), shorter than the 4 to 12 hours at commercial operators. The parameters show the differences between classes, not the behaviour of real equipment. The multirotor's energy use comes from the power model of Zeng, Xu and Zhang (2019). |