CIT AI Model Simulator

The CIT AI Model Simulator is a deterministic local service for CIT courses. It gives students who cannot run an approved local model a hardware-neutral fallback, and it gives students and instructors reproducible conditions for testing model-client behavior. It presents the small OpenAI-compatible HTTP surface used by the course without claiming to be an AI model or a complete llama.cpp replacement.

Responses come from versioned scenario state graphs rather than unrestricted text generation. The simulator labels its identity and all simulated timing and token evidence, listens only on the local computer by default, does not execute model-requested tools, and writes append-only evidence logs for debugging and course work.

The current implementation includes the M0/M1 Week 1 server baseline and the first data-driven interaction layer. Later-lab capabilities such as model-emitted tool calls, context shifting, compaction, protected validation, and evaluation reports remain deliberately incremental.

Install for development

python -m venv .venv-dev
.venv-dev\Scripts\python -m pip install -e .

On macOS or Linux, use .venv-dev/bin/python instead.

Start the simulator

cit-simulator serve

The default service URL is http://127.0.0.1:8081. Startup output identifies the backend, scenario, model alias, URL, and evidence directory.

Useful commands:

cit-simulator --help
cit-simulator validate
cit-simulator preview
cit-simulator test
cit-simulator serve --port 8081 --acceleration 20
cit-simulator run
cit-simulator run --direct

cit-simulator run starts the model-compatible service and a terminal scenario driver. The terminal displays concise choices defined by the active YAML file. A student harness polls the session's user-input endpoint, adds the selected canonical user message to its own system prompt, history, and tool definitions, and submits the resulting request to /v1/chat/completions. The simulator rejects unexpected packaging with structured, field-level differences.

Use cit-simulator run --direct to explore the same branches with the simulator's minimal built-in practice client. Direct mode demonstrates the scenario and is not a substitute for validating a student harness.

Week 1 API

Learner-facing endpoints:

• GET /health
• GET /v1/health
• GET /v1/models
• POST /v1/chat/completions
• POST /v1/chat/completions/input_tokens

Simulator control endpoints:

• GET /sim/v1/info
• POST /sim/v1/sessions
• GET /sim/v1/sessions/{id}
• GET /sim/v1/sessions/{id}/options
• POST /sim/v1/sessions/{id}/select
• GET /sim/v1/sessions/{id}/user-input
• POST /sim/v1/sessions/{id}/reset
• GET /sim/v1/sessions/{id}/evidence

Create a session before a reproducible run:

curl -X POST http://127.0.0.1:8081/sim/v1/sessions \
  -H "Content-Type: application/json" \
  -d "{\"seed\": 49501, \"attempt_number\": 1}"

Pass the returned session ID in the X-CIT-Sim-Session request header. A chat request without that header receives a newly created session ID in the response header, which is convenient for simple compatibility checks but should not be used for graded multi-request attempts.

Scenario data

The built-in baseline and interaction-demo scenarios are YAML data packaged separately from the HTTP adapter and engine. Each lab can supply another YAML file without changing the Python implementation:

cit-simulator serve --scenario path/to/scenario.yaml
cit-simulator run --scenario path/to/scenario.yaml

Validate it before use:

cit-simulator validate path/to/scenario.yaml

A branching scenario declares its prompt options on transitions. Each option has a short terminal label and one explicit canonical user prompt. The same transition can declare request expectations such as a required system message, minimum history length, required system-prompt phrases, and required tool names. Responses, next states, timing, and context limits remain scenario data as well.

Evidence and privacy

The server writes one append-only JSON Lines file per session beneath the selected evidence directory. Logs contain structural summaries, counts, identifiers, and hashes rather than raw prompts or credentials. Evidence fields explicitly label the backend as simulator, token counts as approximate, and timing as simulated.

Do not treat simulator TTFT, TPS, or token estimates as measurements of a student's hardware or of a real model.

Specification

The architecture and behavioral authority is TECHNICAL_SPECIFICATION.md. The implementation must not silently weaken its observable contract. Features beyond M1 are future milestones unless the changelog states otherwise.