An early workflow target for producing mid-sized research output while keeping inference spend legible.
A cost envelope, not a price promise
North uses a sub-$3 target for a mid-sized 2,000–3,000 word research synthesis as an engineering constraint. It is not a guaranteed retail price. Actual spend changes with source count, context size, model routing, retries, and the subject being researched.
Publishing the envelope forces the pipeline to account for every pass instead of treating inference cost as invisible infrastructure.
Spend capability where it matters
Retrieval cleanup and extraction can often use smaller models or deterministic code. Contradiction resolution and final synthesis may deserve a more capable model. Routing by stage reduces repeated context and makes quality-cost tradeoffs explicit.
Cached evidence artifacts also allow a user to revise the framing or regenerate the document without purchasing the same retrieval work twice.
Quality controls remain binding
A workflow does not meet its budget goal if it saves money by dropping necessary sources or compressing away uncertainty. Cost evaluations therefore run beside evidence coverage and claim-support checks.
The target will be reported with prompt, model, and infrastructure assumptions. Without those conditions, a single price number is not reproducible.