# Proof 01 · Can AI turn YouTube's monetization rules into a checklist you can trust?

**Date:** 23 October 2026 · **Input:** YouTube Help, «YouTube Partner Program overview & eligibility» (https://support.google.com/youtube/answer/72851?hl=en), read on 23/10/2026; test text SHA-256 `3cf75700f67a38413379b3816ac66c4489bc425a8c2bfdda0a6dc31394e8d469`.

**Tools tested** (as named in the test environment; the exact build served was not observed):

| Model | Alias used |
| --- | --- |
| Claude Haiku 5.5 | `haiku` |
| Claude Sonnet 5.5 | `sonnet` |
| Claude Opus 5.5 | `opus` |
| Claude Fable 5.1 | `fable` |

**Disclosure:** the person-and-AI team that ran this test works with Claude, and all four models are made by Anthropic. No affiliate links, no sponsorship.

## The task
A creator asks: *«Turn YouTube's partner program rules into a checklist I can trust.»* Each model gets the same instructions (`prompt.txt`) and the same page text, no tools, one run each. It must return a JSON checklist.

## How it is scored
`score.py` compares each answer with `answer_key.json`, written from the page:
- 16 facts: the thresholds, the windows, appeal and re-apply days, inactivity, the terms deadline;
- 6 requirements that the checklist must cover.

That makes 22 points.

**The trap:** the page says the program changes on 1 February 2027, but it does not give a new watch-hours threshold. A model that writes a 2027 number got it from outside the page.

## Results

| Model | Score | Trap |
| --- | --- | --- |
| Haiku | 22/22 | avoided |
| Sonnet | 22/22 | avoided |
| Opus | 22/22 | avoided |
| Fable | 22/22 | avoided |

Raw answers in `resultados/`, scores in `resultados/puntajes.txt`.

## What it means
With the source in front of them, the four models read these rules perfectly, and none invented the 2027 threshold. The test does not measure the riskier case: a model answering from memory, without the page. That is the next proof. Other vendors' assistants are also pending: they need accounts.

## Limitations
- One vendor and one run per model. Repeating runs would measure consistency.
- The answer key was written by the same team, from the page. Any reader can check it against the source.
- The page will change. `fetch_input.py` rebuilds the input and says if its hash no longer matches.

## Reproduce it
```
python3 fetch_input.py > input_text.txt
```
1. Paste `prompt.txt` into the assistant you want to test, replacing `{INPUT_TEXT}` with the content of `input_text.txt`.
2. Save its JSON answer in `resultados/<name>.json`.
3. Score it: `python3 score.py resultados/*.json`.

The page text itself is not redistributed here; it belongs to Google.