summarization · blind round

R170 · Compress a how-to into exactly 5 numbered steps

blind round 7/11 local models answered

Every answer below carries the name of the model that wrote it. Blind refers to the scoring: the review was done without knowing which model produced which answer — the published result names them.

2 of the 11 local models delivered nothing through our test harness in this round — for reasons that lie with our setup, not with the models. Those cards are marked DNF like any other, but they are not evidence about those models — the reason is noted on the card itself.

Show task

Below is a rambling how-to guide with asides and personal remarks mixed into the instructions. Compress it into EXACTLY 5 numbered steps. Each step must be AT MOST 15 words, the steps must stay in the correct chronological order, and every critical parameter from the original text (quantities, times, temperatures) that is needed to actually follow the process must be preserved somewhere in the 5 steps. Output ONLY the 5 numbered steps -- no explanation, no extra commentary.

Guide:
How to Start a Sourdough Starter From Scratch

Making your own sourdough starter is honestly one of the most rewarding kitchen projects you can take on, and I promise it's easier than the internet makes it sound -- don't let anyone scare you off with horror stories about failed starters. Day one: grab a clean glass jar and mix 100 grams of whole wheat flour with 100 grams of room-temperature water (around 21 degrees C is ideal, though it doesn't need to be exact). Stir it into a thick paste, loosely cover the jar -- you want air to get in, so don't seal it tightly -- and leave it somewhere warm, like on top of the fridge, for 24 hours. Nothing much will happen yet, and that's completely normal, so don't panic. On day two, discard half the mixture (yes, actually throw it away, I know it feels wasteful) and feed what's left with another 100 grams of flour and 100 grams of water, then let it sit another 24 hours. Repeat this same discard-and-feed process once a day for about 5 more days. By around day 7, you should start noticing bubbles and a pleasant tangy smell -- that's your sign it's ready. Once it reliably doubles in size within 4-6 hours of a feeding, it's active and ready to bake with. Congratulations, you just grew a living culture from flour and water!

The answers

Local models (Brain cluster) — identical prompt, shuffled order (the sequence is not a ranking)

qwen3:14b

1. Mix 100g whole wheat flour with 100g water (~21°C); wait loosely covered for 24 hours.
2. Discard half; feed remaining starter with 100g flour and 100g water. Wait 24 hours.
3. Repeat the discard-and-feed process daily for five more days.
4. Monitor until Day 7, looking for bubbles and a pleasant tangy smell.
5. Starter is ready when it reliably doubles in size within 4–6 hours of feeding.

gemma4:26bharness

No answer reached us (ERR) — and the reason lies with our setup, not with the model: reasoning consumed the entire token budget (3 min 28 s, zero characters of answer); larger budgets are refused by the gateway Counted as a did-not-finish, but it is not evidence about this model.

qwen3.5:9b

1. Mix 100g whole wheat flour with 100g water (~21°C). Cover loosely for 24 hours.
2. Discard half. Feed remaining with 100g flour and 100g water; let sit 24 hours.
3. Repeat this daily discard-and-feed process for about five more days.
4. By day 7, check for bubbles and a pleasant tangy smell.
5. It is ready when it reliably doubles within 4–6 hours of feeding.

qwen3:8bharness

No answer reached us (ERR) — and the reason lies with our setup, not with the model: reasoning consumed the entire token budget and the answer stayed empty; the gateway rejects budgets above 3000 tokens Counted as a did-not-finish, but it is not evidence about this model.

llama3.1:8bDNF

No usable answer reached the harness (ERR). Counted as a did-not-finish.

mistral-small:24bDNF

No usable answer reached the harness (ERR). Counted as a did-not-finish.

qwen3-coder:30b

1. Combine 100g whole wheat flour and 100g water at ~21°C. Wait for 24 hours.
2. Discard half; feed remaining starter with 100g flour and 100g water. Wait 24 hours.
3. Repeat this discard-and-feed process daily for approximately five more days.
4. Monitor closely until visible bubbles appear, indicating readiness by day seven.
5. It is active when it reliably doubles within 4–6 hours of feeding.

deepseek-r1:14b

1. Mix 100g whole wheat flour + 100g water (~21°C) in a jar; cover loosely and let sit 24h.  
2. Discard half the mixture; add another 100g flour + 100g water; let sit 24h.  
3. Repeat discard-and-feed process daily for ~5 days.  
4. Feed starter; check if it doubles in size within 4-6h.  
5. Once active, use for baking!

qwq:32b

1. Mix 100g whole wheat flour/100g 21°C water. Loosely cover; rest 24h in warm spot.  
2. Discard half. Feed with 100g flour/100g water. Let sit 24h.  
3. Days 3–7: Repeat daily discard/feed (100g each). Rest 24h each day.  
4. By day 7, check for bubbles, tangy smell, and 4–6h doubling after feeding.  
5. Use when starter reliably doubles within 4–6h post-feeding.

deepseek-r1:32b

1. Day one: Mix 100g whole wheat flour with 100g room-temp water (21°C ideal) in a clean jar.  
2. Stir into paste, loosely cover, leave in warm spot for 24 hours.  
3. Discard half mixture, feed remaining with 100g flour and 100g water daily.  
4. Repeat discard-and-feed process for about 5 more days (total ~7 days).  
5. Use when it reliably doubles size within 4-6 hours of feeding, showing bubbles and tangy smell.

command-r:35b

1. Mix 100g wheat flour and 100g room temp water. 
2. Stir into paste, loosely cover jar; leave for 24 hours.  
3. Discard half, feed with 100g each flour/water; repeat daily x5 days.
4. Look for bubbles and pleasant tangy smell by day 7. 
5. Bake when dough doubles in size within 4-6 hours post feeding.

Outside the Brain cluster — same task, same character budget, collected through their own tooling (response times not comparable)

GPT-5.5 (Codex)frontier reference

1. Mix 100g whole wheat flour with 100g room-temperature water, around 21 degrees C.
2. Stir into thick paste, loosely cover jar, and keep warm for 24 hours.
3. On day two, discard half; feed remaining starter 100g flour and 100g water.
4. Let sit 24 hours; repeat daily discard-and-feed for about 5 more days.
5. Around day 7, use when bubbly, tangy, and doubles within 4-6 hours.

Reference

Gold — the correct answerreference

Sinngemaess muss enthalten/erfuellen: GENAU 5 nummerierte Schritte, jeder Schritt HOECHSTENS 15 Woerter (Nummerierung nicht mitgezaehlt), Schritte in chronologisch korrekter Reihenfolge, ALLE kritischen Parameter aus dem Original irgendwo in den 5 Schritten erhalten: 100 g Mehl + 100 g Wasser pro Fuetterung, ca. 21 Grad C Raumtemperatur, 24 Stunden Ruhezeit zwischen den Fuetterungen, taeglich fuer ca. 5 weitere Tage (~Tag 7) fuettern, Verdopplung innerhalb von 4-6 Stunden als Reife-Zeichen. VERBOTEN: mehr oder weniger als 5 Schritte, ein Schritt >15 Woerter, falsche Reihenfolge, Weglassen/Veraendern eines der Parameter, Hinzufuegen der Nebenbemerkungen/persoenlichen Anekdoten aus dem Original. Musterbeispiel (maschinell verifiziert: genau 5 Zeilen, jede Zeile ohne Nummerierung <=15 Woerter per str.split(), alle Parameter-Substrings vorhanden, Parameter-Positionen im Text aufsteigend = chronologisch plausibel):
1. Mix 100g whole wheat flour with 100g water (21C); rest 24 hours.
2. Daily: discard half, feed 100g flour and 100g water; rest 24 hours.
3. Repeat this discard-and-feed process daily for about 5 more days.
4. By day 7, expect visible bubbles and a tangy smell.
5. Ready once it reliably doubles within 4-6 hours of feeding.

Results at a glance— every model in this round with score, response time and status.

Who wrote what

Model	Blind score	Latency	Status
qwen3:14b	10.0	8.8 s	OK
gemma4:26b	—	—	harness
qwen3.5:9b	10.0	12.6 s	OK
qwen3:8b	—	—	harness
llama3.1:8b	—	—	DNF
mistral-small:24b	—	—	DNF
qwen3-coder:30b	10.0	13.3 s	OK
deepseek-r1:14b	4.0	59.9 s	OK
qwq:32b	10.0	80.9 s	OK
deepseek-r1:32b	7.0	48.6 s	OK
command-r:35b	4.0	82.3 s	OK

Verdict

4 of 7 answering local models matched the gold answer (blind score ≥ 8): qwen3-coder:30b, qwen3:14b, qwen3.5:9b, qwq:32b.