What "traceable" looks like

The provenance is the product.

Everyone in this trade says "sourced". Almost nobody shows you what that means when a source doesn't hold. Below is a sanitised extract from a real route package, the method that produced it, and the thresholds the grading actually uses — including the ones we fail.

The artefact

A cover verdict that leads with a failure.

Planning safe movement for a business trip to Baghdad. The destination could not be confirmed — which gate admits visitors was unknown — so the package says so on its cover rather than burying it. Every competitor shows you green checkmarks.

AEGILO ROUTE PACKAGE · EXTRACTmeridian · baghdad · 2026-06
Q: Plan movement — airport → hotel → business site, 15–19 Jun.
▸ COVER VERDICT: DOES NOT PASS — endpoint unconfirmed (Dora gate)
ENDPOINT · business site (destination)
"Which gate admits visitors is unknown."
recon:none — generic pin sits deep inside the perimeter
verdict:NOT CONFIRMED — gate assignment needs host confirmation / ground recon
ARRIVAL · airport curb
"Arrivals curb, civil terminal — not the aerodrome centroid."
recon:map recon · gate + barrier visible; ground-confirm on arrival
verdict:CONFIRMED
ROUTE · primary vs alternate
"An alternate that shares the primary's river crossing is not an alternate."
check:shared-choke fraction measured per route
verdict:RETAINED — alternate is independent

Read the failure carefully, because it is the useful part. A generic map pin for the destination sits deep inside a perimeter, and a plan built on it would put a car at the wrong wall of a compound at the wrong hour. The honest output is not a better guess. It is naming the one thing that has to be confirmed by the host before anybody travels.

The method

Five stages, and what each does to a claim.

01 THE QUESTION

Fix what is actually being decided.

Most enquiries arrive as a topic, not a question. Before anything is collected, the decision gets written down in a form that can be answered and can be wrong — because a question that cannot fail cannot be researched either.

02 THE SOURCES

Fetch the document, and keep it.

A cited source is retrieved and stored, so the citation points at a document we hold rather than at a model's recollection of one. A portal that led us to a record is logged as a route to it, never counted as a source in its own right.

03 TRIANGULATION

Count originators, not articles.

Who said a thing is separated from where it appeared, and from the outlet family that carried it. Only independent, named originators count toward corroboration. Four papers running one wire story is one source, and the system records it as one.

04 THE RED TEAM

Argue against it before you see it.

The conclusion is attacked on its own evidence — what would have to be true for it to be wrong, and whether anything in hand already says so. Judgements that do not survive are weakened or dropped rather than quietly rewritten.

05 THE DECISION LOG

Keep what was rejected.

The alternatives considered, the reason each was set aside, and the red-team verdict stay in the deliverable. A judgement you cannot argue with is not a judgement you can rely on.

How a source is counted

Four newspapers running one wire story is one source.

This is where most "multi-source verified" claims quietly fall apart. If you count articles, a single agency report syndicated across a morning looks like overwhelming agreement. We hold the originator — who actually said it — apart from the publication that carried it and from the outlet family behind that, and only independent named originators count toward corroboration. It is a structural rule in the code, not an editorial intention.

CORROBORATED Two or more independent named originators, reporting it separately.
AMPLIFIED One originator, repeated. Volume of coverage, not weight of evidence — the failure mode that makes a rumour look established.
SINGLE-SOURCE One originator, once. Reported as such rather than smoothed into a confident sentence.
CONTRADICTED Sources disagree. The disagreement is the finding, and it ships as one.

Worth saying plainly: where only one originator has been established for a finding, it is labelled single-source — including when we suspect the story is better supported than our record of it shows. The system is built to under-claim corroboration rather than assert agreement it cannot demonstrate, so a finding is sometimes graded weaker than it deserves. That is the direction we would rather fail in, and the label is what tells you it happened.

The thresholds

Where the lines are, so you can disagree with them.

A firm that grades its evidence and won't tell you where the boundaries sit hasn't really told you anything. These are the actual values the grading uses.

STRONG 3+ independent originators, across 2+ types of source
PARTIAL 2+ independent originators
WEAK Everything below that — and it is labelled weak, not rounded up
CONFLICTED Any opposing account, whatever the count on either side

Separately, a headline measurement is not allowed out at all unless source coverage passes its own rubric — enough of the period actually covered, enough locations resolved precisely, no single provider dominating the picture, and at least two named originators present. Fail that and the finding is reported as limited or unmeasured rather than stated as measured.

What sits behind it

A corpus that is versioned, checksummed and dated.

Aegilo maintains its own body of events and claims rather than querying someone else's index at the moment of asking. Every release is stamped, checksummed and kept, so a judgement made in July can be re-examined against exactly the material that produced it — rather than against whatever the web says a year later. Figures below are from the release promoted on 27 July 2026.

397,854 canonical events One real-world event, deduplicated across every source that reported it
457,530 claims Individual assertions, each carried back to who made it
427,003 publications The pieces those claims appeared in
3,463,796 graph edges How events, claims and actors connect to one another

This supplies the surrounding picture — what has happened around a place, an actor or a route, and who reported it. It is not a substitute for the registers, maps and local records a specific question needs, and for a named location those are read per engagement rather than assumed to be pre-loaded.

What we don't claim

The list a vendor usually leaves out.

A page about provenance that overstated would refute itself. So, explicitly: we publish no accuracy score and no track record, because no completed benchmark exists to support one. We do not issue probabilities or predictive risk scores — confidence is stated in words, with the reason. We do not do parcel-level or cadastral geometry, sanctions and PEP screening, live or always-on monitoring, or investigation of individuals beyond what they have consented to. Where a question needs any of those, we say so and make introductions.

And the honest one about this page: everything above describes how the work is produced. It is not a client report. For a sample of the deliverable itself, ask — we send one with the reply.

Common questions

Asked before every engagement.

How does Aegilo count sources?

By independent named originator, not by article. Who said something is held separately from where it appeared and from the outlet family that carried it, and only independent named originators count toward corroboration. Four newspapers reprinting one wire story is one source. Unnamed placeholders and tertiary aggregators can evidence a claim but can never establish corroboration on their own.

What does Aegilo mean by a triangulated finding?

Three or more independent originators across at least two different types of source is graded strong. Two or more independent originators is partial. Anything below that is graded weak and labelled weak. Where sources disagree, the finding is conflicted and the disagreement is reported rather than resolved by preference.

Does Aegilo publish probabilities or risk scores?

No. Confidence is stated in words with the reason attached, never as a manufactured percentage. Nobody reduces a contested future to a number honestly, and a figure invites a precision the evidence does not carry. Where a measurement exists it is reported as a measurement, and it is not the same thing as a forecast.

How do you know an AI didn't invent part of the report?

Because the cited document is fetched and held, so a claim can be checked against the source it came from rather than a paraphrase. Every model call is recorded with the model, the prompt, the response and its hash, and a build test fails if any code path skips that record. The attribution guards can only ever remove or weaken an originator — they are not able to invent one or manufacture agreement between sources.

Ready when you are

Have a question worth answering well?

Start a conversation

The method above is also published weekly, in public, in The Brief — the same source handling, on a live watchlist.