This is an illustrative sample of the README the Art of the Possible skill ships with every generated project, produced for a sample company, Mastercard. It demonstrates the output format only. Mastercard is not affiliated with and has not endorsed these materials.

# Settlement Reconciliation, Art of the Possible Build

A working pipeline built in one session with the Cheek Art of the Possible skill, for the operating problem named in the interview. Three internal systems record the same settlement activity and do not talk to each other. Finance reconciles them in spreadsheets at month end, exceptions surface weeks late, and the insight, which issuers drive the exception load, never reaches the weekly operations review.

## What this project does

The pipeline runs four stages, each deliberately placed on the earliest rung of the Five Generations ladder that clears the bar. Two stages use no AI at all, because they did not need it.

1. Ingest. Reads settlement extracts from the three systems, or generates realistic sample data standing in for them. No AI, file parsing is a solved problem.
2. Reconcile. Matches records across systems with rule-based matching on transaction reference, amount, and date windows. Generation one, rules, because the matching logic is knowable and auditable.
3. Score. Flags exception records by likelihood of write-off using classic machine learning trained on the sample history. Generation two, classic ML, because the signal lives in structured features.
4. Brief. Assembles the weekly exceptions brief with the top issuers by exception load and the recommended actions. A template fills it, generative AI drafts only the two narrative paragraphs where a template would read like a form letter.

## Run it

    ./run.sh

Then open http://localhost:8787. The No One Works Here web UI animates each stage as it runs, shows the reconciliation counts live, and renders the brief with its recommendations when the pipeline lands.

## The build fork

This sample repository is the fast build. It runs entirely on generated sample data, needs no API keys, and exists so the room can see the shape of the answer in minutes. The real build adds the actual system integrations, a stamped .env for credentials, and the six-seat compliance review panel, which reviews the generated codebase and reports before anything touches production.

## Files

- pipeline/ - the four stages, one module each, with the generation choice documented at the top of each file
- data/ - sample data generators and any uploaded extracts
- server/ - the local web UI
- results/ - reconciliation output, scores, and the generated brief
- docs/ - the interview record and the ladder decisions, stage by stage

Questions at any point are welcome at skill-help@paulcheek.com.