Methodology

Content production & verification

SKYTICAL is an aviation information platform built around trusted sources, fast updates and full traceability.

The system continuously tracks public information released by civil aviation authorities, international organizations, airlines and trusted news outlets, organizes events through an automated data pipeline, and then uses widely recognized language-model families such as Anthropic Claude, OpenAI GPT and Google Gemini to help produce the Traditional Chinese and English reports.

The models are not a news source and never invent information. Every article must be built on source material the system actually obtained, with the original references kept at the end so readers can verify directly.

---

Our core principles

Sources before models

SKYTICAL reporting starts from trusted sources, not from a language model.

The system first obtains official announcements, press releases, RSS feeds, open APIs or trusted media reports, and only then hands the collected material to a model. The model may use only the source content given to it — no outside knowledge, no unsupported inference.

Every key fact is traceable

Core facts in a story must map back to the original material. The system keeps the source URL, the publication time, and the original wording that supports the key statements.

Every article ends with a Sources box, so readers can open the original announcement or report and check for themselves.

Automation is not unlimited generation

Content must pass format, source, date, quotation and risk checks before it can enter the publishing flow.

Content that is insufficient, conflicting or unverifiable is rejected. Accident, flight-safety, legal and other high-risk stories must pass the same, stricter source and evidence checks; passing stories publish directly with their risk flags retained. The system never forces out a story just to keep the update count up.

---

Automated data pipeline

GitHub Actions starts the pipeline on schedule, checking RSS feeds, open APIs and official pages of the FAA, ICAO, IATA, airlines, accident investigators and other trusted sources.

The system strips navigation, ads, duplicated text and dead pages, and checks publication time, body completeness and source validity.

Only material with enough substance and a traceable source moves on.

URL normalization and title-text similarity keep the same piece from being processed twice.

When several sources cover one event, the system identifies what they share and where they differ, then builds a single news event.

Qualified material goes to a multi-model failover system built around widely recognized model families including Anthropic Claude, OpenAI GPT and Google Gemini.

The models are responsible for:

• Extracting the core facts from the sources
• Organizing event timelines and development status
• Rendering source information into Traditional Chinese and English
• Producing the headline and summary
• Tagging the supporting source for each key statement

The system switches to a backup model only when the primary has a technical problem. A conservative call from a model is not treated as a failure, and the system never re-asks other models to get a more aggressive answer.

Everything a model returns must pass code-level checks, including:

• JSON and data-structure validation
• Date and time consistency
• Verbatim quote matching against the source
• Statement-to-source mapping
• Duplicate-event checks
• Risk flags and publish-eligibility rules

Only content that meets the publishing bar reaches the site.

Validated stories are published to GitHub Pages, keeping:

• The original source names
• Links to the original articles or announcements
• The source publication time
• The time this site compiled the story
• An automated-production label

The original sources always prevail over SKYTICAL's compilation.

---

How models take part

SKYTICAL uses language models as tools for organizing information and generating text — never as a source of facts.

Models work only from the source material the system provides. They may not search the web, and may not fill gaps with background knowledge that does not appear in the sources.

For example, if a source only says an aircraft arrived at an airport, the model may not speculate that the flight was:

• A charter mission
• A diversion
• A repositioning flight
• A maintenance flight
• A cargo mission
• A diplomatic movement
• Special passenger transport

unless the sources explicitly support that description.

Words like “confirmed”, “approved”, “delivered” or “entered service” must match the level of certainty expressed by the original source.

Beta exception: some entries on the thrice-daily briefing pages are compiled with a search-capable model - the only place model retrieval is allowed, under hard limits: cited links must come from actually retrieved results, social/video platforms are excluded, and forum sources need a companion source. Such entries are tagged "AI search" and the briefing carries a BETA mark; regular articles are never covered by this exception.

---

Models and providers

SKYTICAL runs a multi-model architecture to limit the impact of any single model or vendor failing.

The main model families and providers disclosed publicly are:

• Claude (Anthropic)
• GPT (OpenAI)
• Gemini (Google)

Every model is bound by the same rules and must pass the same quotation and data checks. Switching models never lowers the publishing bar.

Exact model versions, technical routing and fallback order may change with availability, quality testing and service terms. The Methodology lists only the major model families familiar to general readers; the source-verification rules and publishing bar stay the same regardless of internal routing.

---

Source selection

SKYTICAL prefers first-hand sources, or sources with clear editorial accountability, such as:

• Civil aviation authorities
• Accident investigation agencies
• International aviation organizations
• Official airport and airline announcements
• Airframe and engine manufacturers
• Trusted news organizations
• Open data services with clear licensing terms

Social media posts, reposts, or material whose original source cannot be confirmed are never the sole basis for a key fact in an automatically published story.

---

Update cadence

As a rule, the main news pipeline runs once every hour.

Actual publication timing can be affected by:

• When the original sources update
• Temporary site or API outages
• Data completeness checks
• Cross-source deduplication
• Model service availability
• Human review of high-risk content

A scheduled run therefore does not guarantee a new article every hour. When there is no sufficient, verifiable new material, the system does not generate content to fill space.

---

Transparency & open source

SKYTICAL's core code is public on GitHub; readers can inspect the fetching, deduplication, model calls, validation and publishing flow.

avwire/
├─ .github/workflows/       # schedules & automation
├─ pipeline/
│  ├─ fetch.py              # RSS / API source fetching
│  ├─ dedupe.py             # dedup & event merging
│  ├─ providers.py          # multi-model failover & calls
│  ├─ write.py              # model compilation, quote & data checks
│  └─ build.py              # site generation & publishing
├─ data/                    # pipeline state & public data
└─ site/                    # GitHub Pages site

The actual layout may evolve as the system develops; the GitHub repository is authoritative.

To protect the service, API keys, administrative credentials and other secrets are never published in the source code or the site frontend.

---

Corrections & limitations

Automation increases speed and coverage, but cannot guarantee zero errors.

Factors that can affect the output include:

• The original source issuing a correction
• Incomplete RSS or open-API data
• Inconsistent publication-time or timezone labeling
• A language model misreading the content
• Sources describing the same event differently
• Coverage or update-latency limits in live data

When a problem is found, SKYTICAL may update, retract or flag the content. For anything authoritative, defer to the regulator, airline, airport or other original publisher.