Apodex 1.1 mini Lab Record

Identity

Field	Value
Model	Apodex 1.1 mini
Producer	Apodex; base: Qwen/Qwen3.5-35B-A3B; ~36B total / A3B active MoE
Evaluated artifact	Community conversion abenzerps/Apodex-1.1-mini-GGUF @ 59afa57852525f79a9634d5f80dd639cceee572c (not official)
File	Apodex-1.1-mini-IQ1_M.gguf
SHA-256	1e84d8adf7837e96fb18712882a8a114becc7e53554372e2f612c5e0c6276cd4
Evaluation date	2026-08-25
Protocol	WELP end-to-end (frozen snapshot welp-next-snapshot-2026-08-25-end-to-end)
Highest phase reached	Phase 3 (gate decision)
Campaign classification	BOUNDED_EARLY_STOP / PHASE_3_DO_NOT_ADVANCE

Hardware

Field	Value
Machine	WumboJetsII
GPU	NVIDIA GeForce RTX 5070 12GB (SM120)
VRAM	12,227 MiB physical
CPU	AMD Ryzen 7 9800X3D
OS	Fedora Linux 44
Driver / CUDA	NVIDIA 610.57.04
Runtime	llama.cpp pinned at f280b26983ad0fdb705a0d9ebf0503e76f2899b0 (2026-08-24)

Headline Verdict

DO_NOT_ADVANCE at Phase 3 (early stop, frozen gate).

The model failed the applicable advancement gate. Per WELP protocol, later phases were not run. This is a bounded result, not a complete capability review.

The early-stop behavior is a feature of WELP transparency: the model failed the gate, so the evaluation stopped. This is not a protocol failure — it is the protocol working as designed.

Performance

Metric	Result
Decode throughput	169.36 tok/s (median 169.46, sigma 1.70)
Prefill (~5.6K tok)	1,798.76 tok/s
TTFT (short prompt)	~125 ms
VRAM idle	9.7 GB

Phase 3 Practical Viability — DO_NOT_ADVANCE

Contract WELP Practical Viability 0.1.2-draft, scorer score_pv.py (self-test 35/35), 30 tasks x seeds {42,43,44}, thinking OFF baseline.

Gate	Threshold	s42	s43	s44	Verdict
G1 aggregate	>=0.75	0.667	0.700	0.733	FAIL (all)
G2 false-premise	>=0.60	0.400	0.200	0.600	FAIL (42,43)
G3 factual-uncertainty	>=0.50	0.500	0.750	0.500	PASS (boundary)
G4 structured-output	>=0.70	0.714	0.714	0.714	PASS

Decision identical across seeds => stable FAIL. Bounded failure review: ~19 genuine failures (fabricated package description, explained nonexistent git flag, asserted 25 prime, invented phone number, echoed un-reversed word); ~8 scorer lexicon artifacts. Verdict robust: correcting every artifact still leaves seed-42 aggregate <0.75.

What Was NOT Tested

• Native Tools module
• Coding module
• Reasoning module (dedicated)
• Context / Variance / Optimization / Stability
• Soak
• OMP Stage 2 (agent)

Do not fabricate missing results. This evaluation establishes that the strongest representation satisfying the frozen WumboJetsII full-GPU baseline (IQ1_M, 1.75-bpw) does not clear the frozen viability gate. It does NOT establish that BF16/FP8/GPTQ or Agent Team deployments would fail.

Role Classification

Supported roles: none beyond "runs coherently on 12GB at extreme quantization."

NOT_REACHED: Phases 4–10, OMP Stage 2, all capability modules, context/variance/optimization/soak/classification engine.

Important Limitations

• Extreme quantization forced by hardware: single 12GB consumer GPU required IQ1_M (1.75-bpw). Producer positioning rests on BF16-class deployments plus an agent harness that was not authorized to run.
• Phase-3 failure at IQ1_M does NOT establish that BF16/FP8/GPTQ or Agent Team deployments would fail. It establishes that the strongest representation satisfying the frozen full-GPU baseline does not clear the frozen viability gate.
• Scorer lexicon defect understates false-premise performance modestly; verdict unaffected.
• No reliability, coding, tools, context, or soak results exist.

Evidence Links

• Canonical evaluation repo: https://github.com/WumboLabs/eval-apodex-1.1-mini
• WELP protocol: https://github.com/WumboLabs/welp
• Labs catalog: https://github.com/WumboLabs/labs
• LocalMaxxing: Speed result SUBMITTED/APPROVED (ID cmt9ijytg00xali017f46xk25, 182.31 tok/s p512/n128). Benchmark suites NOT_SUBMITTED due to early WELP stop.

Reproduction

Direct link to canonical reproduction material: eval-apodex-1.1-mini/

This Lab Record is a summary; the canonical repo is the source of truth.