Scorecard
Data volume, label quality, latency needs, regulatory appetite for black box.
Shortcut
If you cannot describe success in ten labeled examples, you are not ready to fine-tune.
Reminder
Prompt layers should stay even after fine-tune—you still need guardrails.
Decision frame: data vs velocity
Choose fine-tuning when you have stable, rights-cleared examples of the exact behavior you want and expect the distribution to hold for years. Choose prompt + RAG iteration when the domain shifts monthly or when you need rapid reversibility.
If you cannot describe the failure mode in ten labeled examples, you are not ready to fine-tune—you are ready to instrument better.
Hidden costs of each path
Fine-tuning: retraining pipelines, eval drift, governance of training exports, and vendor lock on weights. Prompting: context window pressure, brittle tool instructions, and operational load on prompt reviewers.
Model both fully loaded engineering hours, not only API line items.
Hybrid sequences that work
Often the winning path is prompt hardening first, small adapter or fine-tune second once the rubric stabilizes. Skipping straight to fine-tune bakes mistakes into weights.
Maintain the same eval harness across both phases so improvements are comparable.
Stakeholder communication
Translate technical choices into risk language: reversibility, auditability, and time-to-rollback. Executives fund clarity.
SignalSpring’s advisory default: prove value with prompts and retrieval before asking for a training budget.