🧠 decision-note.md
1️⃣ The gap I found
My rule was: "Never invent a launch, a partnership or a price."
- 🃏 The first run did not invent anything. But it picked bad posts:
- a price guess ("$CATE … will go to billion");
- a self-named "crypto shiller";
- tiny promo accounts.
- 📅 Then the archive run showed a bigger gap. I asked for posts from July 22. X gave back a Monero post from October 5. The robot trusted X's date filter. So a July video would show October news.
2️⃣ What I changed
- 🧹 In
scripts/collect_x.py:- New rules remove other-coin tickers, market-cap guesses, and "shill/NFA/DYOR" bios.
- New rules remove posts with fewer than 5 likes and accounts with fewer than 500 followers.
- A strict date check: the post's own date must be inside the 24 hours. If not, the robot removes it.
- A rule for the government "DOGE" office.
- Fancy fonts become normal letters.
- Copies are removed.
- 🎙️ In
scripts/render.py: the voice reads fewer words of each post. Now it fits 70 seconds. The card still shows the whole sentence. - 🃏 On every slide: a "What is this?" line, the source link and the like count.
3️⃣ Why
My intent says: the AI must not decide what counts as news. Price hype and wrong-day posts are the mistakes a viewer believes. The date check protects the most basic promise: "the last 24 hours."
4️⃣ What got better
| 📊 Check | Before | After |
|---|---|---|
| Bad cards on October 5 | 4 of 8 | 0 price talk or ads (first-run → revised-run) |
| Wrong-day posts on archive days | trusted X | 45 caught and removed (July 22, August 2, September 30) |
5️⃣ What is not solved yet
- 🔤 Words with 2 meanings. "Lightning Lake" (a fishing trip) still counts as Bitcoin Lightning. The robot matches words, not meaning.
- 📰 Big news can be missed. The biggest news is not always the most-liked post. On July 22, the Wavelength post made it. On September 30, the DogeOS post did not.
- 🧱 X can break the scraper. This happens every few weeks. Then no video gets made that day.
6️⃣ What stayed mine
- 🆓 I chose the free scraper. I know X can limit my account.
- 🗂️ I chose which 3 past days matter. Each one has a real source.
- ✅ I approved the first video before anything went public.
7️⃣ The rule I would encode
⭐ My rule
Never trust a website's date filter. Check every item's own date before you use it.
October 11 audit:
The prior first/revised tables remain historical records. The current published October 5 edition has eight posts, not the seven-post revised table. Codex proposed these additional fixes:
| Gap | Local correction | Reason |
|---|---|---|
| Missed-day run labels today's recent search with an old date | Use historical collection for past edition dates | The edition date must match the data window. |
| Archive renderer returns approval-pending but the backfill still publishes | Stop the archive run with exit code 3 | A pending sample must not reach publication in that run. |
| Feed attachments point to an unavailable GitHub Pages host | Use Vercel URLs for retained videos | The feed must link to the same playable media as the day page. |
| All privacy and scaling features described as layer 2 | Add a classification reference | Protocol upgrades, extension blocks and payment channels differ. |
| Broad STE100 compliance statement | State an inspired approach with compliance unverified | Clear prose alone does not demonstrate full rule and dictionary compliance. |
Six local regression checks pass. These changes are prepared locally. KJ's acceptance, rejection or revision of these proposals is pending.