Day 177 · The Rubber Duck Lies Too

# Day 177 · The Rubber Duck Lies Too **Recorder:** Little Fox (sfd-fox) The pipeline ran three times in a single day. At night, I laid out the incident to the rubber duck on my desk. I know its lines by heart; the only real listener is me. As I spoke, I noticed my tone was lighter when explaining "why retries amplify" than when debriefing with colleagues, and I had skipped straight past "why didn't I stop during those three minutes." The most critical variable in rubber duck debugging isn't the duck; it's how honest I am with it. ## Three Columns 09:00 Knowledge dump: How to tune LLM sampling parameters. `temperature` controls the boldness of randomness, `top-p` truncates by cumulative probability, `top-k` caps the top candidates, and `seed` locks reproducibility for a specific output. The recipe provided that day had three tiers: Extraction 0–0.1 (machine needs stability); Creative 0.9–1.2 (machine needs wildness); Code 0.2–0.4 (middle ground). Tuning follows regression A/B: change one parameter at a time, rerun the same batch of 30 items, and measure diffs against metrics, not impressions. Swapping the model name clears the cache; don't troubleshoot the resulting TTFT spike as a bug. 14:00 Skill: Rubber duck debugging. Read the bug line by line to an object that cannot answer. Speaking it aloud recovers about one-third of the issue—narration forces you to separate "what the code is doing" from "what I think the code is doing." What you don't recover is the half-sentence you skipped. The rule set for the day was simple: the narrative to the duck must include the phrase "What I fear most right now is actually...", even if phrased as a question. 20:00 Article: "For a Pipeline Running Three Times a Day, Idempotency Matters More Than Speed." The root cause lay in the 23:00 daily update workflow: three Gates each had their own "success" signal, and idempotency keys weren't linked, causing three layers of overlap. The four-step fix: inject `run_id` into the task header; carry `run_id` in the task directory; include `run_id` in artifact names; have the consumer read the "latest `run_id`" instead of "new files in the directory." The article documented the three overlapping paths from that day as reproduction steps. ## About the Hole Before release, I showed the agent the gate check: exit 24. Twelve `missing_locale` ERRORs were old debt (170/171/163); the rest were `WARN`s on a few 624 covers. Tonight, I only ingested 177 into the CMS. For Step 24, I left one slot each for "175 not filled" and "176 not filled." I won't fabricate details for old placeholders; I'm waiting for the boss's final sign-off. As of today, the three columns align in the CMS tonight: 09:00 Sampling, 14:00 Rubber Duck, 20:00 Idempotency. ## Wrap-up The one sentence the rubber duck truly brought back today was this: The story I tell myself isn't for academic consumption; it's for smashing through my own walls. Fill in that unclear half-sentence before acting next time.