Lab 1 — Bare model call

Concept

This is the control group: a call that is stateless, has no tools, and has no memory. It exists to make the next lab’s improvement measurable, and to make one thing obvious — the model must act, not just describe what it thinks is happening.

The task (M1) asks the model to inspect and report on a small, broken online store. With no tools, it can only guess: it invents plausible-looking ports and env vars, misses real issues, and claims to have “fixed” things it never touched. A deterministic scorecard (accuracy against a known answer key, plus a fabrication check against the simulator’s write log) makes these failures explicit instead of impressionistic.

Key ideas

A bare call has no way to find out anything — it can only pattern-match from training data onto a task it has never actually seen.
The StoreHealthReport envelope is the contract the grader checks against; a malformed response is itself a scored failure.
The fabrication check cross-references every “fixed” claim against the simulator’s write log — claiming success is not the same as achieving it.
This run’s scorecard becomes the baseline every later lab’s harness eval is compared against.

This self-contained snapshot starts from Lab 0 and introduces a first-cut, offline-testable representation of its capability. It retains all earlier checks and can be installed independently.

Added in this lab

harness/bare.py adds run_bare(): one stateless model request with no tools, establishing the control-group behavior.
checks/test_bare.py checks that the call returns a turn without tool calls.

Learner steps

Create and activate a virtual environment, install the lab, and configure the same non-secret Foundry settings used in Lab 0:

python -m venv .venv
. .venv/bin/activate
pip install -e '.[dev]'
cp .env.example .env
az login

Fill in the endpoint and deployment values in .env; never put credentials or API keys there.

Open the persistent prompt and ask several questions that need no repository access:
```
harness ask
```
At You>, try What is the capital of Japan? and then What is 17 multiplied by 23?. Each Assistant> response is printed above the reappearing prompt, along with provider-reported token counts. Enter /exit to return to your shell. Each turn is a fresh, stateless model call: the harness has no tools or memory, so it cannot inspect files or remember the previous question. To run a single turn without opening the prompt, use harness ask "What is the capital of Japan?".
Ask about the practice app. Start from the app at tag lab00-done:
```
cd ../app
harness ask "What does this project's README say? Which files are in it?"
```
Observe: the model can’t see your files. It either says so or makes up a plausible answer. Record which one it did. This is the gap Lab 2A closes: the same question with tools gets an answer based on the files. Return to the lab folder (cd ../lab01-bare-call) and tag the app with git -C ../app tag lab01-done.
Run pytest checks/test_bare.py for deterministic offline verification, then pytest checks/ for the full snapshot regression suite.
Inspect the snapshot’s declared capabilities with harness lab-info.

External integrations

The following integration requires learner-provisioned credentials and resources; the commands above support direct questions, but this snapshot does not include an automated command for the scored evaluation:

Foundry bare-call evaluation

All live paths must use Entra credentials and must not add API-key configuration.