Credits

Sybil Local AI registry

The idea of a recipe — one exact launch of model X on GPU Y, with proof it worked — comes from the Sybil Local AI registry by 0xSero (sybil-solutions/local-ai-registry). That project is MIT-licensed, © 2026 0xSero.

We did not import their AMD card files or their recipe proofs on this ship. We re-implemented the idea in our own words and Astro pages: evidence tiers, a pass/fail suite, a static JSON API, withdrawals, and a contribute path. Their site’s prose and UI code stay theirs.

MIT License

Copyright (c) 2026 0xSero

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

Full license text lives upstream. If we later ingest their AMD Q4+ rows, each cell will say “Cited external” and link the recipe file. That import is a separate task.

local desktop / our test rig

Official test-run cards on /results are credited to AMDai (verified, local desktop). Schema amdai-bench/run@1. Official 2026-10-13 set: 35B UD-Q5_K_XL 102.39 / 90.52 / 66.43 / 44.4, Q5_K_M 109.73, Q4_K_M 117.69, Q6_K 103.23 at 4K, dual split 82.78 at 4K / 42.03 at 256K, dual-instance ~100 / 172 / ~177. Flash / Strata / Q3 are pending a DDR5 rerun. Label: Measured on our rig, 2026-10. A measured Qwen 3.8 Flash Q3 run is an approved one-time exception — it will show as measured when the jsonl line lands.

Shannon / @iamMrDuncan

The community lab write-up on /labs/shannon is sourced from Shannon (@iamMrDuncan), X Article How I run my Lab (2026-10-12). Every claim there is labeled Cited, not measured by AMDai. We did not invent tok/s, prices, or a Thermite repo URL. Herdr and Coolify appear in /harnesses because of that write-up. The P100 Qwen 3.8 27B ~33 tok/s with MTP figure is NVIDIA cited context only — Oct 5 post.