The live rebuilds — one brief, two on-air runs
EP5 · 2026-10-17
Built live on air against the incumbent's own board. GLM-5.3-Flash on Hermes (UI/UX Pro Max + Anti-Slop) shipped a faithful rebuild in ~30 minutes at 55 tok/s — docked on originality for shadowing the #1 layout and inheriting its orange theme from agent memory. Nemotron 3.5 Lightning 30B-A3B on the Bionic harness took 10 minutes flat and paid for it: no vision check, light-gray text on white, a broken YouTube link. Both artifacts await their file drop before View-site lights.
harness: hermes pro max + bionicmodels: glm-5.3-flash · nemotron-3.5-lightninghw: 2× GB10 · RTX 5090 laptopepisode: ▶ EP5
Ep.5 ratifies the ladder — the crown holds
EP5 · 2026-10-17
The on-camera review settled every open verdict: the incumbent #1 stands, unanimously ratified at 8/10 aesthetic, 7/10 accuracy. The Ep.3 one-shots tie at 6.20 — GLM takes #2 on the faster clock; the pi and DeepSeek Ep.4 builds tie at #6. No win against the champion, so the ladder holds and the composite reorders the rest. The board's ranks now shift from the data alone. One hardware record corrected in the same pass: the crown build ran on 2× GB10, not 1×.
status: ratified on cameracrown: heldrows: 8 · ties: 2episode: ▶ EP5
The rebuild — GLM-5.3-Flash takes the crown
EP4 · 2026-10-04
The board's first turnkey rebuild: a new champion takes #1 and every prior entry shifts down a rung. Built end-to-end by GLM-5.3-Flash (EXL3, 4 bpw, TensorFold) under the Hermes Agent harness on the gx10-a6b9 rig — 2× NVIDIA GB10, 242 GB unified (hardware record corrected at the Ep.5 review; first shown as a single 121 GB machine). Direction committed before line one: the press-ledger you're reading. Zero external dependencies, six pages, offline-clean.
harness: hermesmodel: glm-5.3-flash-exl3-4bpw-tensorfoldhw: 2× GB10 · 242 GBstatus: crowned
The first prototype — qwen3.8-flash-next-nvfp4
EP1–3
Reigned since Ep.1: the live domain itself, built by qwen3.8-flash-next-nvfp4 (6B active, NVFP4) on a single DGX Spark. Ratification of its rubric rows was still pending when the rebuild landed; Retired: the human-in-the-loop prototype build was deleted from the live tree when the GLM-5.3-Flash rebuild took the domain — it never sat on the ratified ladder.
harness: hermesmodel: qwen3.8-flash-next-nvfp4hw: 1× DGX Spark (GB10)status: displaced
GLM 5.3 — the 20-minute one-shot on dual Sparks
EP3 · 26/09/2026
Same shared brief as the overnight build, finished in 00:20:00 across a 2× Asus GX10 cluster (256 GB combined) at 3-bit. Aesthetic 6/10 (prints 3/5), accuracy 2/10 (prints 1/5) as judged on camera — the accuracy column is where one-shots bleed.
time: 00:20:00aesthetic: 6/10accuracy: 2/10hw: 2× GX10episode: ▶ EP3
Qwen 3.8 27B — the deliberate overnight build
EP3 · 26/09/2026
The slow-motion control group: 02:13:00 on a Mac Studio M3 Ultra (96 GB), fp16. Same brief, same judges. Seven times the wall clock for the same printed scores — parameter efficiency did it no favours either.
time: 02:13:00aesthetic: 6/10accuracy: 2/10hw: Mac Studio M3 Ultra 96GBepisode: ▶ EP3
▶ opens the episode that judged a run on youtube.com/@localhost_live. Nothing here is scored until the verdicts are ratified.