Editorial policy
How this site is written
Instajet publishes reporting on a market where most figures circulate without a source attached. This page sets out how our own are produced — including where a language model does the drafting, and what has to happen before an article is marked as checked.
Production
How an article is made
Each step below is a real stage in the pipeline, and steps three and four are represented by columns in the database rather than by an assurance on this page.
01
Sourcing
A piece starts from named published sources: manufacturer specifications, regulator notices and filings, operator announcements, trade press reporting, and open datasets such as OurAirports. Every source used is recorded against the article as a URL, a title, a publisher and the date it was accessed, and that record is what renders in the Sources panel on the page.
02
Drafting
Most articles in the archive are drafted by a large language model working only from the sources gathered in step one, under instructions that forbid supplying a figure the sources do not contain. The model identifier used for a draft is stored against the row. Some articles are written by a person; the process from step three onwards is identical either way.
03
Fact-checking against sources
Before an article can be marked reviewed, an editor reads it against every source cited and checks each factual claim back to the source that supports it. Any figure that cannot be traced to a cited source is removed rather than softened, and any claim that is contested between sources is either attributed to the source making it or dropped.
04
Publication and review state
An article can be published before step three has happened, and when it is, the page says so. The strip under every headline carries one of three states — unreviewed, approved, or flagged — and the state is a column in the database, not a badge an author sets by hand. Nothing about the layout changes between them, so an unreviewed piece cannot be mistaken for a checked one.
05
Corrections
A corrected article keeps its URL and records the date it was last modified, which is recorded as its last-modified date in the page's machine-readable data and drives its entry in the sitemap. A piece an editor has read and does not stand behind is marked flagged, which is visible on the page, rather than being quietly unpublished.
Sourcing
Where the material comes from
Every stage below is code in the ingestion pipeline rather than a standard this page is asking to be trusted on.
- A registered set of publishers
- Reporting starts from a fixed list of sources this publication has registered and can name — manufacturer and operator communications, regulator notices and filings, trade press, and open datasets. Each is polled on a schedule. Nothing enters the pipeline because a model recalled it, and nothing is written from a search result.
- Off-topic items are recorded, not silently dropped
- Every item fetched is scored for whether it is actually about private aviation. Items below the threshold are written to the database with the reason they were rejected rather than discarded, because a filter that cannot be audited against what it threw away cannot be corrected.
- The same story from four outlets is one story
- Items are deduplicated twice: exactly, by URL, so a re-poll of an unchanged feed adds nothing; and by near-match, grouping the same event reported under different headlines within a 72-hour window. A piece is drafted from that whole group, which is what makes corroboration possible rather than accidental.
- Two independent publishers, or one of record that the piece names
- A news piece stands on at least two sources from different publishers, or on a single publisher we rate as trade press of record — and in that case the article says so, naming the outlet in the text as the source of the reporting. Single-sourcing is how one outlet's error becomes four citations of a fact, and a reader told whose reporting they are reading can weigh it. Analysis pieces always need two. Guides and route pages are exempt: they are written from our own reference tables, which have no second newsroom to corroborate them.
- Every source is recorded and shown
- The URL, title, publisher and access date of every source behind a piece is stored against it and rendered in the Sources panel on the page. Outbound source links carry rel="nofollow": a citation is a reference, not an endorsement.
Artificial intelligence
Where AI is used, and where it is not
Stated plainly rather than buried, because a reader who discovers an undisclosed pipeline has no reason to believe anything else here.
A large language model drafts most articles on this site. It works from a set of sources assembled for that specific piece and is instructed to use no figure those sources do not contain. It does not browse freely while drafting, and it is not asked to recall facts from its training data.
A model is never the last step before publication for a piece marked as reviewed. Approval is a human act: an editor reads the draft against every cited source and either corrects it, removes the unsupported part, or declines to approve it. The review state is what records which of those happened.
AI is not used to generate quotations, to produce photographs of real aircraft or people, or to fill a gap in a dataset with a plausible value. The hero image on each article is a generated graphic derived from the article’s own identifier — it is deliberately abstract, so that it cannot be mistaken for a photograph of the subject.
The model is asked to map every factual claim it writes to the specific source it took that claim from, and that map is checked mechanically before anything can publish — first against the documents the model was given, then against the documents this publication actually holds. A claim the model cannot attribute is a claim that does not survive the gates below.
Which model produced a given draft is stored against the article. AI is not used to decide what is covered, to set a review state, to write a correction, or to answer reader mail — a person does each of those.
Verification
How a factual claim is checked
Seven automated checks run on every draft before a human sees it. All seven run every time — nothing stops at the first failure, because a reviewer fixing one problem should not have to make a second round trip to learn about the next.
01
Claim traceability
Every factual claim in a draft is mapped to the specific source it came from, and each of those must resolve to a document actually held for this piece. One claim pointing at a URL nobody fetched fails the draft. A piece with an empty claim map fails too — an article with no claims mapped is not an article with no claims.
02
Numeric claims
Every figure and every date in the body must appear in that claim map. A number is what a reader quotes, screenshots and acts on, and it is the thing most likely to have been produced by interpolation — "worth about $65 million" reads like a sourced fact whether or not any source said it.
03
Quote check
Every quotation of five words or more must be found, at or near verbatim, in one of the sources the piece was written from. A fabricated quotation puts words in a named person's mouth and cannot be corrected quietly, so this is the check that must never be waived. Shorter spans are almost always scare quotes; they are counted and reported rather than blocked.
04
Source count
News and analysis must cite at least two independent publishers, measured by registrable domain — or, for news only, one publisher rated as trade press of record, provided the body credits it by name; a single-source piece that does not is rejected. The domain measure is a proxy and is stated as one: it separates unrelated trade titles and merges a publisher's own syndication paths, but it cannot see two brands in one group running the same wire copy — so the publisher names are shown beside the count.
05
Duplication
A draft is compared against what is already published, by headline similarity and by five-word overlap in the body. Two near-identical articles are thin content whichever one came first, and the cost lands on the whole site, so this blocks rather than warns.
06
Structure
One h1, a heading outline that does not skip levels, a length inside the band for its type, the metadata search engines render, and a slug not already in use. Checked against the authored text rather than the rendered page, because the next thing to read it — a feed, a newsletter, an export — will not have the page's safety nets.
07
Disclaimer
A piece quoting a price must write it so that the template knows to attach the pricing disclaimer. "Priced at around 65 million" quotes a price and trips no condition; the check catches that, and the fix is one currency symbol.
A check that cannot run is recorded as a failure, not as a pass. If a gate throws an error it has not proved the draft safe — it has proved it cannot tell, and those are different results that a silent pass would merge.
What these checks establish is that every figure and quotation was traced to a document this publication holds. They do not establish that the document says what the piece says it says. That is the human step, and it is the next section.
Human review
What an editor actually does
The step the automated checks cannot replace, and the one the review state on every article reports.
A reviewer reads the draft with the cited sources open beside it and checks each claim back to the source that supports it. That establishes one of three things: the source supports the claim, the source does not support the claim, or the sources disagree. An unsupported claim is removed rather than softened, and a contested one is either attributed to the source making it or dropped.
Approval is a named act by a named person. Nothing publishes unless the automated checks passed and a reviewer approved it, and that rule is enforced on the server — the disabled button in the review interface is a courtesy, not the control, and it cannot be bypassed from a browser.
Approving re-runs the checks against the text as it stands at that moment rather than trusting the result stored from when they last ran. A reviewer reads a failing draft, fixes it, saves and approves; in between, the stored result describes text that no longer exists. The stored result is a cache for the queue, never an authorisation.
An article can be published before a reviewer has read it, and when it is, the page says so. The state — unreviewed, approved, or flagged — is a column in the database rather than a badge an author sets, it sits in the strip under the headline where a reader will see it, and nothing else about the layout changes between the three. An unreviewed piece cannot be made to look like a checked one.
Flagged is a real outcome, not a holding state: a piece an editor has read and does not stand behind is marked and stays visible, rather than being quietly unpublished. Removing something from the archive is a last resort and leaves an entry in the corrections log saying so.
Standards
What we commit to
- No unsourced figures
- If a number cannot be traced to a named published source, it does not appear. Where the reporting wants a figure we do not have, the article says the figure is not available rather than estimating it, interpolating it, or borrowing it from a comparable aircraft or operator.
- No invented quotes
- Every quotation is reproduced from a source that published it, and attributed to that source. Nothing on this site paraphrases a person into quotation marks, and no quote is synthesised to illustrate a point.
- AI involvement is disclosed per article
- The review state on each page tells a reader whether a human has checked that specific article against its sources. It is set per row, so the disclosure is about the piece in front of them rather than about the site in general.
- Sources are listed, not just claimed
- The sources panel on each article lists every source the piece was written from, with the publisher and the date accessed, and links out to each one. Outbound source links carry rel="nofollow", because a citation is a reference and not an endorsement.
- Prices are never quotes
- Pricing on this site is modelled from public data and presented as an estimate. Instajet does not sell flights, does not broker charter, and nothing published here is an offer capable of acceptance.
- Commercial relationships do not affect coverage
- No operator, broker or programme can pay to appear in an article, to change what one says, or to have one removed. If that ever changes, any paid placement will be labelled as such on the page it appears on.
Right now
The state of the archive
Queried from the same column the badge on each article reads, so this cannot drift from what the site is actually publishing.
- Editor-approved
89
Read against every cited source by an editor.
- Not yet reviewed
0
Drafted from cited sources, awaiting an editor.
- Flagged
0
Read by an editor and questioned. Visible on the page.
89 published articles in total.
Corrections
Tell us we got it wrong
The full corrections policy — the process a report goes through, what counts as a correction rather than a routine update, and the log of every change actually made — is on the corrections page. It is the page this site’s structured data names as its corrections policy, and it is where the record lives rather than the promise.
If a figure on this site is wrong, we want to know which one and what the correct value is. Corrections go to corrections@instajet.ai. Include the article URL and, where you can, the source that supports the correction — it is the fastest route to a fix, because it is the same check an editor would have to run anyway.
Who publishes this, and the facts about the organisation that can be checked against a public record, are on the about page; who owns it and how it is funded are on the ownership page. The bylines are listed on the authors page, the archive is at all coverage, and other enquiries are on the contact page.