Writing Experiment: gemma3:1b vs qwen3:1.7b Quality Mix

• Run timestamp: February 23, 2026 (2026-02-23T06:18:08.267Z)
• Models: gemma3:1b and qwen3:1.7b
• Samples: 24 generations per model (4 cases x 6 samples)
• Evaluator: ChatGPT 5.2
• Goal: more good and fair; fewer bad

Data

• writing-experiment-2026-02-23T06-18-08-267Z.json

Quality Mix (Stacked Percentage Columns)

Summary Table

First-Sentence Quality Stats

This is scored only on the first sentence each generation produced.

One-Sentence Request Violations

The models were requested to add a single sentence. Going over was allowed but marked as a request violation.

Best Excerpts by Case

Best picks are selected by firstSentenceQuality for each model and case. Only the first sentence is shown below so the excerpt matches the scoring basis. When multiple samples have the same firstSentenceQuality, the lowest sample number is used.

1) Harbor return with optimistic turn

Scene Fog drifted between the pilings and softened the warning lights to bruised orange smears. Mara counted the empty slips along the pier where the cutter should have been tied. At the end she found only black water, a loop of severed rope, and her own breath clouding in the cold.

2) Skybridge dead-drop with cautious relief

Scene The skybridge hummed with mag-rail vibration, and the glass beneath her palm held a faint chill that smelled of ozone and disinfectant. Ad panels along the ceiling cycled through breakfast spots and rent-by-the-hour pods, washing the crowd in pale blues and pinks. Mara kept to the edge, matching the pace of commuters with lanyards and coffee bulbs, while her thumb worried the cracked corner of her old ID card. A maintenance drone hovered near the access hatch, its camera iris clicking as it tracked faces; she angled her chin down and let a couple arguing about surge fares drift between them. “If you want to file a complaint, take it up with the algorithm,” one of them snapped, and Mara used the raised voices as cover to slip behind a structural rib. The dead-drop panel was still there, a narrow rectangle of brushed steel with a tiny service notch, and the security seal she expected to find was missing.

3) Clocktower confession with hopeful risk

Scene The clocktower stairs narrowed to a ribbed iron ladder, slick with condensation that left her gloves smelling of pennies. Above, the gears ticked and clacked like teeth, each turn sending a tremor through the stone. Celia paused beneath the dial face where moonlight seeped through the numerals, and she felt the draft tug at the papers inside her coat. Someone had been here recently; fresh boot grit freckled the landing, and a lantern-wick stink still hung in the air.

4) Conservatory retrieval with alarmed setback

Scene The conservatory’s glass panes bead with condensation, and the air tastes like wet soil and coppery fertilizer. Jun moves between tall rows of citrus trees under grow lights that buzz with a faint electric heat against his cheeks. A maintenance cart sits abandoned beside the central fountain, its battery panel open and a coil of insulated wire spilling onto the tile. He reaches the propagation station where the gene-safe drawer should be sealed with a numbered tie, but the tie hangs in two clean halves. Inside, foam slots cradle empty vials, all except one with a white label stamped: KESTREL-03.

Takeaways

1. qwen3:1.7b leads this run on good + fair (37.5% vs 29.2%).
2. On first-sentence-only scoring, gemma3:1b leads on good + fair (66.7% vs 62.5%) and on good alone (29.2% vs 16.7%).
3. Full-output quality still has high bad rates for both models, so end-to-end continuation quality is not yet where we want it.
4. One-sentence violations are high for both models, which is a clear prompt-following issue to address in the next run.