Inside Jean

What happens between your question and Jean’s answer.

Every question runs through eight governed stages inside your own environment, from the Governance Gateway to the ledger. Every part it touches is yours.

How Jean answers: every AI answer traced to its source

  1. The finding

    The date, in one line, before the reasons.

    Reconciliation lands on 12 March, not 26 February. The answer opens with the fact and the cause, names the model that produced it, and says that nothing left your perimeter.

    • Sovereign model
    • In perimeter
    • Answered in 1.9s
  2. The reasoning

    The reasoning is a path on your map, not a paragraph.

    Dotti holds the vendor change, the two gates it moved and the quarter cut-off as one graph. Jean shows every hop. Where a word means two things in your company, it says which one it used.

    • Six hops on the graph
    • Your term: cut-over window
  3. What it read

    Every source, open at the line Jean used.

    The migration plan, the vendor thread, the UAT calendar and the Slack channel. Jean reads only what your access allows, and says plainly what stayed out of scope and why.

    • SharePoint
    • Outlook
    • Jira
    • Slack
    • Two rules kept the rest out
  4. What you can do

    Next steps are drafted, logged and waiting for a person.

    Tell finance and risk, or move the gate. Both are written before they run, both sit in the ledger, and the one that changes a date waits for someone in finance to approve it.

    • Gateway PASS, logged
    • One route waits on finance

How your question travels through Jean, stage by stage

When does reconciliation actually land?

Finance / Payments migration

Internal Sovereign model Allowed
  1. Question in
  2. 1. Ask Simon decides
  3. 2. Gateway Policy gate
  4. 3. Recall Automatic
  5. 4. Map Automatic
  6. 5. Route Simon decides
  7. 6. Research Simon decides
  8. 7. Compose Simon + policy
  9. 8. Record Automatic
  10. Answer out

1. Ask it like you would ask a colleague.

Simon decides

Jean already knows who you are, where you are working and what you can see. Simon reads your label and picks the path.

What the question carries with it

When does reconciliation actually land?Asked 10:41
Signed in as M. Farrow, FinanceYour IdP
Working in Payments migrationMate
Cleared for internalLabel
Path: internal, one of fiveSimon
Stage 1 of 8Logged

2. Nothing runs until your policy says so.

Policy gate

A small model in your tenancy classifies the prompt and returns one decision.

The Gateway’s decision on this prompt

Touches payment schedules and vendor datesInternal
Policy: Finance data handlingOwner: Finance
Highest model tier allowedSovereign
DecisionAllow
Stage 2 of 8Logged

3. It recalls what your company already knows.

Automatic

Every fact comes back with where it came from and when.

Facts on file

Migration plan v11 sets the reconciliation gateSharePoint
The vendor moved the SDK buildEmail, 14 February
UAT window is 18 to 22 MarchJira
Quarter close cut-off is 22 MarchFinance calendar
Stage 3 of 8Logged

4. Dotti places the question on your map.

Automatic

People, systems and terms, and how they connect in your company.

Where the question sits on the map

Cut-off in FinanceQuarter close
Cut-off in EngineeringCut-over window
Reconciliation gateDepends on UAT pass
UAT passDepends on the SDK build
Stage 4 of 8Logged

5. The right model, never a bigger one than you allow.

Simon decides

AI model orchestration under your policy: Simon picks the tier your classification allows. Company, department and your own limits can only narrow it, and the reason is written down.

Routing this question

Frontier model, outside the perimeterNot for internal data
Sovereign model, in your perimeterChosen by Simon
Internal small modelRuns the checks
Reason for the choiceIn the ledger
Stage 5 of 8Logged

6. Jean goes and looks, as you.

Simon decides

Simon picks the tools the question needs. Connectors read with your permissions, and anything Jean could not open is named.

The tools Simon chose, and what they read

Jira PAY-412, the SDK buildRead
Outlook, the vendor threadRead
Slack, #payments, 6 MarchRead
Finance restricted folderNot in your scope
The webNot needed
Stage 6 of 8Logged

7. Checked claim by claim before it reaches you.

Simon + policy gate

Simon judges whether each claim rests on what the tools returned, and the Gateway reads the answer on the way out.

The answer, claim by claim, judged by Simon

Reconciliation lands 12 MarchGrounded
The vendor moved the SDK buildGrounded
Two dependent gates moveGrounded
Outbound checkPass, nothing masked
Stage 7 of 8Logged

8. In the ledger before you read it.

Automatic

Every stage writes a hash-chained entry. Anything that acts waits for a person.

Ledger entries for this question

10:41 prompt received, Gateway allowSealed
10:41 routed to the sovereign modelSealed
10:42 answer served, sources attachedSealed
10:43 draft to finance created, unsentAwaiting Amina, Finance
Stage 8 of 8Logged

What you own when you run Jean

Nine parts, one perimeter. Every one runs where you put it and writes to the same ledger.

Your perimeter

Governance Gateway

One decision on every action

ledger hash-chained

  • Scored on the way in and out
  • Your classification sets the ceiling
  • Unreachable means nothing runs

Runs in your tenancy

Dotti

The map of your company

  • Built from your own systems
  • Retrieval by relationship, not wording
  • Every hop shown in the answer

Runs on your servers

Knowledge Base

Every fact, with its source

  • Facts, not passages that looked relevant
  • Corrections stay on the record
  • No source, no fact

Runs on your servers

Synthia

A test company with known answers

ANSWER KEY
  • Whole synthetic companies, generated
  • The answers planted in advance
  • Scored before your data is touched

Runs on your servers

Models you own

Yours inside, frontier on your terms

FrontierSovereignInternalyour ceiling
  • The weights are yours to export
  • Frozen core, versioned adaptive layers
  • Nothing you ask trains anyone else

Runs on your servers

Connectors

Read and act, within limits you set

Microsoft 365SlackSalesforceJiraGoogle DriveSAPGitHubZendeskNotionSnowflake
  • Your permissions, never a service account
  • Per-user scopes on every tool
  • Live events start workflows

Runs on your network

Simon

The decision, without the essay

read onceLicence validYesRisk tierBOwner namedNo
  • Answers from your own options
  • Same case, same answer, every time
  • Fifty checks for the price of one read

Runs on your servers

Hallucination Guard

Every answer read against its evidence

TOOL RESULT
  • Checked against what the tools returned
  • Jean never vouches for Jean
  • No source means a flag

Runs on your servers

Workflows + Specialists

Fixed work, proven decisions

codecodemodelcodeΣSpecialist solveOPTIMALgap 0%
  • Built from a sentence or a chat
  • Solvers prove the plan, or prove none exists
  • Refuses when no skill fits

Runs on your servers

Synthia

Proof that Jean tells you the truth

  • Built in house by our research mathematicians
  • Whole synthetic companies with a verified answer key, so accuracy is measured, not demonstrated
  • Every gap it finds tunes retrieval, prompts and the Gateway’s policy adapters before go-live
  • A wrong answer is traced back to the fact it should have used
  • Tuned before a single real record is connected
Jean
→ MODULE DAG→ SYNTHIA→ THE TWIN
Module DAG
Synthia
The twin
How the twin is generatedcomplete
The digital twin is ready

Every claim traces back to the module that owns it.

29
modules
5,637
claims
11,838
Q&A
Answer accuracy
96%
Path correctness
91%
Questions per run
1,480
Refusals when ungrounded
100%
Dotti

The map of your company

  • The ontology layer: who does what, which system holds which record, and how it all connects
  • Knows what your words mean in each team, and resolves two names for one thing
  • Jean walks the map to understand a question, and shows every hop
  • Carries Synthia’s proven accuracy over into your actual systems
  • The Digital Twin is built on top of it
Jean
The MapDOTTI
202
Entities
5,135
Facts
94%
Verified
29
Systems
10
Sources
Mining live · 4 min ago
PeopleTeamsCustomersSystemsArtefactsConcepts
acct_idrenewal_datecust_refnotice_periodAcme GroupSalesforceSharePointFinance ledgerNotice periodPriya RamanMSA v4.2Delivery pod 3FinanceRenewal dateSOW 14K. MensahNorthwindTradingM. FarrowLegalHarborLogisticsOutlookL. BrennanDana OkaforT. NovakChurn riskOmar HaddadBrightlineRetailCustomerSuccessSofia LindJiraData warehouseMigrationplan v11
CustomerAcme Group

One node, nine spellings across four systems. Resolved once, so one client is one node and never nine rows.

ACME GRPAcme Group LLCacme_grpACM-0042+5 more
Mined fromSalesforceSalesforceSharePointSharePoint
Governance Gateway

Your rules decide before any model does

  • Every request, human or agent, is checked against your policy before any model sees it
  • One of four decisions: allow, ask a named person, re-route in perimeter, or refuse
  • Classification decides the route, not convenience
  • Every decision, reason and model goes to a ledger you own, your one AI audit trail
  • It fails closed: an unavailable control is a failed control

See the Gateway in a security review

Jean
Youmodel.chatrequested Frontier · external

“Send the Acme renewal exposure, with the ledger export, to the frontier model for a board summary.”

Gateway · internal SLM resolved in 38ms
  1. Classify
    customer terms + ledger rows detected
    Confidential
  2. Match policy
    Confidential stays in tenancy · owned by L. Brennan
    P-07
  3. Verdict
    frontier denied for Confidential
    Reroute
  4. Route
    in tenancy, 38ms
    Internal SLM
  5. Ledger
    entry written · hash 9f3c…a71e
    #48,214
Recent verdictsGateway log
Allow
Summarise Acme thread
Internal · ran on Sovereign · EU
2m
Reroute
Draft a customer note
Confidential clause · internal SLM
6m
Approval
Send the renewal notice
Held for Priya Raman
11m
Block
Paste payroll to a public model
Restricted · refused, logged
24m
Checked · 7d
48,213
Blocked · 7d
312
Re-routed · 7d
1,899
Median decision
38ms
Simon

The decision, decoupled from the essay

  • Makes the calls inside every question: which classification path, which model tier, which tools, and whether the answer rests on their evidence
  • Returns the answer itself, picked from the options you defined, with a number for how certain it is
  • Reads the case once and answers every question in the same pass, so fifty checks cost close to what five used to
  • Same case, same question, same answer, reproducible months later against the part of the file that produced it
  • You set the line: above it the answer acts, below it the case goes to a person. Policy, not hope
Jean
Case 4417 · onboarding pack, 38 pages
Read once · 5 questions in one pass · 0 words written
your line 0.80
Trade licence current?
YesNo
0.99 · acts
Sanctions match on a director?
YesNo
0.97 · acts
Bank details match the licence?
YesNo
0.95 · acts
Risk tier
ABC
0.88 · acts
Beneficial owner disclosed?
YesNo
0.61 · to a person
Same case, same questions, same answers. Reproducible from the case hash, each answer shown against the page that produced it.
Hallucination Guard

A second reader checks every answer against its evidence

  • An open-weights groundedness judge, on your own hardware, reads each answer against what the tools actually returned
  • Catches what string-matching misses: the right invoice number with the wrong amount attached
  • Jean’s earlier answers never count as evidence, so a claim cannot vouch for itself
  • Today it scores answers offline in the validation pipeline. In the live path it runs after the last token, so the answer still streams with no wait
  • A flag sits on the sentence and says what is missing, so you can open the source and check it yourself
Jean
What the tools returned
xero.get_invoice
INV-2026-0451
amount £47,300.00
status PAID 2026-09-14
clickup.get_task
86ab1 Chase remittance advice
assignee R. Nandi
due (none set)
Jean’s answer, sentence by sentence
INV-2026-0451 was paid on 14 September.0.01
The amount settled was £47,300.0.02
R. Nandi owns the follow-up, due this Friday.0.94
No due date in the task. May be from the model's own knowledge.
Scored in the validation pipeline, on your own hardware. Nothing sent to a third party.
Workflows

Work your people repeat, done one fixed way

  • Say it, keep it, or let Jean notice: built from a sentence, from one click on work just done in chat, or from work Jean sees repeated
  • Fixed steps run as code. Only the steps that need judgement call a model, so the same input takes the same path
  • Summarising 10 emails into Slack: 13 model calls in chat, 3 as a Workflow, and 4 to 5 times faster
  • Lives in Dotti as how your company works, reusable by anyone with permission, and it stays when people leave
  • Never live on its own. A Workflow Jean built waits for a person, and sensitive steps pass the Governance Gateway
Jean
Renewal watch Active
Run once
no
Contract date minus 45 daystrigger · 41 contracts watched
Fetch contract, invoices, ticketstool_call · ×3SharePointJiracode
Read clause 7, ageing, open P2stransform · ×1code
Governance Gatewaypolicy · 0.1spolicy
Anything unresolved?checkcode
Draft the renewal noticellm_step· asks M. Farrowmodel
Microsoft TeamsPost a 4-line brief to Priyatool_call · ×1code
End 4 of 6 on code · −94%
Specialists

Decisions a language model cannot prove, solved

  • Solvers for the constrained calls run by a senior person with a spreadsheet: capacity, rosters, on-call, pricing, panels
  • Each model is authored once, reviewed and frozen. The AI reads the ask and explains the result, and never touches the maths in between
  • Every answer says what it is: proven best, within a stated gap, or proven impossible with the exact conflicts named
  • A fix that bends a rule you cannot bend is filtered out before it is ever offered
  • When no skill fits, Jean refuses and offers to build one. A wrong answer is worse than no answer
Jean
Bench
Assignments
CP-SAT
Calendar
CP-SAT
Commitmentssoon
MILP · HiGHS
Packingsoon
MILP · HiGHS
Matchingsoon
Min-cost flow
Sequencingsoon
CP-SAT
Networkssoon
Max-flow
Routessoon
OR-Tools routing
OptimalCP-SAT · 3,424 vars · 6,758 constraint rows · search space 10^1,031
1,323
weighted crew strength. No legal roster scores higher
Feasible in 12s, optimised in 120s
dayjobruleassigned
Mon 04Harrow site surveyTwo-personA. Whitfield · M. Osei
Mon 04Croydon commissioningCertifiedR. Nandi
Tue 05Slough retrofitTwo-personM. Osei · J. Bagley
Wed 06Reading fault callCertifiedA. Whitfield

How the Governance Gateway, Dotti and Synthia work together

Synthia proves accuracy, Dotti maps your company, the Gateway governs every call, Simon decides, the Guard checks, Workflows and Specialists do the work. Each makes the others more reliable, and every correct answer makes the next one cheaper, inside your perimeter, on your balance sheet.

See Jean answer a question from your business.

See Jean answer a question from your own systems. If Jean is wrong for you, we will say so in the call.

Book a demo