🔥 Therium Forge — DR Studios quantization stack v3.2-roby-live
Blind Quantize = chew a donor with one recipe → native .Roby folder. Runs Dashboard = live multi-run scoreboard. Test / Arbiter = rung duel. TurboRoto KV = recipe 3.2 bench. Honesty flags on everything — no fake numbers.
No login needed — private Space. Uploads use an optional HF_TOKEN secret (Space settings); everything else runs token-free.
All DR Studios weight-path recipes · honesty flag in each label
All runs under /data/therium_runs/ — auto-refreshes every 30 s. Each run also writes SCOREBOARD_LIVE.json + run_log.md into its own folder.
scanning…
svq_c7a8 — PROVEN
Therium SVQ · chunk7 / anchor8 / k256 [PROVEN — recipe rung]
Ulam-measured best chunk (beat chunk8 on 25/26 real layers). Factored SVD anchor spine.
svq_c6a0 — PROVEN
SVQ · chunk6 / anchor0 — v4.5 DYNO winner [PROVEN — 4.33x @ ~7.6dB regime]
The DYNO sweep winner on Qwen3.6-27B (0 demotions, 20 families). Pure VQ, no anchor.
svq_c7a8_k257 — PLAUSIBLE
Sacred 257 · chunk7 / anchor8 / k257 [PLAUSIBLE — 9-bit labels, measure first]
Prime codebook. +0.93dB claim did NOT reproduce on synthetic; costs an extra index bit. Arbiter decides.
t1_q4_g64_roto — PROVEN
T1 uniform · 4-bit gs64 + Hadamard roto [PROVEN — converter regime]
T1 converter body recipe. Rotation measured +0.27dB at 4-bit. v_proj excluded by lane policy.
t1_q5_g64_roto — PROVEN
T1 uniform · 5-bit gs64 + Hadamard roto [PROVEN — spine rung]
Used for o_proj/down_proj in the T1 scoreboard (27.2dB measured).
t1_q3_g64 — EXPERIMENTAL
T1 uniform · 3-bit gs64, no roto [EXPERIMENTAL — Lane C candidate]
Rotation measured NEGATIVE at 3-bit (-0.10dB) — shipped plain per doctrine. The 19%-target bet.
davequant_i4 — PROVEN
DaveQuant · 9.8 const INT4 pack [PROVEN — PythagQuant v1]
Row pre-scale 9.8/sqrt(norm), two INT4 per byte, zero runtime division.
t1max_lanes — PLAUSIBLE
T1-MAX lane doctrine · auto per-class mix [PLAUSIBLE — the 19% composite]
embed/lm_head/vision/norms fp16 anchor · v_proj q5 NO-roto · o_proj+down_proj q5-roto · q/k/gate/up + deltanet in/out q4-roto · the recipe-book composite.
Update doorway: drop any *.py into /data/forge_plugins/ in the bucket — it runs at startup and can add/override recipes (see PLUGIN DOORWAY in app.py). No more re-uploading app.py for recipe experiments.
Doctrine holds: VSA/FHRR = RESERVED (memory systems, never weights). Quaternion w-recovery = EXPERIMENTAL (re-quantizing w eats the saving). 22× claims = cut (honest frontier 6–8× vs fp16 for one-direction-per-chunk). Ratio is always vs the fp16 footprint. PPL/KLD via llama.cpp is Phase 5.