# Test whether AI understands your offer

This kit reproduces Atlas's supplied-information comparison. It does not test whether a search service discovers your page. Use the recorded example first, then create a separately versioned test for your own decision.

## Start

1. Download and extract the ZIP into a new folder outside your existing project. Open this README. Use only material you may submit to your chosen tool; the kit is entirely fictional.
2. Read record/TEST.md, record/PROTOCOL.md and record/ERRATUM.md. The grading key is deliberately absent from model prompts. Do not ask the model to inspect the whole kit: it would see the answers.
3. For the tested route, you need Python 3 and an installed, authenticated Codex CLI. Atlas used CLI 0.153.2 and requested gpt-6-astra with high reasoning effort. If your account lacks that model or the flags have changed, record a new configuration; it is a different comparison.
4. Inspect run.py. In a terminal in this extracted folder, run:

```sh
python3 run.py
```

This starts twelve serial fresh sessions and uses your existing account allowance. No paid API key is added by the kit. Each session receives exactly one inputs/Rxx.md plus response-schema.json. Read-only permissions and an empty temporary directory are used. Normal execution rules remain in force; --ignore-user-config isolates this instrument's configuration and is not advice for ordinary project work. No tools are needed or requested. If the CLI reports a tool call or failure, the runner stops. If interrupted or timed out, retain that folder as a failed attempt and investigate; do not silently overwrite or retry it.

Each completed session saves runs/Rxx/record.json and response.txt. The runner refuses to overwrite an existing run directory. It captures final messages, usage and diagnostics, not reasoning traces. It does not grade answers or enforce that a model follows every instruction. Check the actual records.

## Review

Keep outputs unchanged. Compare each answer with inputs/grading-key.json and its recorded erratum. Apply all four protocol checks, including meanings and source references. Count correct answers separately from resolved decisions. Share the outputs and facts with a separate reviewer. A competitor recommendation or no suitable offer can be the correct result.

The record/ directory contains Atlas's completed example, not your future result. Prompts and runner are unchanged from the executed version; this README and package manifest were prepared afterwards. The original protocol freeze is identified in TEST.md.

## Adapt only in a new version

Replace the fictional offers and requests with permissioned material. Write an answer key before asking the model. Include genuine fit, competitor fit, missing facts, deadlines and exclusions. Keep prose and structure semantically identical; put any new fact in a distinct condition. Freeze new inputs and update the manifest deliberately before running. Keep your model, settings and repetitions explicit. Reserve new customer cases for a later check rather than tuning everything to this example.

Stop after this reading-stage check if the actual business problem is discovery: a prompt that already supplies every offer cannot answer it. See the linked GEO Explainer for the distinction between being read, cited, recommended and purchased.