Omnel by Selione Open the console
For laboratories · clinicians · AI companies · institutions

How faithfully does a machine represent a person? Measured, not declared.

Selione builds a guided archive of one human life and tests, blind, how well a model can stand in for that person. Omnel offers the measurement, the corpus and the instruments behind it to the people who have to trust a number about a human being — under rules that do not bend for a contract.

Request access See the six offers
  • Never a sentence. Counts, proportions and judgments leave; raw text and voice never do.
  • Consent per purpose. Three separate questions to every traveler, each withdrawable.
  • Nothing under the floor. No figure is published on fewer than thirty judged items.
omnel.selione.ai / console / overview
OverviewFloor · 30 judged items
Consenting travelers248research purpose
Judged trials, quarter1 204+ 312 vs last
Open campaigns21 publishable
Abstention rate23%within band
Recognised, by month in the archiveper archive, never averaged across people
M1M3M6M9M12
Journal entries today412 · written before each answer
Withdrawals honoured3 · the same day
Raw text exported0 · no such export exists
2 290

questions on a scale fixed, and anchored in a public calendar, before the first measurement.

3

numbers and no average: recognised, abstained, confidently wrong — each with its own floor or ceiling.

30

judged items per archive before anything is published; below that, no number at all.

0

sentences in anything that leaves: judgments, counts and proportions only, in groups of twenty or more.

What we measure

The Standard: a scale fixed before the first measurement.

In astronomy a distance is never measured directly: a star of known brightness is compared with how bright it looks. The Standard is that object for representations of people. Someone who knows the person is shown two answers to the same question, theirs and the model's, without being told which is which. Three proportions come out, and a floor under which nothing is published.

i

Recognised

How often the person who knows them picks the person's own answer over the model's. Reported per archive; never averaged across people.

Target band · above 70%
ii

Abstained

How often the model said it did not know. A high rate is the instrument working; a rate below its floor is an alarm, because a model that answers everything is inventing somewhere.

Floor · at least 5%
iii

Confidently wrong

Answers given with no hedge, and wrong: the one error that counts when the subject is a human being, because the person who loves them cannot catch it. Counted on its own; above its ceiling the result is not publishable, whatever the score says.

Ceiling · under 5%
Recognised
71%
Abstained
23%
Confidently wrong
6%
Illustrative values. The marker on each track is the band's floor, target or ceiling. What the Standard does not guarantee is written on the same page as the Standard: it measures recognisability on the questions tested, not identity; a score earned on one archive says nothing about the next; it cannot see what was never asked.

Read the whole Standard, with its proof

Six offers

What Omnel provides, and what each one requires.

Every offer below shows the console screen it comes with. The console is live; the screens are illustrations of it, and every figure shown is an illustration, not a measurement.

01 · For AI companies

Evaluation against the Standard

Have your model of a person judged the way ours is: blind, by the people who know that person, with the same three numbers and the same floor. The result is published about the system, never about a person.

You receive
A campaign report: the three bands per archive and overall, the publishable flag, the protocol as anchored before the campaign.
Delivered as
A console campaign with a signed PDF; trials run on our infrastructure, your model called through your endpoint.
Requires
Travelers who said yes to the "judge other models" purpose and are paid for their time; at least five archives; thirty judged items each.
console / campaigns / K-7
Campaign · model K-7 · 150 blind trials · 5 archivesPublishable
71%recognised
23%abstained
6%confidently wrong
ArchiveTrialsRecog.Abst.Conf. wrongStatus
A-013277%18%5%publishable
A-023065%30%5%publishable
A-033171%19%10%at ceiling
A-0424———under floor
A-053374%22%4%publishable
Protocol anchored2026-09-30 · public calendar
Reported aboutthe system · never a person
02 · For research and model builders

The verdict corpus

Humans judging how faithfully a machine represented them, with a label no other corpus carries: true, but not to them — the disclosure a person refuses even when the fact is right. Offered as judgments, never as sentences.

You receive
Labelled judgments with their context class and confidence; counts by label, by purpose and by cohort; a withdrawal feed.
Delivered as
An annual licence per purpose, read through the console or the API; the export has no text field, by construction.
Requires
A thousand lines consented for the "teach" purpose before the first licence; a contract that forbids re-identification; withdrawals honoured the same day.
console / corpus / licence 2027-Q1
Licence · purpose "teach" · 2027-Q1Judgments only · no sentence
True61%
True, but not to them7%
Invented19%
Abstained rightly13%
Lines in scope4 110consented "teach"
Withdrawn this month9deleted the same day
Cohort countries14shown above 20 people
Export fieldslabel · class · confidence · month · cohort
Text fieldnone exists
03 · For products that must not invent

The judge that says "I don't know"

A decision model that scores an answer against a known archive: a calibrated confidence, “not calibrated yet” said plainly until enough judged decisions exist, and abstention as a first-class answer. The first level is built and in production.

You receive
A verdict with its confidence and the count of supporting passages, or an abstention; a monthly calibration report.
Delivered as
An API scoped to the archives that consented; keys stored as fingerprints; the journal written before the answer.
Requires
Your use case stated in writing and bound to the consented classes; until enough judged decisions exist, the report says “not calibrated yet”.
console / judge
Judge · keys & calls · 30 daysJournal before answer
Active keys3fingerprints only
Calls18 42030 days
Abstentions23%4 191
Calibrationnot yetsaid plainly
Reliabilitypredicted vs observed
{
  "verdict": "abstain",
  "confidence": 0.18,
  "support": 0,
  "reason": "no passage in the consented classes",
  "journal": "2026-10-06T21:14:02Z"
}
04 · For laboratories

Instruments for studies

The instruments of the Standard, read over months. Aggregate access on our infrastructure over consenting travelers; protocols before results; null results reported.

You receive
Cohort-level series over months: fidelity, abstention, calibration, consistency; preregistered add-ons scored by us.
Delivered as
A study workspace in the console; aggregates computed on our side; groups of twenty or more; no raw text ever leaves.
Requires
Your own ethics approval for add-ons; a protocol anchored before the first reading; co-applications welcome.
console / studies / cohort-2027
Study · consistency over months · cohort buildern = 37 · above the floor
Consent · researchMonths in archive · 6+Age band · 55–74Country · hidden under 20
HorizonPairsConsistencyCalibration (Brier)Anchored
30 days2120.820.142026-09-30
180 days960.750.172026-09-30
365 days———opens 2027-10
Preregistrationanchored · hash 7f3a…c2
Null resultsreported in the same series
05 · For organisations

Tacit — what your best people know

When an expert leaves, the procedures stay and the judgment goes. Tacit asks three to ten of them a few questions a day over ninety days, dictated or written, and turns what they say into cards. The expert sees every card first; only what they publish reaches the library. An echo answers colleagues from published cards only, and abstains otherwise; a monthly certificate is measured in the Register. It is a feature of the Omnel console, under the same rule as everything else here: the person decides what leaves the room.

You receive
A library of published cards, searchable by situation; a pilot report on coverage and gaps.
Delivered as
A three-month pilot for five experts, run in the console; the library stays yours, exportable in a documented format.
Requires
Experts who volunteer; a sponsor who accepts that managers see only what the expert published.

Tacit, in detail — the expert's screen, the library, the certificate

console / tacit / maintenance-site-b
Pilot · Maintenance, site B · month 2 of 3Experts decide what leaves the room
Experts7volunteers
Days answered61 / 90a few minutes each
Cards written184seen by the expert first
Published13153 kept private
Cards published131
Kept private53
Colleagues' questions answered from cards212
Abstained · no published card61
Visible to managers or HRonly the 131 published
Exportdocumented format · yours
06 · For partner systems

Ask the archive

A server any system can connect to. A question returns one bounded passage with its provenance — what was said, when, with what confidence — or an abstention that says the person never said it and we refuse to guess.

You receive
One passage per question with its date and confidence, or an abstention; a journal of every question asked of your connection.
Delivered as
A connection per partner system, through an open protocol; consent per class of question, per connection, revocable by the traveler.
Requires
A partner with a legitimate relationship to the traveler — care, family, legal — and the traveler's own yes to each class.
console / connections
Connections · 2 activeConsent per class · revocable
ConnectionWorkPlacesHealthMoneyQuestions, 30 d
Care system · ward 4yesyesnono1 204
Family appyesyesnono318
question received · class "places"
consent checked · granted by the traveler on 2026-08-14
journal written
answer · one passage · 2019-05 · confidence 0.81
question received · class "health" · refused · no consent
Tacit · the fifth offer, in detail

What your best people know, and never wrote down.

When an expert retires or leaves, the procedures stay and the judgment goes: the early signal they noticed, the exception they knew was safe, the mistake they had learned never to make. Tacit asks them a few questions a day over ninety days, in their own words, and turns what they say into cards a successor can ask for. Colleagues ask an echo that answers from published cards only and abstains otherwise; every month a certificate is measured in the Register. It runs inside the Omnel console, and nothing an expert says is visible to anyone until they publish it.

  1. 1

    You name the people

    A pilot covers three to ten experts whose departure would hurt: approaching retirement, single points of failure, the ones everyone calls.

  2. 2

    They answer, at their pace

    A few questions a day over ninety days, dictated or written, whenever they have a few minutes.

  3. 3

    Their words become cards

    What they say becomes cards, keeping their numbers, thresholds and tools. Nothing is added that they did not say. An echo answers colleagues from the published cards only, and abstains when none covers the question.

  4. 4

    They decide what leaves the room

    An expert sees every card first. Only what they publish reaches the organisation's library. Unpublished answers are never visible to their manager, to HR, or to us.

tacit / expert / day 41 of 90
Today’s question · day 41 of 90Only you see this screen

One question a day, about the work you actually do. Dictate it or write it; a few minutes is enough.

Your answer · dictated · 2 min 40

…the pump sounded a half-tone higher, days before the pressure drifted. The gauge said fine. I had the seal changed anyway, and the next time I waited for the gauge we lost a shift…

Card · from your words

A half-tone rise in the pump's pitch precedes a pressure drift by two to four days, before the gauge moves.

Publish to the libraryKeep private
Card · from your words

When the sound changes and the gauge does not, trust the sound: change the seal at the next stop.

Publish to the libraryKeep private
Written from your wordsnothing added that you did not say
Seen by your manager or HRonly what you publish
console / tacit / library
Library · Maintenance, site B131 published · 53 kept private
Ask the echo“the pressure drifts but the gauge reads normal”
Published · 131Experts · 7Answered from cards · 212Abstained · 61
CardFromPublishedCited by the echo
A half-tone rise in pitch precedes a drift by two to four daysExpert 3day 4118 times
Never reset the alarm history before the night shift has read itExpert 1day 129 times
Supplier B's “urgent” means Thursday; plan the week around itExpert 5day 276 times
The Friday checklist exists because of the 2019 flood; keep itExpert 2day 634 times
The echo answers frompublished cards only · abstains otherwise
Exportdocumented format · yours
Models trained on your contentnone, ever

What Tacit is for

Tacit is for the knowledge that leaves with the person: it keeps what an expert actually knows, in their own words, where a successor can ask for it.

Request a Tacit pilot Expert sign-in
A pilot includes
SeatsThree to ten experts, ninety days, in the Omnel console.
LibraryEvery published card, searchable by situation and by expert; exportable in a documented format that opens without us.
EchoAnswers colleagues from the published cards only, and abstains when no card covers the question.
CertificateMeasured every month in the Register, with its denominator; nothing is published under the floor.
ConfidentialityStorage separate from any consumer product; an answer is private to the expert until they publish the card drawn from it; no model is trained on your content.
ReportCoverage and gaps, as counts — never a sentence an expert kept private.
PricePrices are given in writing, on request; invoiced by the company.
The console

Counts, never a word. Every figure with its denominator.

What a partner sees: the instruments, the corpus, the keys, the pilots and the consents — as numbers with their floors, never as anything a person said. The console is live: an organisation requests access, is seated, opens readings, makes keys, and reads the ledger — open it, or read the public register of readings. The three views below are illustrations of it; the figures in them are not measurements.

omnel.selione.ai / console
Consents · the ledgerOff by default · each a separate yes
Research cohort248travelers
Corpus · improve / teach / judge211 · 133 · 58separate questions
Withdrawals31honoured the same day
PurposeAskedYesWithdrawnEffect of a withdrawal
Improve their own Echo4122116lines deleted the same day
Teach a model for others41213314deleted · trained models cannot unlearn, and we say so
Judge other companies' models4125811removed from every future campaign
Keys & billingKeys exist only as fingerprints
KeyScopeCalls, 30 dLast used
sha256:9f3a…e1judge · classes work, places12 980todayactive
sha256:41c0…77corpus · licence 2027-Q1443 days agoactive
sha256:b812…0dconnection · family app318todayrevoked by traveler
Invoicesissued by the company · by transfer
Usage above planstops the key, never bills silently
GovernancePublished · reviewed yearly
Impact assessmentpublished · next review 2027-09
EU representativenamed before the first European traveler
Ethics committeeindependent · required before any clinical publication
Raw text leaving the infrastructure0 · no export path exists
Smallest group reported20
Governance

The rules are the product. They do not bend for a contract.

i

Never a sentence

What leaves is a judgment, a count, a proportion. No raw text, no "anonymised" text, no voice. Groups of twenty or more.

ii

Consent per purpose, off by default

Three separate questions to each traveler, each withdrawable. A withdrawal deletes the lines the same day; a model already trained cannot have them taken back out, and we say so.

iii

Protocols before results

What will be measured is written and anchored before it is measured. Null results are reported. Nothing is published under the floor.

iv

Pilots first, prices in writing

Every engagement starts small enough to be stopped. We would rather lose a contract than a promise made to the people in the archive.

Kind of dataLeaves?Under what condition
A traveler's words, voice, photographsneverNo export path exists for a partner. The traveler alone exports their own archive.
A human verdict on a machine's answeryesAs a labelled judgment, under the purpose the traveler said yes to, withdrawable.
Counts and proportionsyesGroups of twenty or more; smaller groups are shown as "under 20".
One bounded passage, to a connected systemby consentOnly for the classes the traveler opened to that connection; journal written before the answer.
An expert's Tacit cardsby the expertOnly the cards the expert published; private cards never reach the library.
How an engagement runs

Four steps, each one small enough to stop.

  1. 1

    A conversation

    You tell us what you would measure, build or license, and what you would need to see first. We answer within a few days, and we say no when the rules say no.

  2. 2

    A protocol, anchored

    What will be measured, on whom, with which floor, is written and anchored in a public calendar before anything runs. You keep a copy.

  3. 3

    A pilot

    A campaign, a study month, a licence quarter or a Tacit pilot for five experts — priced in writing beforehand, invoiced by the company.

  4. 4

    A report, and a decision

    The figures with their denominators, the null results included. Then you decide whether to continue; so do the travelers.

Prices are given in writing, on request, after the first conversation. Pilots are priced to be stopped without regret; licences are annual and per purpose. Consent, not technique, sets the pace: nothing is sold on travelers who have not said yes.

Research

Protocols before results, and partners before publications.

Laboratories

Psychology of memory and ageing, bereavement, human–computer interaction, bioethics. Long, consented, month-by-month records of how well a model of one person predicts that person, sealed and dated before the fact.

Clinicians

Reminiscence and life review, grief. Instruments that respect the floors you would set yourselves, and a product that says plainly where it stops being an archive.

AI companies

Anyone building a representation of a person who wants a number they did not produce themselves, and a public scale to be measured against.

Institutions

Standard-setting bodies, regulators and funders who need a measurement that was defined before the first result, and an archive format that opens without us.

The research protocol Research partners Research data

Request access

A person reads it, and writes back.

Write to us with who you are, what you would measure or build, and what you would need to see first. We answer within a few days.

selioneapp@gmail.com Read the Standard first

What to include

  • Your institution and the person who will own the engagement.
  • Which offer, or which question, you are coming with.
  • Your timeline, and whether you need an ethics approval of your own.
  • What a good first result would look like to you.

We do not run demonstrations on real archives. Everything shown on this page is an illustration of the console; the first real figures you see will be yours, under a protocol you anchored.