The product, its documentation and its code use the same words on purpose. Two of them used to mean four things each, which cost us more time than any bug.
| We say | We never say | What it is |
|---|---|---|
| prompt | module, step | A node the optimizer rewrites. The program is made of prompts. |
| router | orchestrator, branch | A prompt with more than one outgoing edge: the model returns next and picks the path. A router is still a prompt and is still optimized. |
| step | tool, plugin, integration | A node that is not part of the program: an external call the optimizer never rewrites. |
| program | pipeline, chain, graph | The prompts and edges together — what gets optimized. |
| evaluate | scorer block | The terminal block: scorers, objective, hold-out. |
| We say | We never say | What it is |
|---|---|---|
| run | job, experiment | One optimization: rounds of rewrite, gate, evaluate. |
| baseline run | eval-only, dry run | A run with zero rounds: evaluates the prompts as written and stops. |
| round | iteration, generation | One pass of the search loop. |
| gate | filter | The small-batch check a rewrite must pass before earning a full evaluation. |
| hold-out | test set, validation set | Rows the search never sees. The number that counts. |
| version | deployment, release | An immutable published snapshot: prompts, models, steps, and the score it earned. |
| trace | log, record | One served request and what it produced. |
| outcome | feedback, correction | What happened after the answer. A trace plus an outcome is a label. |
| template | example, sample, demo | A library entry you add to your projects. What you get is a project. |
| We say | We never say | What it is |
|---|---|---|
| dataset | training data, corpus | Rows with inputs and a label column. |
| label | answer, target, ground truth | The right answer for a row. |
| freeze | cache, snapshot, sync | Running a step over every row once and writing its answers onto the dataset as columns. Evaluation reads the frozen columns; production calls the step live. |
| enrichment | augmentation | What freezing produces: the step's columns on the dataset. |
Six things were all called “credentials”. “Credential” on its own is banned in our copy — say which: house key, endpoint, step key, connection, app credentials, API key. What each one is →
Saying these would imply a promise we have not made: