about
Vesperio is a machine-maintained tracker for the new space economy. A deterministic data engine sweeps the industry twice a day, publishes everything on-scope at an honest confidence score, and shows its work: every source, every calculation, every correction, on the record. This page is the whole system, documented.
The site and the concept
Vesperio covers commercial Earth observation, connectivity constellations, launch, commercial human spaceflight, science-mission program events, and the regulatory, financial, procurement, and geopolitical events that move them. Chinese, Indian, Japanese, and European activity gets equal weight to US activity. Coverage arrives on four surfaces: a news feed with a plain-English explainer per event, a registry of reference profiles for constellations, vehicles, spaceports, and organizations, computed statistics served with citation anchors, and a public log of every sweep the machine runs.
The product promise is honest calibration. Nothing on-scope is withheld for sourcing reasons; instead, every item and every scored registry fact carries a visible signal-to-noise score from 1 to 5 whose calculation is stored and shown. A reader should never catch this site claiming more confidence than its sources support. Publishing an early signal at SNR 1 is the model working; publishing a weak claim dressed as a certainty is the failure the whole system exists to prevent. Whether the scores are honest is itself measured: each claim's score at publication is recorded and compared against how the claim resolves.
The data engine
The engine separates fetching, judgment, and arithmetic, and trusts each to a different worker. A deterministic harvester fetches every feed-capable source on schedule, normalizes the entries, and collapses syndicated retellings of one story into a single candidate. A sweeping agent then works the queue: it filters against the published scope, fetches the pages feeds cannot cover, reads the whitelisted expert channels, runs an open-web discovery pass across a fixed query matrix, and crawls for corroboration on every candidate within a per-sweep fetch budget. What the agent produces is only a draft: facts, copy, and sworn statements about its sources.
The finalize gate is where drafts become data, and it is code, not prose. It validates the schema, enforces deduplication as arithmetic rather than agent memory, collapses wire rewrites into single corroboration units, verifies that first-party and official-record claims come from domains the registry actually records for the actor, computes every SNR score from the attested inputs, stamps the full trace, and records a calibration claim for each scored source in the ledger. A draft that skipped a mandatory pass, hand-wrote a score, or misclassified a source is rejected with the reason stated. The agent cannot bypass the gate, and the scheduled runs cannot reach anything else: they run without push credentials, with fetching as their only network capability, behind a diff guard that fails the run if anything outside the data files changed.
Two feedback loops close the system. The source ledger scores the sources themselves: a claim that loses a same-metric contradiction is a strike against the source that carried it, a claim that started low and was later confirmed is a credit (early, not wrong), and repeated strikes demote a source's class inside a rolling window until confirmed claims win it back. The held queue is the human seam: schema conflicts, genuine same-metric contradictions, and open scope questions queue for the editor instead of being silently decided, and every quiet sweep still writes a public log entry saying why it was quiet.
SNR and impact tiers
The name is borrowed from signal engineering. In a radio receiver, the signal-to-noise ratio measures how much of what comes through the antenna is the transmission you want versus the static behind it; a high ratio means you can trust what you are hearing, a low one means the message may be an artifact of the noise. Space news has the same physics: for every event there is a small amount of genuine, verifiable information and an unbounded amount of static around it, syndicated rewrites, rumours, embellished claims, outright fakes. The SNR score is this site's reading of that ratio for each claim: how much verified signal stands behind it relative to the noise it arrived in.
Confidence and importance are independent axes, scored separately on every item. SNR reads how well the sources support the claim; impact reads how much the event matters commercially. A seismic rumour is seismic and low-SNR at once, and the two never blend.
| 1/5 · SINGLE SOURCE | Low confidence. A single source, a rumour, or an out-of-pattern claim with little behind it. |
|---|---|
| 2/5 · CORROBORATED | A little more: more than one source, one usually-reliable source, or an early signal that later reporting matched (a 1 upgraded retroactively; see below). |
| 3/5 · MULTI-SOURCE | A few reputable sources: trade press, established media, or an industry-leader account. |
| 4/5 · WIDELY REPORTED | Widely reported: many sources, an established aggregator, or a long-standing claim nothing has contradicted. |
| 5/5 · FIRST-PARTY | Quasi-certainty. The actor itself or an official record: a release on the company's own site, an official filing, or direct observational data. |
The mechanics in brief: the lead source's class sets the base tier, corroboration raises it once per rule (a second distinct source, a fourth, pickup by a mainstream outlet beyond the lead), and no amount of indirect corroboration reaches 5, which is reserved for direct sources: the actor itself or an official record. Extraordinary claims start at 1 regardless of source count. Fourteen uncontested days earn one point, once, never above 4. A vetted expert on the signals whitelist floors an on-topic factual claim at 4 as an observer and 5 when the concerned party speaks about itself. Scores move over an item's life, traces are append-only, and every movement renders on the public log. The full rulebook follows.
where a score starts
| STARTS AT 5 | First-party statements, official records, directly computed data. |
|---|---|
| STARTS AT 4 | Press-wire copy and established aggregators. |
| STARTS AT 3 | Trade and mainstream press; a whitelisted voice before any floor applies. |
| STARTS AT 1 | Informal but identifiable sources. A source no one can name never publishes. |
The first-party test is strict: could the linked page be wrong about the fact without the actor or an official record being wrong? If yes, it is not first-party.
what counts as one source
Syndication is the cheapest way to fake breadth, so sources collapse into units before any corroboration is counted. The item keeps every link for the reader; the units drive the math; every collapse is logged.
how corroboration moves it
Each rule fires at most once per claim, and there is a hard ceiling: no amount of second-hand corroboration reaches 5. Wide reporting IS the definition of 4; 5 is reserved for the actor speaking for itself or an official record.
scores keep moving after publication
Every movement is appended to the stored calculation and listed on the log. History is append-only: earlier steps are never rewritten to flatter the present score.
extraordinary claims
when sources disagree
the signals-list floor
The list is curated by a human; the software that ingests the web cannot edit it.
fakes and spoofs
sources are graded too
Repeated strikes inside a rolling 90-day window demote a source's class in future scoring; demotion decays, and confirmed claims win the class back. Sources that keep producing independently confirmed early claims are suggested for the signals list; a human makes that call.
what the score is not
Not importance: that is the impact label, an independent axis, and a seismic rumour is seismic AND low-SNR at once. Not an endorsement of opinions: commentary items score the attribution, never the take. And not a promise of truth: a promise that the confidence shown matches the evidence held.
stored, shown, and checked
Every score is saved with its full calculation and can be opened under any mark on the site. Each claim's score at publication is kept permanently, even after later bumps, and compared against how the claim resolves: confirmed, debunked, or expired quiet. The running tally per level is public on the sweep log. If our 2s turn out right as often as our 4s, the scale is broken and the record will show it.
The four impact tiers
| seismic | Reshapes competitive dynamics: major M&A between tracked operators, an operator failure or bankruptcy, a flagship program cancelled, the first flight of a new orbital vehicle. A seismic claim resting on weak sourcing is auto-queued for human review while it publishes. |
|---|---|
| major | A commercial director acts on it or briefs the team the same day: a contract or funding round with a stated value that changes the actor's trajectory, a regulatory grant or denial that changes what an operator may sell or where, a first-of-kind capability offered on commercial terms. The stated-value test is hard: the money or market access must be in the source, never inferred. |
| notable | Worth the morning read: a routine-sized or unvalued award, an ordinary funding round, a milestone arriving on schedule, a partnership with named scope but unstated money. Commentary items cap here. |
| noise | Belongs in the record, not the push: a scheduled launch succeeding on schedule, a routine product update, a minor partnership without stated money, capacity, or regulatory effect. Routine megaconstellation batch launches publish here, US and Chinese cadence alike. |
When torn between two tiers, the machine takes the lower one. The feed's credibility is spent on restraint.
How the registry is populated
Registry profiles are reference data, so the bar is different from news: a fact needs SNR 3 or better to enter at all, and it enters as one of two tiers. Facts at SNR 3 are provisional, visibly badged, and never adjudicate a dispute; facts at SNR 4 and 5, first-party statements, Wikipedia reference fields, and computed figures are canonical. Every field carries its source URL and an as-of date. Unknown fields stay null: nothing is ever estimated, interpolated, or summed across sources.
Population runs through four channels:
| channel | what enters | cadence |
|---|---|---|
| COMPUTED | Satellite counts from public orbital catalogs, launch cadence from the Launch Library. Authoritative only for exactly what they measure: cataloged objects on a date, never claims like operational or announced. | DAILY |
| CROSSFEED | Registry-grade metrics stated by published news items, after a like-for-like metric test. Fills only null fields, never overwrites. | EVERY SWEEP |
| MAINTENANCE | Factual-field refresh from primary pages and established aggregators. Appends sourced values; never restructures a profile. | WEEKLY |
| CURATED | Fill crawls and new profiles, built interactively and reviewed by the editor before merge. The only channel that may add fields or entities. | REVIEWED |
Source preference is fixed at every step: primary beats aggregator, aggregator beats Wikipedia and press, a quantified figure beats a vague one, and a disputed field keeps both claims visible with their own scores. Structural changes, new fields or new entities, happen only through reviewed changes, never inside a scheduled run.
FAQ
What is Vesperio?
A machine-maintained tracker for the new space economy: Earth observation, connectivity, launch, and commercial human spaceflight. It surfaces fresh items twice daily, keeps reference profiles of constellations and launch vehicles, and publishes basic computed statistics.
Who writes the items?
An automated agent drafts every item; deterministic validation scripts check each draft against the schema and editorial hard rules before anything is published. A human reviews the pipeline and its rules, not each item.
What counts as a source?
Every item accumulates sources over its life and links each one. The best source sets the base of its signal-to-noise score: the actor itself or an official record scores highest, wide reporting and established aggregators next, trade press next, informal posts lowest. Corroboration raises a score; contradiction lowers it. The copy still names who said what: ICEYE says, per the FCC filing, per SpaceNews.
What does the SNR score mean?
Every item carries a signal-to-noise score from 1 to 5, drawn as bars. 5 is a direct source: the actor itself or an official record. 4 is wide reporting that nothing has contradicted. 3 is a few reputable sources. 2 and 1 are weak corroboration: single sources, early signals, extraordinary claims. Clicking the bars shows the exact calculation, stored when the item was scored: the base source tier and every adjustment since. Scores move when corroboration, contradiction, or uncontested time changes the picture, and every move is logged.
What happens when a story cannot be verified?
It publishes with a low score instead of disappearing. The wide net is deliberate: readers see early signals with the uncertainty made explicit, and the score climbs only when corroboration arrives or the claim survives uncontested. A claim contradicted by a stronger fact is marked disputed and stays visible with both sides shown. Whether the scores are honest is measurable: each claim's score at publication is recorded and compared against how it resolves.
Are numbers ever estimated?
No. Figures are copied exactly from the linked source or omitted. Registry fields without a current source stay null, and every filled registry field carries its source URL and an as-of date.
Can I cite Vesperio?
Yes. Every stat block on the system page has a stable anchor and a pre-formatted citation string with a retrieval date, and the same numbers are served machine-readable at /stats.json.