AI Model

Origin & evolution

The discipline of documenting a model as a reportable artefact crystallised with Model Cards for Model Reporting, introduced by Margaret Mitchell and colleagues (including Timnit Gebru) at the 2019 ACM Conference on Fairness, Accountability, and Transparency. A model card records intended use, training data, evaluation metrics🔢MetricStrategyA unified metric that measures progress, health, or behaviour across the productView reference → broken down by group, and known failure conditions. The format has since been adopted by Hugging Face, Google, and most foundation-model providers as the default way to ship a model with its provenance attached.

The foundation-versus-fine-tuned split followed the rise of large pre-trained models. A foundation model is trained once at scale; teams then fine-tune or prompt it for a specific task☑️TaskProduct SpecificationA unit of work within a story or epicView reference →. That changes the dependency graph: a product may rely on a vendor's base model and a separate fine-tune layer, each with its own version and eval history.

How it works in practice

A support product routes tickets through claude-sonnet-4 for triage. The team pins the version, records its evaluation score against a 500-ticket benchmark (87 per cent correct routing), and tracks cost per thousand tokens. When the vendor releases🚢ReleaseProduct SpecificationA shipped version of the productView reference → a newer checkpoint, they re-run the same benchmark before switching: the new model scores 91 per cent on routing but regresses on multilingual tickets, so they hold the upgrade for non-English queues. The model card makes that decision auditable.

AI Model vs. its neighbours

• AI Dataset🪆AI DatasetAI & Machine LearningA dataset for AI training or evaluationView reference → is the corpus a model learns from. The dataset is the input to training; the model is the trained output. Contamination or licensing problems in the dataset surface later as model behaviour, which is why the two are tracked as separate, connected assets.
• AI Guardrail🚦AI GuardrailAI & Machine LearningA guardrail for AI safetyView reference → sits around the model at runtime, filtering inputs and outputs. The model decides what to say; the guardrail decides what is allowed through. Swapping the model does not change the guardrail, and vice versa.
• Eval Benchmark🎖️Eval BenchmarkAI & Machine LearningA benchmark for evaluating AI qualityView reference → is the measuring instrument. The benchmark is fixed; the model's score against it is what moves. Comparing two models means holding the benchmark constant.

In the graph

In the Unified Product Graph, an AI model lives in the AI and intelligence region as a first-class dependency. A product reaches it through product_powered_by_ai_modelProductpowered byAI Modelhierarchy; its behaviour is shaped by ai_model_prompted_via_prompt_versionAI Modelprompted viaPrompt Versionhierarchy, measured by ai_model_benchmarked_by_eval_benchmarkAI Modelbenchmarked byEval Benchmarkhierarchy, and metered by ai_model_costed_by_ai_cost_trackerAI Modelcosted byAI Cost Trackerhierarchy. Modelling the version, eval, and cost as distinct edges means an upgrade decision can be traced end to end: which prompts break, which benchmark moved, and what the new spend looks like.