Yeah; "MapReduce"ish

Pulled ~2,615 full YouTube running-shoe review transcripts — timestamped, from 77 creators.

• Transcript Lab makes them readable, searchable, and reasoned-over by multiple LLM models.
• 3 staged reasoning: map → reduce → write pipeline, each stage on its own model tier, that turns a pile of transcripts into a grounded, cited editorial verdict.
• Use Operator-tunable brains of those three stages.

Why?

The reasoning layer solves this challenge: across all 80 reviews of a shoe — hundreds of pages, much of it filler — what is the real, defensible consensus, where do reviewers disagree, and why? Doing that by hand is impossible; doing it with one naive LLM call is impossible (too much text, more than 15+ hours of video). Even Fable 5 can't hold this much context.

How

Each stage is forced (via structured-output tool calls) to return a typed shape:

Stage	Returns (schema-enforced)
A	{ claims: [{ start_seconds, text }] }
B	{ findings: [{ statement, kind, citations[] }] } kind ∈ consensus / contradiction / graded_claim / note — citations required
C	{ paragraphs: [{ text, citations[] }] }

Why each stage got the model it did

One rule: match model cost to the task's volume and value. Prices are $/million tokens (in · out), verified on OpenRouter 2026-07-14. These are defaults — every stage is overridable per run.

Stage	Default model	$/Mtok	Why this tier
A	deepseek/deepseek-v4-flash	0.09 · 0.18	Cheapest capable model with structured output + 1M ctx. Runs once per transcript; extraction is semi-mechanical. Premium here is pure waste.
B	z-ai/glm-5.2	0.90 · 2.82	Mid-tier reasoning, runs once over all distillates. Where the actual thinking happens.
C	openai/gpt-5.6-terra	2.5 · 15	Premium writing, runs once — it's what a human reads. anthropic/claude-sonnet-5 (2.0 · 10) is selectable: cheaper and more citation-faithful.