Paper Trail: try the check number

A voice recording of a walk becomes a puzzle hunt. This page lets you try the part that checks the answers.

What this page is, and isn't. It uses an invented route, so my real one stays private. The clues below are hand-written samples that show the format. In the real tool, Whisper and Gemma 4 run locally on a GPU laptop and write the clues; they can't run in a web page. What does run here is the checking logic: letter rules, the check number and the debrief. It is a small port of the same arithmetic as the Python code in the repository.

1. Walk the sample hunt

A team gets a sheet like this. At each place they read the right sign and write down one letter. The letters, in order, spell a word.

2. Fill in your letters

The sheet says the check number is 20. Give each letter a value (A=1 ... Z=26), add them up, and while the total is above 26 subtract 26. If your number differs from the sheet, a letter is wrong or a sign has changed.

3. Debrief: what did a team write?

Type the word a team wrote on its sheet, right or wrong. The tool compares it with the key and points at the stop that was probably misread or has changed.

4. Try the rules on your own sign

Each stop gives one letter by a rule. Counting letters only, with no spaces, apostrophes or digits, the sheet also tells players how many letters the right sign has, so they never read the wrong sign.

Rule:

How the hunt is made

I scout a loop with a voice recorder, saying the sign text and one visible detail at each stop. Whisper transcribes it, Gemma 4 structures the stops, and plain code finds words that the signs' letters can spell in walking order. Gemma writes the clues, and a solver pass and a truthfulness pass reject clues that point to two places or describe things I never said. The rule I held to: the models propose and write, and code decides anything that can be checked.

The code, the tests and the write-up are linked from the DEV post for the Hacktoberfest Open-Source AI Challenge, Week 1.