Term
Ground Truth
A reference taken to be correct, against which model outputs are measured. In AI code analysis it is a piece of code with deliberately injected, known defects, so you can measure how many real problems a model actually finds.
Ground Truth — explained in detail
Ground Truth denotes a reference that is treated as correct by definition and against which a model’s outputs are measured. It is the benchmark used to decide whether a model is right or wrong.
In the context of AI-assisted code analysis, ground truth is a piece of code into which a known set of defects has been deliberately injected. Because you know in advance which and how many problems are actually present, you can measure what share of them a model detects (see recall). Without this known truth, you can only assess what a model claims to find — not what it misses.
That is precisely the value: ground truth makes the invisible visible. False positives (reported issues that are not real) and false negatives (real issues that go undetected) only become measurable once a reliable reference exists.
Example / Practical use
You prepare a codebase with 20 deliberately injected weaknesses — for instance an SQL injection, an off-by-one error and a missing null check. These 20 are the ground truth. If a model then runs its analysis and correctly reports 12 of them, recall is 60 percent.
Without that reference you would only know: “The model reported 12 issues.” Whether that is good or poor, whether eight real defects slipped through — that would remain unknown. Ground truth turns a mere claim into a verifiable hit rate.
Distinction from related terms
Ground truth must not be confused with groundedness or faithfulness. Groundedness asks whether a model’s statement is supported by a provided source text (i.e. not hallucinated). Ground truth, by contrast, is the external, correct-by-definition reference against which results are evaluated. One concerns fidelity to the source, the other the benchmark of evaluation — two distinct concepts that are easily mixed up because of their similar word roots.
Discover more
AI Workflows by Keyword: How We Make Recurring Routines Enforceable
A typed keyword triggers a fixed AI routine — and every single step must be committed before the next one appears. Why that's the actual trick.
GlossaryRecall
Recall (hit rate, sensitivity) measures the share of all actually existing relevant cases that a system finds. Formula: relevant cases found divided by all relevant cases that truly exist. Requires a known ground truth.
EncyclopediaAgent Evaluation — Task Success Rate, Eval Harnesses, Benchmarks
How to evaluate AI agents: eval harnesses, task success rate, benchmarks like GAIA and Tau-Bench, LLM-as-a-judge — and how it differs from pure model quality.