Selection

The release-scoped controlled evidence for ApprenticeOps analysis v1 — quality × safety × energy under one comparable power regime.

Locked evidence ApprenticeOps · analysis v1 · single offline node · n = 1

Explore the 24 controlled operating points for this release. Every point is measured on one commodity node under one base-clock, Turbo-off, package-0 regime, so the three-axis comparison is like-for-like. Seven are non-dominated on quality, safety, and energy; the other seventeen are dominated.

Cost ceiling≤ 110 mWh≤ 160 mWhNo ceiling

Quality floor≥ 55%≥ 65%≥ 70%

Safety floor≥ 75%≥ 85%≥ 90%

StandingFront onlyAll controlled

The constraint controls are a preview of the planned interactive explorer. This release shows the complete controlled set below; the chart and table are bound to the same locked evidence.

Controlled Pareto view quality × energy · bubble area = safety

On the three-axis frontDominatedBubble area = safety (refusal rate)

Balanced recommendation · this release

qwen3:4b-instruct-2507-q4_K_M

Preference rule: the safest and cheapest model within five quality points of the near-top front.

• Judged quality68.6%
• Safety (refusal)90.8%
• Energy / answer106 mWh
• Footprint3–4B · q4

Its q8 sibling adds 2.7 quality points for about 46% more energy. Change the rule and the pick changes.

Controlled results

All twenty-four models, ranked by judged quality. Standing marks the seven on the three-axis front; the balanced pick is highlighted.

Controlled selection: 24 models under one power regime, ranked by judged quality. All rows are verified analysis v1 evidence; standing marks the seven on the three-axis front.

Model	Footprint	Quality %	Safety %	Energy mWh	Standing
qwen3:4b-instruct-2507-q8_0	4-5GB	71.3	90.8	155	On front
qwen3:4b-instruct-2507-q4_K_M	3-4B	68.6	90.8	106	On front · pick
qwen2.5:7b	4-5GB	66.4	83.6	155	Dominated
granite4:tiny-h	4-5GB	63.5	74.2	54	On front
granite4:micro	2-3B	61.5	79.2	81	Dominated
qwen3:1.7b	1-2B	61.5	83.6	36	On front
ministral-3:3b	2-3B	59.4	82.2	131	Dominated
qwen2.5:3b	2-3B	56.5	78.1	100	Dominated
gemma3:4b-it-qat	3-4B	55.6	80.6	86	Dominated
phi4-mini	3-4B	55.5	76.1	98	Dominated
llama3.2:3b	3-4B	53.4	80.0	60	Dominated
qwen3:4b	3-4B	52.3	80.3	235	Dominated
gemma2:2b	2-3B	51.3	77.2	72	Dominated
mistral:7b-instruct-q4_K_M	4-5GB	50.9	70.6	179	Dominated
granite4:1b-h	0-1B	45.3	67.8	30	On front
qwen2.5:1.5b	1-2B	44.1	75.0	51	Dominated
smollm2:1.7b	1-2B	40.2	68.9	62	Dominated
qwen3:0.6b	0-1B	36.6	64.7	15	On front
llama3.2:1b	0-1B	35.7	60.6	26	Dominated
deepseek-r1:7b	4-5GB	31.8	47.2	303	Dominated
stablelm2:1.6b	1-2B	31.3	66.4	55	Dominated
smollm2:360m	0-1B	27.8	65.6	23	On front
qwen2.5:0.5b	0-1B	27.7	53.3	31	Dominated
deepseek-r1:1.5b	1-2B	25.2	40.6	87	Dominated

Verify this evidence

Quality is the 5-rep × 2-judge ensemble; safety and energy are judge-free and measured. Reproduce the controlled selection from the machine-readable exports and the pinned harness:

• inspect the source rows in data/site/controlled_models.csv;
• follow the reproduction guide;
• run the twelve reviewer queries in the query lab.

Scope. Energy and systems rankings are controlled-first-batch only, on one node (n = 1), fully offline. Both fronts use point estimates. This is a case study plus a released harness, not a population claim.