From the architect
Architect's Note
How TPA Picks Its Ten Points, and How Far You Should Trust Them
I built TPA, which means I wrote the rules that decide what reaches you. What a story must prove before it qualifies. How the ten points get ordered for you specifically. What happens on the days the standard leaves me with four. You should know all of it, because the only sound reason to trust a briefing is that you understand how it was assembled.
What happens, step by step
- 1
I collect the sources.
Every run, TPA reads the standing RSS and Atom feeds for your chosen area. For crypto, that means CoinDesk, Reuters and their peers. On top of that sit wire searches built from your configurator answers: your watchlist, your region, your industry. Anything older than 72 hours is discarded, duplicates are stripped, and a typical run leaves me with around 180 raw items to work from.
- 2
The two-source rule decides what survives.
This is the strongest guarantee in the product. A story earns a place only when two different outlets have independently reported the same event. The AI proposes the stories. My code then re-checks every source it claims against the articles actually fetched. A source that fails to match a real URL gets dropped. Any story left with fewer than two distinct outlets gets deleted, even when that takes the edition below ten points. Your crypto edition ran with four points for exactly this reason. I chose the honest number.
- 3
I rank what survives.
The ranking model orders the surviving stories 1 to 10 by what is most consequential for you, drawing on your configurator answers, your chosen depth, and your must-include and exclude themes. Read this ordering for what it is: an informed editorial opinion, shaped by your brief and carrying all the character of judgment.
- 4
The code holds the hard lines.
Ten stories maximum. Headlines capped at twelve words. Word counts inside set bands. A valid category on every story, and only whitelisted columns. URLs and outlet names come from the ingested articles themselves, with outlet names overwritten by the real ones, which closes the door on invented sources.
My honest confidence rating
The facts stated are real events
High (9/10)
Two-outlet corroboration plus verbatim source linking. Invented sources and URLs are blocked at the structural level. The residual risk is two outlets repeating the same upstream error.
The sources cited actually support the story
High
Verified in code against the fetched items, independently of what the model asserts.
The selection is the ten biggest stories
Moderate (7/10)
Bounded by what the feeds carried in 72 hours. A story covered by a single outlet is excluded by design.
The 1 to 10 ordering matches your priorities
Moderate (6 to 7/10)
The model reads your brief well, though ranking nuance stays subjective. Two human editors would disagree with each other too.
Completeness (always finding ten)
Deliberately low
Quiet news days produce shorter editions. I built it that way. A short edition carries information; filler carries none.
"TPA buys verifiability with volume. Four bulletproof points beat ten shaky ones every time."
The facts are strong, the ranking is a well-briefed editor's call, and the more precisely you answer the configurator, the sharper that call gets.
Vasudevan Kidambi
Editorial Architect, TPA
I designed the rules TPA uses to admit a story and to rank it for you. If a point in any edition fails the standard I have described here, tell me and I will publish the correction in the next one: contact@navoinc.com.