Lesson 4 · 18 min
base.md is the heart of the channel with sources. Every number that appears on screen goes in with who said it, the video, the date, and a quote copied from the transcript without edits. A script checks each quote; if one isn’t there, the video is discarded. Here you’ll build your database and checker.
Start the lesson
Goal: build base.md for a video in your niche, with an exact quote on every row, and use a script to check each quote against the transcript.
Click each block.
The check doesn’t judge whether the number is true in the real world. It confirms that someone said exactly that in the cited video. That’s why everything is attributed ("according to him").
Illustrative examples based on an invented transcript for the real estate niche.
Copy directly from the transcript, exactly as written, even if the speech is informal. Use 6 to 30 words, enough to prove the number.
"the average rent here in the area went up twelve percent in one year"
Changing words to make them sound better breaks the verification. The script can’t find them, and the video fails.
"the average rent in the area increased 12% over the last year"
Joining the beginning of one sentence to the end of another creates a sentence no one said.
"average rent went up… and will keep rising in two thousand twenty-seven"
Two or three words can appear anywhere and don’t show the context. Use at least 6.
"twelve percent"
Run it in the agent’s folder. Replace <URL> with the selected video.
yt-dlp -x --audio-format mp3 -o "audio.%(ext)s" "<URL>" # with inemavox (or another local transcriber): python transcrever_v1.py --in audio.mp3 --outdir tx --whisper-model large-v3
| Number used | Who said it | Video | Date | Transcript excerpt | |---|---|---|---|---| | | | <id do YouTube> | dd/mm/aaaa | "<6 a 30 palavras copiadas>" |
#!/usr/bin/env python3
"""usage: python3 confere.py base.md transcript.txt
Is each quoted excerpt from the source in the transcript? Exits 1 if any are missing."""
import re, sys, unicodedata
def n(s): # lowercase, no accents, no punctuation
s = unicodedata.normalize("NFD", s.lower()).encode("ascii", "ignore").decode()
return re.sub(r"\s+", " ", re.sub(r"[^a-z0-9%]+", " ", s)).strip()
base = open(sys.argv[1], encoding="utf-8").read()
tx = n(open(sys.argv[2], encoding="utf-8").read())
fora = 0
for trecho in re.findall(r'"([^"]{20,})"', base):
ok = n(trecho) in tx
fora += not ok
print("OK " if ok else "FORA", trecho[:70])
sys.exit(1 if fora else 0)OK average rent here in the area went up twelve percent in one year FORA average rent in the area increased 12% over the last year
One line is FORA: fix it by copying it again from the transcript. The radars’ checar.py does the same for every video cited.
The transcript and the source can differ in accents, commas, or capitalization without changing the meaning. Normalizing both sides avoids false failures and still catches changed words or stitched-together sentences.
Each line in the source goes through the checker. Interrupt it halfway and run it again.
Illustrative browser simulation. The failing line is intentional: its excerpt was “improved” and is no longer verbatim.
Exercise 3 in sequence 4.
Transcribe a video from one of your 5 channels and build base.md with the number, who said it, video, date, and exact quote. Run confere.py.
Done when: each line has the exact quote in quotation marks, and confere.py prints only OK.
Full exercise: jornada-agentes/exercicios/04-redes.md.
The explanation appears after you answer.
Question 1 of 3
You changed "twelve percent" to "12%" in the quote in the source file to make it cleaner. What happens during verification?competency
Question 2 of 3
What does verification guarantee?
Question 3 of 3
Which quote works for the source file?
Choose an option. The explanation appears after your attempt.