A claim, its evidence.
The target is to trace the story back to its sources, rather than asking the audience to trust only the voice it hears.
THE TARGET ARCHITECTURE
From source to release. An entirely AI cycle.
In the model we are building, no person chooses the headlines, rewrites the script or approves an individual edition. Agents with separate roles research, report, challenge and decide whether to publish. AI also performs the final review.
TARGET ARCHITECTURE · IN TESTING0
human interventions in the editorial cycle
Selection, script, review and clearance of each edition: in the target model, the AI system decides.
SEPARATE ROLES. ONE CYCLE.
How an edition takes shape.
Open a step to see the agent’s task and what it hands to the next. This illustrates the target, not a live run.
Find the storiesSources, dates, events.
At the scheduled time, start collecting from feeds, bulletins and pages through permitted access. Keep origin and timing, group duplicates and distinguish a new event from a recirculated story. Trends suggest questions; they do not prove facts.
HAND-OFFCandidate events with provenance and dates.
Trace the evidenceDocuments, corroboration, uncertainty.
Run targeted searches, open sources and seek separate corroboration. Link each claim to available evidence and flag contradictions, missing information and reports that depend on the same original source.
HAND-OFFA claim dossier: sources, corroboration and unresolved points.
Choose what mattersRelevance, new developments, two editions.
Assess public impact, recency, evidence quality and variety of topics. Compare earlier editions to avoid repetition. Build Italia and World for their respective audiences within the planned duration.
HAND-OFFTwo reasoned rundowns, with priorities and time for each story.
Write to be understoodFacts, context, conversational tone.
Write an original script, distinguishing facts, statements and analysis. Preserve attribution, state uncertainty and prepare tone and emotion cues appropriate to each passage, without turning a serious story into spectacle.
HAND-OFFA script with references, timing and expressive direction for voice and avatar.
Challenge the accountThe script doesn’t clear itself.
Re-examine claims and sources, look for significant omissions and challenge misleading or partisan wording. Request more research, require a rewrite or reject a story. In the target model, human approval does not replace this verdict.
HAND-OFFA reasoned result: clearance, revision or block.
Give it voice and presenceVoice, mood, avatar and editing.
Prepare pronunciation, pacing, voice and captions. Associate moods with spoken passages and synchronise expressions, body and gestures. Compose the studio with readable headlines in landscape and portrait formats.
HAND-OFF16:9 and 9:16 videos, audio, captions and release materials.
Check what you will seeCheck the finished result too.
Assess the finished package: alignment with the cleared script, audio, images, captions, readability and references. Technical checks support AI judgement. If content changes after clearance, the package must go through review again.
HAND-OFFClearance tied to the exact version of the files to distribute.
Schedule the releasePublication, confirmation, archive.
Follow the schedule, adapt descriptions and formats to configured channels, and send only the cleared version. Verify publication confirmation, avoid duplicates and retry temporary errors within defined limits. Update the site only with confirmed releases.
HAND-OFFActual release links, per-channel status and a public edition page.
PLANNED DESTINATIONS
- YouTubeFull editions and clips
- SpotifyAudio and video editions
- InstagramVertical Reels
- TikTokVertical videos
Target channels to activate and test. On the website: video, transcript, sources, release status and corrections. A submission is not a release until the channel confirms it.
Needs revision? Go back.
The reviewer sends issues back to the research, writing or production agent. The result is checked again, with a limited number of attempts.
Unresolved? The edition stops.
If evidence is insufficient or checks keep failing, the system blocks release. It does not seek human approval to bypass the check.
09 / AFTER RELEASE
A story has a memory.
The target includes comparison with what Daily has already reported: new developments, updated sources and contradictions. An update or correction goes through AI research and review again, then remains visible on the edition page.
↳ That context feeds research for the next edition.WHY BUILD IT THIS WAY
Autonomy should leave a trail.
A story, its history.
Daily aims to distinguish developments from repetition: continuity between editions and recognisable corrections.
Not publishing is a decision too.
A schedule should not force gaps to be filled. The system must be able to drop a story or an edition that does not pass its checks.
Two AIs do not guarantee truth.
They can share mistakes, biases or incomplete sources. That is why the project focuses on checkable references, attribution, explicit release blocks and visible corrections. There is no promise of absolute impartiality.
A story can change.
A correction must be checked by a separate AI before publication, with a visible date and explanation. Audience reports are leads to investigate, not direct edits to content. You can already report issues with the site or introduction.
Report an issue ↗The scope of autonomy.
The goal is zero human interventions in the editorial production of individual editions. People build and maintain the system, define its general rules and remain responsible for it. In the target model, they are not a selection, rewriting or approval step in a release.
THE GOAL AND THE PRESENT
What we are building.
The project introduction.
The launch video, the project story and this architecture. Creative pilots help develop the programme’s voice, avatar and visual language.
An edition without human hand-offs.
The entire recurring cycle, from discovery to confirmed releases and corrections. Fully autonomous daily production is not yet announced as operational.