box 242a…d21b · 2× RTX 5060 Ti + 50 GB RAM · generated 2026-10-16 14:16 UTC · spec: SOVEVAL-SPEC-v1 · law: BOX-MODE-LAW-v1
current mode
THINKING
thinker: qwen35-122b — loaded; GPU0=4129MiB GPU1=3520MiB; lineup: bolt | qwen35-122b
thinker lineup (swappable)
bolt · qwen35-122b · glm-elite*
*downloading · switch: thinking_mode.sh on <model> / off
slot matrix
212 matrix · 100 HIER · 16 thinkers · 8 volume
336 computed slots, fit-checked · receipt: BOX-SLOT-MATRIX-v1
| model | size | disposition |
|---|
| qwen3.5-122B:q3xl | 56 GB | 72.7% (real, retake 2026-10-16) — first pass 0.0% was artifact | PROVEN model — think split verified 841 chars, '391' exact |
| nemotron-3.5-bolt:74b-iq2xxs | 22 GB | 90.9% gate ✅ | |
| muse-glimmer:30b-q3xl | 13 GB | 90.9% gate ✅ (retake after budget 3072) | |
| qwen3-next:80b | 50 GB | 100% (legacy graders) | |
| qwen3.5-35b-a3b:iq4_xs | 17 GB | 90.9% legacy | |
| ornith:9b | 5.6 GB | 90.9% legacy + D1 reproduced | |
| kat-coder-v2.5-dev:q4_K_M | 21 GB | 90.9% legacy | |
| qwen3-coder-next:latest | 51 GB | 90.9% legacy | |
| gpt-oss:20b | 13 GB | 90.9% legacy | |
| ornith-1.5:35b | 22 GB | 90.9% legacy | |
| qwen3.8-27b:q3xl | 13 GB | 90.9% legacy | |
| gemma4:12b | 7.6 GB | 72.7% FAIL (gauntlet C1, official settings temp1.0/64/0.95) | out of band at recommended config |
| mixtral:8x7b-instruct-v0.1-q4_K_M | 28 GB | 27.3% FAIL (gauntlet C2, ctx 4096) | out of band; ctx cap 4096 |
| devstral-small-2:24b | 15 GB | 63.6% FAIL (gauntlet C3, first-ever, ctx 8192 cap cures the old 500) | out of band; serve ctx <= 8192 |
| phi-4-mini-reasoning:latest | 2.5 GB | — | TOOL TRACK — do-not-bench on chat grader (operator 2026-10-16) |
| agents-a1:q4_K_M | 21 GB | — | TOOL TRACK — same as phi-4 |
| mxbai-embed-large:latest | 669 MB | — | embedding — not chat |
{"PASS": 10, "FAIL": 16, "NOT-INSTALLED": 7, "NOT-BUILT": 13, "BLOCKED": 53, "CONFLICT": 1, "UNVERIFIED": 3, "PENDING": 1}
HIER formation
GO
+1.8% mean (GO threshold 15%) · 0 errors · 30 orchestrator queries during swarm
D1 consensus
voting wins 4/6
D1: voting wins 5 tasks, never loses · 0 errors — reproduced lane A
seed probe
GO
ollama honors options.seed at temp 0.7 — consensus measurable
tool-caller
Bolt ✅
valid OpenAI tool_calls · Lightning not called