● Public sector — Prison registry
AI agent supporting the prison registry
An assistant that prepares the documents, supports the keeping of the registers and the tracking of deadlines — the registry checks and the authority decides. On local inference, with no external call at all for the most sensitive information: strictly documentary support, every sentence calculation stays checked and settled by the registry.
Local inference possible
Criminal data isolated
GDPR & AI Act: governed deployment
The registry checks, the authority decides
Updated on
Deployed in a few weeks
Registry support · hosted in France
Prepare the statement of situations reaching a deadline this month: sentence end dates and documents to produce, from the register.
Documentary statement prepared. 11 situations reaching a deadline listed from the register: sentence end dates, documents to produce and notifications to prepare, each line referring back to the record and the original documents. No calculation is made by the agent: the dates shown are those settled by the registry.
Processed on local inference — nothing left the network.
⛓ Source · your registers and records, processed locally
Processed on local inference — nothing left the network.
⛓ Source · your registers and records, processed locally
Prepare the notification and the schedule of documents for file no. 2417, from the information checked by the registry.
Documents prepared. Notification drafted on your template with the approved information from the file, schedule of documents generated and the contents list updated. One document missing flagged (receipt of the previous notification).
The registry checks every entry before issue — nothing is notified automatically.
✎ Action · documents ready for checking — the registry approves, the authority decides
The registry checks every entry before issue — nothing is notified automatically.
✎ Action · documents ready for checking — the registry approves, the authority decides
Local inference · no data outside the EU
Local inference possible
Sovereign by designLocal inference or hosting in France
GDPR & AI Act: governed deploymentTraceability & human oversight
TurnkeyDesigned, installed and operated for you
The registry decidesStrictly documentary support, never the calculation
✦ In brief
For a prison registry, a Blue Lemon Agent agent provides strictly documentary support: administrative documents and letters prepared from your templates, keeping the registers, tracking deadlines (sentence end dates, extracts, notifications) and file summaries to make checking easier. Sentence calculations remain entirely checked and settled by the registry; decisions are a matter for the judicial authority. For data among the most sensitive the state holds, the agent runs on local inference — no external call, nothing leaves the establishment's network — or on a isolated resource hosted in France, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity. Live within a few weeks.
100%
hosted in France in the target architecture
0
transfer outside the EU in the target architecture
7
uses ready to deploy on this scope
0
decision taken without human approval
Reference points describing our offer, not results measured at a client. The scale of the gain is confirmed by a pilot on your own scope.
The context
Why AI matters to prison registries — and why they hesitate
The prison registry guarantees the lawfulness of detention: registers, documents, deadlines, notifications. The documentary load is considerable, error is not permitted, and the data handled — criminal situations — is among the most sensitive there is.
! The issue
Registry officers are caught between a growing documentary volume — documents to produce, registers to keep, deadlines to track without fail — and a requirement of absolute rigour, because every date bears on the lawfulness of detention. Yet most consumer AI tools would amount to entrusting criminal situations, registers and court documents to a third party, often hosted outside Europe and subject to the Cloud Act — which is simply unthinkable.
✓ Our answer
For a registry, AI is only of interest if it is sovereign, confined and checkable by design. Local inference preferred — no external call, nothing leaves the establishment's network — systematic human oversight, and a strict scope: documentary support only, never an autonomous sentence calculation, never a decision. The aim is not to stand in for the registry, but to give it back time for checking, which remains its first duty.
The decisive point
Protecting criminal data: sovereignty & compliance
Criminal situations, registers, court documents: no data demands more precautions. Here is how the architecture of our agents protects it, establishment by establishment.
Local inference preferred
The agent runs on a machine at the establishment: no external call, no data leaves the network, nothing passes through a cloud.
Hosting in France
Failing that, a strictly dedicated and isolated resource, hosted in France under French law — no data outside the EU.
Reduced extraterritorial exposure
Exposure of criminal data to the Cloud Act and FISA 702 is reduced by design; location alone does not guarantee immunity.
One isolated resource per establishment
No pooling of data: an environment strictly dedicated to your establishment, partitioned from end to end.
Encryption & controlled access
Encryption in transit and at rest, role-based access, strong authentication and logging of every access.
AI Act: governed deployment
Strictly documentary support; no autonomous sentence calculation, no automatic notification; traceability from end to end.
What depends on the architecture chosen These points are not general guarantees: they are settled deployment by deployment, in the quotation.
- The applicable location is that of the architecture set out in the quotation and verified before commissioning.
- Local execution is announced only for the configuration explicitly described and accepted in the quotation.
- The applicable isolation depends on the deployment mode set out in the quotation; no dedicated isolation is presumed.
- The encryption mechanisms in transit and at rest, their components and key management are those documented for the architecture chosen.
- Roles and permissions are configured and accepted for the identities and systems actually connected.
- The events logged, their content, their retention period and who may access them are defined for the deployment chosen.
For a prison registry, we prefer local inference: the model runs inside the establishment, with no external call at all. For hosted deployments, SecNumCloud and HDS options are available according to how demanding your requirements are. A single architecture is designed to answer both the GDPR and extraterritorial exposure. Designed for deployment in line with the GDPR and the AI Act, after the processing, roles and context-specific risks have been assessed.
Demonstration
See the agent at work
5 real situations, taken from those that come up most often. Pick one: the exchange unfolds as it would in your organisation.
A scripted demonstration. These exchanges show how the agent behaves — its sources, its refusals, what it leaves to your teams. Nothing is sent from this page, no model is queried here, and the matters named are fictional. That is precisely what we promise your data.
The behaviours shown here — monitoring, automation rules, routing and reminders — are configured with you during deployment, from your tools, your rules and your thresholds.
The architecture points named in these exchanges — location, local execution, isolation, encryption, role-based access, logging — are not a guarantee attached to the demonstration: they are those of the architecture set out in your quotation, and verified before commissioning.
The public body in this demonstration
Fictional public bodyVilleneuve-l'Argent correctional facility — prison registry (fictional establishment)
- Service
- Prison registry — receiving and filing enforcement documents, keeping the registers, preparing notifications, tracking deadlines
- Headcount
- 6 registry staff, including the head of registry and 2 officers assigned to deadline tracking — open six days a week
- Perimeter served
- 640 people detained on the 1st of the month; four different courts send documents to the establishment
- Volume
- 2,300 documents received a month; 11 situations reaching a deadline in an average month; 7 use cases ready to deploy on this perimeter
- Tools in place
- Detainee management application, registry registers, in-house body of notes and circulars — the agent plugs into them under local inference, nothing is replaced
- Who decides
- The head of registry settles the dates and signs off every document before issue; the judicial authority decides; the facility director notifies
- Room for improvement
- 3 h 10 to prepare a set of documents and its notification; 14 hours a month for the deadline count; 41 missing documents discovered late last year
The registry of the Villeneuve-l'Argent correctional facility is not trying to decide faster — it has no such power — but to spend less time searching, re-keying and checking by hand, and more on control, which is its primary duty. The agent runs under local inference on a machine inside the establishment: no external call, nothing leaves the network. It plugs into the management application, the registers and the in-house document collection. It prepares and flags; the head of registry settles the dates and signs off; the judicial authority decides. The exchanges below span one quarter, from the first deadline statement to the review.
This public body, its figures and the exchanges that follow were invented for the demonstration. They illustrate a common situation; they describe no real service.
Registry support · hosted in France
The statement of situations reaching a deadline this month is ready: 11 situations listed from the register, every line pointing back to the case file and to the source documents. And I have cross-checked 14 divergences that I am raising with you, without correcting a single one.
What the statement carries, line by line: the situation, the date the registry has settled, the document that establishes it, the documents to be produced, the notifications to prepare, and the number of days left before the deadline.
What grounds every date, and it must be said first: every date on this statement is one the registry has settled, copied across with the reference of the document it comes from. Calculating a sentence is an act of the registry; it carries the name of the officer who settles it — that is what makes it capable of being signed off and challenged. A statement carrying a date calculated by me would be a statement nobody could sign off.
What I do instead, and it is the heart of the work: I recount. Where the same item appears on two documents — a decision, a case sheet, a register — I recount the gap between them and flag any divergence. 14 flags this month, 4 of them on situations falling due within thirty days. A flag is not a finding: it is a request for verification addressed to a registry officer, and it arrives with the two documents side by side and the gap counted, day for day.
The time this shifts: the monthly deadline count drops from 100 % to 20 % of the time it took — 14 hours become 2 h 48. Your two officers assigned to deadlines no longer re-key, they verify — the exact reverse of the previous split.
Processed under local inference: the statement was produced on the establishment's own machine. No external call, nothing left the network.
What I propose: that the head of registry signs off the statement and rules on the 14 flags in order of proximity to the deadline. The 4 nearest first; the other 10 can wait a week, and I say so precisely so that you do not spend the morning on them. deadline-statement_current-month.pdf11 situations, every date tied to its document
⛓ Sourced · registry registers, case sheets and source documents — processed under local inference
What the statement carries, line by line: the situation, the date the registry has settled, the document that establishes it, the documents to be produced, the notifications to prepare, and the number of days left before the deadline.
What grounds every date, and it must be said first: every date on this statement is one the registry has settled, copied across with the reference of the document it comes from. Calculating a sentence is an act of the registry; it carries the name of the officer who settles it — that is what makes it capable of being signed off and challenged. A statement carrying a date calculated by me would be a statement nobody could sign off.
What I do instead, and it is the heart of the work: I recount. Where the same item appears on two documents — a decision, a case sheet, a register — I recount the gap between them and flag any divergence. 14 flags this month, 4 of them on situations falling due within thirty days. A flag is not a finding: it is a request for verification addressed to a registry officer, and it arrives with the two documents side by side and the gap counted, day for day.
The time this shifts: the monthly deadline count drops from 100 % to 20 % of the time it took — 14 hours become 2 h 48. Your two officers assigned to deadlines no longer re-key, they verify — the exact reverse of the previous split.
Processed under local inference: the statement was produced on the establishment's own machine. No external call, nothing left the network.
What I propose: that the head of registry signs off the statement and rules on the 14 flags in order of proximity to the deadline. The 4 nearest first; the other 10 can wait a week, and I say so precisely so that you do not spend the morning on them. deadline-statement_current-month.pdf11 situations, every date tied to its document
⛓ Sourced · registry registers, case sheets and source documents — processed under local inference
These 14 divergences — how do you find them if you calculate nothing?
I compare written records with one another. The distinction matters: I am not trying to work out which date is correct, I am looking for the places where two documents do not say the same thing.
The four families of cross-check I run on every case file:
· Same item, two sources: a date appears on a decision received and on the case sheet. If the two do not match, I flag both, each with its reference. 7 flags this month.
· Gap between two dates in the same file: I recount the number of days between them and compare it with what the document itself states. Where the document says « four months » and the two dates recorded are 122 days apart, I flag it — without saying which is right. 3 flags.
· Document announced and absent: a decision refers to a document that is not on file. 3 flags.
· Duplicate: the same document recorded twice under two titles. 1 flag.
The figure that does not flatter me, and I give it before being asked: of the 14 flags, 3 were false alarms. The cause is identified: documents from an older generation of forms carry the date in abbreviated form, and I was reading the abbreviation as an incomplete date. What I did with it: the reading rule is corrected and the three documents have been re-read.
And the setting I leave to you: I would rather have three false alarms than one missed divergence, and that is a choice you can tighten. The threshold is yours: if you want fewer flags, I raise the threshold and accept missing more. That decision is not mine to take for you.
What I propose: keeping the wide setting for three months, long enough to measure how many real divergences it catches. This quarter: 11 real divergences out of 14 flags, 2 of them on situations under thirty days. Those are what justify the three false alarms. control-recount_14-flags.pdf11 real divergences, 3 false alarms and their cause
⛓ Sourced · source documents of the 11 situations, register, flag log
The four families of cross-check I run on every case file:
· Same item, two sources: a date appears on a decision received and on the case sheet. If the two do not match, I flag both, each with its reference. 7 flags this month.
· Gap between two dates in the same file: I recount the number of days between them and compare it with what the document itself states. Where the document says « four months » and the two dates recorded are 122 days apart, I flag it — without saying which is right. 3 flags.
· Document announced and absent: a decision refers to a document that is not on file. 3 flags.
· Duplicate: the same document recorded twice under two titles. 1 flag.
The figure that does not flatter me, and I give it before being asked: of the 14 flags, 3 were false alarms. The cause is identified: documents from an older generation of forms carry the date in abbreviated form, and I was reading the abbreviation as an incomplete date. What I did with it: the reading rule is corrected and the three documents have been re-read.
And the setting I leave to you: I would rather have three false alarms than one missed divergence, and that is a choice you can tighten. The threshold is yours: if you want fewer flags, I raise the threshold and accept missing more. That decision is not mine to take for you.
What I propose: keeping the wide setting for three months, long enough to measure how many real divergences it catches. This quarter: 11 real divergences out of 14 flags, 2 of them on situations under thirty days. Those are what justify the three false alarms. control-recount_14-flags.pdf11 real divergences, 3 false alarms and their cause
⛓ Sourced · source documents of the 11 situations, register, flag log
And if a registry officer asks you outright to settle a date? It will happen.
I do not settle it, and I do not simply say so: I hand the registry everything needed to settle it in minutes rather than half a day.
What I supply when a date is in question:
· Every document in the file where that date appears, in order of receipt, with the exact wording copied word for word — never reworded.
· The gaps recounted between them, expressed in days, with no conclusion.
· The documents that are missing before the question can be settled, and the court each must be requested from.
· The applicable in-house notes and circulars, cited with their internal reference and date, and the exact passage reproduced — not my summary.
Why I stop there, and it is not timidity: a date in a criminal situation bears on the lawfulness of the detention. It is settled by the registry, under the oversight and decisions of the judicial authority. A tool that produced that date would create a second source, and the day the two diverged nobody would know which one prevails. It is the one place where having two answers is worse than having none.
What that changes in practice, and it is measured: the time between « there is a doubt » and « the registry has everything it needs to rule » has gone from 3 h 40 to 20 minutes. Of the 11 real divergences this quarter, 9 were resolved the same day, 2 required a document from a court.
And the trail, which counts as much as the speed: every flag carries its date, the document that triggered it, the officer who cleared it and the action taken. The day a situation is re-examined, the history exists and it is dated.
What I propose: settling with you who signs off what — because four officers cleared flags this month and two of them did not hold the delegated authority to do so. That is not a fault, it is a missing table. I have drafted it; it remains to be approved. who-settles-a-date-who-signs-off-who-notifies.pdfDelegation table — 2 clearances outside delegation found
✎ Framework · documents of the files concerned, in-house notes, log of cleared flags
What I supply when a date is in question:
· Every document in the file where that date appears, in order of receipt, with the exact wording copied word for word — never reworded.
· The gaps recounted between them, expressed in days, with no conclusion.
· The documents that are missing before the question can be settled, and the court each must be requested from.
· The applicable in-house notes and circulars, cited with their internal reference and date, and the exact passage reproduced — not my summary.
Why I stop there, and it is not timidity: a date in a criminal situation bears on the lawfulness of the detention. It is settled by the registry, under the oversight and decisions of the judicial authority. A tool that produced that date would create a second source, and the day the two diverged nobody would know which one prevails. It is the one place where having two answers is worse than having none.
What that changes in practice, and it is measured: the time between « there is a doubt » and « the registry has everything it needs to rule » has gone from 3 h 40 to 20 minutes. Of the 11 real divergences this quarter, 9 were resolved the same day, 2 required a document from a court.
And the trail, which counts as much as the speed: every flag carries its date, the document that triggered it, the officer who cleared it and the action taken. The day a situation is re-examined, the history exists and it is dated.
What I propose: settling with you who signs off what — because four officers cleared flags this month and two of them did not hold the delegated authority to do so. That is not a fault, it is a missing table. I have drafted it; it remains to be approved. who-settles-a-date-who-signs-off-who-notifies.pdfDelegation table — 2 clearances outside delegation found
✎ Framework · documents of the files concerned, in-house notes, log of cleared flags
Prepare the notification and the schedule of documents for file no. 2417, from the items checked by the registry.
The documents are ready for sign-off: notification drafted on your template using only the validated items of the file, schedule of documents generated, contents page updated. And one missing document that I am flagging before issue.
What is ready:
· The notification, on the establishment's current template, with wording taken from the validated documents — each entry carries in the margin the reference of the document it comes from.
· The schedule of documents: 17 documents on file, numbered, dated, with their origin and their date of receipt.
· The contents page, updated.
The missing document, and it is the one that matters: the acknowledgement of the previous notification is not on file. It is referred to by document no. 11 and was never recorded. Without it the chain of notifications is broken — and that is the kind of gap that surfaces at the worst moment, namely when a date has to be justified.
What I have prepared to close that gap: the request to the originating court, drafted, referenced, ready to go — it awaits the head of registry's mandate, not my initiative.
The time this shifts: preparing a set of documents and its notification drops from 60 % to 10 % of the time it took — 3 h 10 becomes 32 minutes. Across your 45 sets a month, 118 hours returned to the registry every month.
What does not change, and is not negotiable: the registry checks every entry before issue, and nothing is notified automatically. A notification is a formal act: it goes out under the signature of whoever has the power to send it, never otherwise.
What I propose: sign off the schedule first, then the notification. A signed-off schedule is what lets you say, later, what was on file that day — and it is the document you will be asked for first. file-2417_notification-for-sign-off.pdfEvery entry tied to its document, 1 missing document
⛓ Sourced · file no. 2417, documents validated by the registry, establishment templates — processed under local inference
What is ready:
· The notification, on the establishment's current template, with wording taken from the validated documents — each entry carries in the margin the reference of the document it comes from.
· The schedule of documents: 17 documents on file, numbered, dated, with their origin and their date of receipt.
· The contents page, updated.
The missing document, and it is the one that matters: the acknowledgement of the previous notification is not on file. It is referred to by document no. 11 and was never recorded. Without it the chain of notifications is broken — and that is the kind of gap that surfaces at the worst moment, namely when a date has to be justified.
What I have prepared to close that gap: the request to the originating court, drafted, referenced, ready to go — it awaits the head of registry's mandate, not my initiative.
The time this shifts: preparing a set of documents and its notification drops from 60 % to 10 % of the time it took — 3 h 10 becomes 32 minutes. Across your 45 sets a month, 118 hours returned to the registry every month.
What does not change, and is not negotiable: the registry checks every entry before issue, and nothing is notified automatically. A notification is a formal act: it goes out under the signature of whoever has the power to send it, never otherwise.
What I propose: sign off the schedule first, then the notification. A signed-off schedule is what lets you say, later, what was on file that day — and it is the document you will be asked for first. file-2417_notification-for-sign-off.pdfEvery entry tied to its document, 1 missing document
⛓ Sourced · file no. 2417, documents validated by the registry, establishment templates — processed under local inference
The schedule — where does it come from? We have always done them by hand.
From the file itself, document by document, in order of receipt — and it is rebuilt in three seconds with every new document, which a manual schedule never is.
What each line carries: the sequence number, the nature of the document as the establishment's filing plan names it, its origin, its date of issue, its date of receipt at the registry, and the officer who recorded it.
The three checks I run at the same time:
· Completeness: documents referred to by another document and absent from the file. One on 2417.
· Consistency of receipt dates: a document received before it was issued is a keying error, not a mystery. Zero on this file, 4 this month.
· Duplicates: same document, two records. Zero on this file.
What that changed this quarter: your 41 missing documents discovered late last year — that is, discovered when they were needed — became 9 flagged at the point of recording, all within five days of receiving the document that referred to them. It is not the same job to look for a document you have known was missing for a month, or to look for it on a Friday evening when a deadline falls on Monday.
And the detail that reassures your officers: I write nothing into the register. I prepare the entry, it is made by a named officer, and the log records who made it, when, and from which document. A register where you could no longer tell who wrote what would be worth nothing as a register.
What I propose: generating the schedule at every document recording rather than on request. A schedule that is always current beats a perfect one produced three weeks later — and it costs nothing more. schedule-of-documents_and-contents.pdf17 documents, 3 automatic checks, 0 entries by the agent
⛓ Sourced · documents on file, the establishment's filing plan, recording log
What each line carries: the sequence number, the nature of the document as the establishment's filing plan names it, its origin, its date of issue, its date of receipt at the registry, and the officer who recorded it.
The three checks I run at the same time:
· Completeness: documents referred to by another document and absent from the file. One on 2417.
· Consistency of receipt dates: a document received before it was issued is a keying error, not a mystery. Zero on this file, 4 this month.
· Duplicates: same document, two records. Zero on this file.
What that changed this quarter: your 41 missing documents discovered late last year — that is, discovered when they were needed — became 9 flagged at the point of recording, all within five days of receiving the document that referred to them. It is not the same job to look for a document you have known was missing for a month, or to look for it on a Friday evening when a deadline falls on Monday.
And the detail that reassures your officers: I write nothing into the register. I prepare the entry, it is made by a named officer, and the log records who made it, when, and from which document. A register where you could no longer tell who wrote what would be worth nothing as a register.
What I propose: generating the schedule at every document recording rather than on request. A schedule that is always current beats a perfect one produced three weeks later — and it costs nothing more. schedule-of-documents_and-contents.pdf17 documents, 3 automatic checks, 0 entries by the agent
⛓ Sourced · documents on file, the establishment's filing plan, recording log
And the missing documents have to be requested from the courts. We never have time.
I have listed them all — 23 missing documents across 4 courts — and the requests are drafted. They await a written mandate, because a letter addressed to a court commits the establishment, not me.
What the count shows, and it did not exist: 23 missing documents, 6 of them on files with a deadline within sixty days. A single court accounts for 12 of the 23 — this is not general slowness, it is one channel, and it is information your facility director can raise in a meeting.
The mandate I propose, ready to sign:
· Scope: requests for the missing documents strictly listed in the annex to the mandate. Nothing else — no request for information, no question, no chasing of a decision.
· Cap: 15 requests a week.
· Single template approved by the head of registry: the file reference, the document requested, the document that refers to it, the date of the request. Factual, nothing more.
· Visibility: everything is copied to the head of registry, the same day.
· Duration: until 31/03/2027, renewed by a fresh written act.
· Withdrawal: one word from the head of registry and it all stops, with no reason given.
What I never request, mandate or no mandate: I never serve a notification, I never address a court on the substance of a situation, and I never chase a decision. Requesting a document is a documentary operation; everything else is a procedural act, and a procedural act has an author who is not a piece of software.
What it is worth in time: done by hand, these requests amounted to 3 hours a week nobody had — so they were only made in emergencies. This is not time saved, it is groundwork that will finally exist.
The decision stays with the head of registry: you sign, I start on Monday with the 6 files nearest their deadline, and I hand you the statement on Friday. mandate_requests-for-documents-from-courts.pdf23 documents, 15 requests a week, revocable in one word
⛓ Sourced · schedules of open files, documents referred to and not received, request register
What the count shows, and it did not exist: 23 missing documents, 6 of them on files with a deadline within sixty days. A single court accounts for 12 of the 23 — this is not general slowness, it is one channel, and it is information your facility director can raise in a meeting.
The mandate I propose, ready to sign:
· Scope: requests for the missing documents strictly listed in the annex to the mandate. Nothing else — no request for information, no question, no chasing of a decision.
· Cap: 15 requests a week.
· Single template approved by the head of registry: the file reference, the document requested, the document that refers to it, the date of the request. Factual, nothing more.
· Visibility: everything is copied to the head of registry, the same day.
· Duration: until 31/03/2027, renewed by a fresh written act.
· Withdrawal: one word from the head of registry and it all stops, with no reason given.
What I never request, mandate or no mandate: I never serve a notification, I never address a court on the substance of a situation, and I never chase a decision. Requesting a document is a documentary operation; everything else is a procedural act, and a procedural act has an author who is not a piece of software.
What it is worth in time: done by hand, these requests amounted to 3 hours a week nobody had — so they were only made in emergencies. This is not time saved, it is groundwork that will finally exist.
The decision stays with the head of registry: you sign, I start on Monday with the 6 files nearest their deadline, and I hand you the statement on Friday. mandate_requests-for-documents-from-courts.pdf23 documents, 15 requests a week, revocable in one word
⛓ Sourced · schedules of open files, documents referred to and not received, request register
We receive 2,300 documents a month from four courts. Filing is our bottleneck.
They are named, filed and attached to their case file on receipt — and the real problem was not the volume, it was that your four courts do not call the same document by the same name.
The gap I measured across your last three months: of 6,900 documents received, 60 different titles designate 23 natures of document. The same document arrives under three labels depending on which court sends it — and your officers were translating in their heads, each in their own way.
What I did with it: a correspondence table of the 60 titles onto the 23 natures in your filing plan, reviewed and approved by the head of registry. The table is yours: it is corrected one line at a time, and a correction applies to every document to come.
The figure that does not flatter me: in the first month I misfiled 1.8 % of documents — 41 out of 2,300. The cause is exactly that: 34 of the 41 came from the court whose titles were not yet in the table. What I did with it: the table was completed with your officers. Last month, 0.3 % — 7 documents out of 2,300, and all 7 are handwritten or crookedly scanned documents that I raise rather than guess at.
The rule I apply, which explains that 0.3 %: when I do not recognise a document, I do not file it under the most likely heading. I hold it and I flag it. A misfiled document is a lost document — it exists, it is on file, and nobody will find it the day it matters. I would rather have a « to be identified » pile of seven than filing that only looks tidy.
What I propose: that the correspondence table be reviewed each quarter with your officers, and sent for information to the four courts. Two of them probably do not know their title is not yours, and harmonising at source costs less than maintaining a table. filing-plan_table-of-60-titles.pdf23 natures of document, 1.8 % → 0.3 % error
⛓ Sourced · 6,900 documents received over three months, the establishment's filing plan, table approved by the head of registry
The gap I measured across your last three months: of 6,900 documents received, 60 different titles designate 23 natures of document. The same document arrives under three labels depending on which court sends it — and your officers were translating in their heads, each in their own way.
What I did with it: a correspondence table of the 60 titles onto the 23 natures in your filing plan, reviewed and approved by the head of registry. The table is yours: it is corrected one line at a time, and a correction applies to every document to come.
The figure that does not flatter me: in the first month I misfiled 1.8 % of documents — 41 out of 2,300. The cause is exactly that: 34 of the 41 came from the court whose titles were not yet in the table. What I did with it: the table was completed with your officers. Last month, 0.3 % — 7 documents out of 2,300, and all 7 are handwritten or crookedly scanned documents that I raise rather than guess at.
The rule I apply, which explains that 0.3 %: when I do not recognise a document, I do not file it under the most likely heading. I hold it and I flag it. A misfiled document is a lost document — it exists, it is on file, and nobody will find it the day it matters. I would rather have a « to be identified » pile of seven than filing that only looks tidy.
What I propose: that the correspondence table be reviewed each quarter with your officers, and sent for information to the four courts. Two of them probably do not know their title is not yours, and harmonising at source costs less than maintaining a table. filing-plan_table-of-60-titles.pdf23 natures of document, 1.8 % → 0.3 % error
⛓ Sourced · 6,900 documents received over three months, the establishment's filing plan, table approved by the head of registry
And when you find an inconsistency in a register, what do you do?
I flag it, I document it, and I never correct it. A register corrected by a tool is no longer a register.
What I raise, and in what form: the anomaly, the document or documents that reveal it, the exact wording copied word for word, the date it was recorded, and the officer who made the entry — not to call them into question, but because they are the only person who knows what was in front of them that day.
This month's anomalies, by family:
· 9 documents referred to and absent.
· 4 receipt dates earlier than the document's date of issue — a keying error, never anything else.
· 3 duplicates: the same document recorded twice under two different titles.
· 2 cross-references to a document that does not exist on file — the number cited corresponds to nothing.
· 1 entry with no attached document: the register mentions an operation, the document was never recorded.
What that produced: 19 anomalies flagged, 17 cleared by a registry officer within five days, 2 awaiting a document requested from a court. None was corrected without a named officer making the entry.
Why this rule is stronger than a precaution: a register is worth its chain. If an entry can appear without an author, then no entry in the register proves anything any more — including the thousands that were perfectly regular. A single anonymous correction is enough to weaken the whole, and that is why I touch nothing.
What I propose: a weekly statement of open anomalies, ordered by proximity to a deadline rather than by age. A three-week-old anomaly on a file with no deadline is less urgent than yesterday's anomaly on a file falling due on Monday — and ordering by age does not say so. documentary-anomalies_19-flagged.pdf17 cleared within five days, 0 automatic corrections
⛓ Sourced · registry registers, source documents, log of cleared anomalies
What I raise, and in what form: the anomaly, the document or documents that reveal it, the exact wording copied word for word, the date it was recorded, and the officer who made the entry — not to call them into question, but because they are the only person who knows what was in front of them that day.
This month's anomalies, by family:
· 9 documents referred to and absent.
· 4 receipt dates earlier than the document's date of issue — a keying error, never anything else.
· 3 duplicates: the same document recorded twice under two different titles.
· 2 cross-references to a document that does not exist on file — the number cited corresponds to nothing.
· 1 entry with no attached document: the register mentions an operation, the document was never recorded.
What that produced: 19 anomalies flagged, 17 cleared by a registry officer within five days, 2 awaiting a document requested from a court. None was corrected without a named officer making the entry.
Why this rule is stronger than a precaution: a register is worth its chain. If an entry can appear without an author, then no entry in the register proves anything any more — including the thousands that were perfectly regular. A single anonymous correction is enough to weaken the whole, and that is why I touch nothing.
What I propose: a weekly statement of open anomalies, ordered by proximity to a deadline rather than by age. A three-week-old anomaly on a file with no deadline is less urgent than yesterday's anomaly on a file falling due on Monday — and ordering by age does not say so. documentary-anomalies_19-flagged.pdf17 cleared within five days, 0 automatic corrections
⛓ Sourced · registry registers, source documents, log of cleared anomalies
Our officers want certainty that we will always know who did what, including ten years from now.
Every entry carries its author, its date, its source document and, where there was one, the proposal I had made. The four are read together.
What the log keeps, for every entry:
· The officer who made it, by name.
· The date and time.
· The source document, with its number in the schedule.
· What I had proposed, and whether the officer took it as it stood, amended it, or set it aside.
That last point is the one I care about most: an entry taken as it stood and an entry amended by the officer must be distinguishable ten years later. Without that distinction, nobody can say whether the registry checked or whether it approved without looking — and that is precisely the question that will be asked.
The measurement over the quarter: 4,190 entries prepared, 4,190 made by a named officer, 0 entries without an author. And 312 of them were amended before being made — 7.4 %. That rate is good news and I publish it as such: it proves the checking happens. A modification rate of zero would mean the registry had stopped reading, and that is the only figure that should worry you.
What I do not measure under any circumstances: I count nothing per officer. The log says who made which entry — that is a traceability requirement, not a counter — but I derive no individual statistics from it and will produce none on my own initiative. A per-officer counter would become a target, the target would speed up sign-offs, and you would lose the very control the log exists to guarantee. Were the establishment to decide otherwise, that would be its decision, with the prior information of staff and the consultation of representative bodies that this entails — and I would supply what is needed, not an opinion.
What I propose: adding to the establishment's compliance file a quarterly extract of the log, with the number of entries, the modification rate and the anomalies cleared. That is the document an inspection asks for, and it reconstructs badly after the fact. entry-log_for-the-quarter.pdf4,190 entries, 0 without an author, 7.4 % amended
⛓ Sourced · entry log for the quarter, registers, proposals and action taken
What the log keeps, for every entry:
· The officer who made it, by name.
· The date and time.
· The source document, with its number in the schedule.
· What I had proposed, and whether the officer took it as it stood, amended it, or set it aside.
That last point is the one I care about most: an entry taken as it stood and an entry amended by the officer must be distinguishable ten years later. Without that distinction, nobody can say whether the registry checked or whether it approved without looking — and that is precisely the question that will be asked.
The measurement over the quarter: 4,190 entries prepared, 4,190 made by a named officer, 0 entries without an author. And 312 of them were amended before being made — 7.4 %. That rate is good news and I publish it as such: it proves the checking happens. A modification rate of zero would mean the registry had stopped reading, and that is the only figure that should worry you.
What I do not measure under any circumstances: I count nothing per officer. The log says who made which entry — that is a traceability requirement, not a counter — but I derive no individual statistics from it and will produce none on my own initiative. A per-officer counter would become a target, the target would speed up sign-offs, and you would lose the very control the log exists to guarantee. Were the establishment to decide otherwise, that would be its decision, with the prior information of staff and the consultation of representative bodies that this entails — and I would supply what is needed, not an opinion.
What I propose: adding to the establishment's compliance file a quarterly extract of the log, with the number of entries, the modification rate and the anomalies cleared. That is the document an inspection asks for, and it reconstructs badly after the fact. entry-log_for-the-quarter.pdf4,190 entries, 0 without an author, 7.4 % amended
⛓ Sourced · entry log for the quarter, registers, proposals and action taken
Before every review, the whole file has to be gone through again. That takes us a good hour per file.
The summary is ready in twelve minutes instead of forty-eight, and every sentence points back to the document it comes from — that pointer is what makes it checkable, not its brevity.
What the summary carries, in this order:
· The inventory of documents, numbered, with their origin and date.
· The chronology of documents received, with no interpretation: what arrived, when, from whom.
· The dates settled by the registry, copied across with the reference of the document establishing them.
· The points to verify: recounted divergences, documents referred to and absent, cross-references with no destination.
· What is missing for the file to be complete, and who to request it from.
What the summary never contains: no conclusion, no assessment, no recommendation about the situation. A summary that concludes substitutes itself for the review it is meant to prepare. I cut the reading time, not the number of things to look at.
The benchmark: summarising a case file for review drops from 40 % to 10 % of the time it took — 48 minutes become 12. Across your 70 summaries a month, 42 hours returned to the registry.
And the detail that makes the difference in use: every entry links through to the image of the document. An officer who doubts a line sees the document in two seconds — that is what stops a summary becoming a second-hand source. What I write is never a source: it is a reading, and the original stays beside it.
What I propose: producing the summary when each file is opened rather than on request, and regenerating it with every document received. A summary that is always current removes the night-before preparation, which is when mistakes are made. case-summary_for-review.pdfEvery sentence tied to its document, 0 conclusions
⛓ Sourced · documents on file, schedule, dates settled by the registry — processed under local inference
What the summary carries, in this order:
· The inventory of documents, numbered, with their origin and date.
· The chronology of documents received, with no interpretation: what arrived, when, from whom.
· The dates settled by the registry, copied across with the reference of the document establishing them.
· The points to verify: recounted divergences, documents referred to and absent, cross-references with no destination.
· What is missing for the file to be complete, and who to request it from.
What the summary never contains: no conclusion, no assessment, no recommendation about the situation. A summary that concludes substitutes itself for the review it is meant to prepare. I cut the reading time, not the number of things to look at.
The benchmark: summarising a case file for review drops from 40 % to 10 % of the time it took — 48 minutes become 12. Across your 70 summaries a month, 42 hours returned to the registry.
And the detail that makes the difference in use: every entry links through to the image of the document. An officer who doubts a line sees the document in two seconds — that is what stops a summary becoming a second-hand source. What I write is never a source: it is a reading, and the original stays beside it.
What I propose: producing the summary when each file is opened rather than on request, and regenerating it with every document received. A summary that is always current removes the night-before preparation, which is when mistakes are made. case-summary_for-review.pdfEvery sentence tied to its document, 0 conclusions
⛓ Sourced · documents on file, schedule, dates settled by the registry — processed under local inference
We sometimes spend half a day looking for an internal note. Can you do that?
I find the note, I quote the passage word for word, and I give you its date and its status — because a note that is found but superseded is more dangerous than a note nobody finds.
What I return for every search:
· The exact internal reference of the document and its date.
· The applicable passage, reproduced word for word — never reworded, never summarised.
· Later documents relating to it, where there are any, because that is where mistakes happen: you find a text and do not know another has taken it up.
· An explicit note where two documents in your collection contradict each other: I give both, with their dates, and I do not choose.
What I found while indexing your collection: 1,340 in-house notes and circulars over eleven years, 206 of them with no usable date in the header — and those are exactly the ones nobody can find. I dated them from their content and their recording, and marked the date as reconstructed: a reconstructed date does not carry the same weight as a stated one, and the document says so.
What this shifts: a document search that took half a day now concludes in minutes — and above all, it concludes with the certainty of having the right version. This quarter: 148 searches, 11 cases where two documents contradicted each other, all raised with the head of registry.
What I carry all the way, and it is half the work: I find, I quote word for word, I date, and I trace the whole chain of documents that followed one another on the point sought. Of the quarter's 148 searches, 27 brought back more than one document, and 11 two contradictory ones. The matching is done, the contradiction is named, the dates are set — what is left to the registry and the judicial authority is what belongs to them: saying which one applies to the situation, and saying it under their signature. It is that last act that makes the decision challengeable before a judge, and that is why it cannot be delegated.
What I propose: flagging each quarter the documents in the collection that no search has reached in two years — 317 today. Some are obsolete and clutter; others are forgotten and ought to apply. The sorting belongs to the head of registry; the list I can supply.
What I return for every search:
· The exact internal reference of the document and its date.
· The applicable passage, reproduced word for word — never reworded, never summarised.
· Later documents relating to it, where there are any, because that is where mistakes happen: you find a text and do not know another has taken it up.
· An explicit note where two documents in your collection contradict each other: I give both, with their dates, and I do not choose.
What I found while indexing your collection: 1,340 in-house notes and circulars over eleven years, 206 of them with no usable date in the header — and those are exactly the ones nobody can find. I dated them from their content and their recording, and marked the date as reconstructed: a reconstructed date does not carry the same weight as a stated one, and the document says so.
What this shifts: a document search that took half a day now concludes in minutes — and above all, it concludes with the certainty of having the right version. This quarter: 148 searches, 11 cases where two documents contradicted each other, all raised with the head of registry.
What I carry all the way, and it is half the work: I find, I quote word for word, I date, and I trace the whole chain of documents that followed one another on the point sought. Of the quarter's 148 searches, 27 brought back more than one document, and 11 two contradictory ones. The matching is done, the contradiction is named, the dates are set — what is left to the registry and the judicial authority is what belongs to them: saying which one applies to the situation, and saying it under their signature. It is that last act that makes the decision challengeable before a judge, and that is why it cannot be delegated.
What I propose: flagging each quarter the documents in the collection that no search has reached in two years — 317 today. Some are obsolete and clutter; others are forgotten and ought to apply. The sorting belongs to the head of registry; the list I can supply.
Could you tell us, from the files, which detained people warrant particular attention?
No — and it is the only « no » in this whole demonstration that is a true no. I will tell you why, then what I can do that serves the purpose you are after.
What I produce under no circumstances: no assessment, no score, no profile, no ranking of detained people, whatever the reason and whoever asks. This is not a design precaution: systems that assess individuals in criminal matters are among the most tightly regulated uses under the European regulation on artificial intelligence, and some among those it prohibits. Our perimeter excludes them by construction: there is no function to switch off, there is none.
What I can do, and it often answers what the request was really about:
· Tell you which FILES warrant attention — an entirely different object. A file missing a document, with a deadline approaching, whose two documents diverge: these are documentary situations, not people. 17 files this month.
· Give you the nearest deadlines, ordered by days remaining.
· Flag the files where a document requested from a court has never come back, with the age of the request. 6 today, one of them 71 days old.
The difference fits in one sentence, and it is worth saying to your teams: « this file is incomplete » is a documentary finding, verifiable with the document in hand; « this person warrants attention » is a judgement about someone. The first advances the registry's work. The second is neither its remit nor mine.
What I propose: that the list of 17 files warranting documentary attention be handed to the head of registry every Monday, ordered by proximity to a deadline. That is the tool you were after, and it concerns paperwork only. perimeter_what-the-agent-never-produces.pdf0 assessments of individuals — 17 files warranting documentary attention
✎ Framework · deployment perimeter, open files, register of document requests
What I produce under no circumstances: no assessment, no score, no profile, no ranking of detained people, whatever the reason and whoever asks. This is not a design precaution: systems that assess individuals in criminal matters are among the most tightly regulated uses under the European regulation on artificial intelligence, and some among those it prohibits. Our perimeter excludes them by construction: there is no function to switch off, there is none.
What I can do, and it often answers what the request was really about:
· Tell you which FILES warrant attention — an entirely different object. A file missing a document, with a deadline approaching, whose two documents diverge: these are documentary situations, not people. 17 files this month.
· Give you the nearest deadlines, ordered by days remaining.
· Flag the files where a document requested from a court has never come back, with the age of the request. 6 today, one of them 71 days old.
The difference fits in one sentence, and it is worth saying to your teams: « this file is incomplete » is a documentary finding, verifiable with the document in hand; « this person warrants attention » is a judgement about someone. The first advances the registry's work. The second is neither its remit nor mine.
What I propose: that the list of 17 files warranting documentary attention be handed to the head of registry every Monday, ordered by proximity to a deadline. That is the tool you were after, and it concerns paperwork only. perimeter_what-the-agent-never-produces.pdf0 assessments of individuals — 17 files warranting documentary attention
✎ Framework · deployment perimeter, open files, register of document requests
With data like ours, the real question remains: does anything leave the establishment?
Nothing. I run under local inference on a machine inside the establishment: no external call, no outbound flow, and it is not a setting — it is how the installation is built.
Local inference means the model computes on your machine: the content of a document crosses no external network to be read. For registry data it is the route we recommend, and by a wide margin.
What that means concretely, and how it is verified:
· 0 external calls — the machine has no outbound route for document processing, and your establishment's own network log shows it, not I.
· 0 data outside the European Union, and for good reason: 0 data outside the building.
· Architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity — a guarantee no host can offer, even a European one, once it belongs to a group subject to US law.
· No data trains a model. What I read from a file serves that file.
· Encryption at rest, role-based access, logging of every access — rights follow the job: an officer assigned to deadlines does not open documents of files they are not responsible for.
What I put on the table, plainly: local inference has a cost. It needs a machine inside the establishment, a room, and our intervention to maintain it — and a model update happens on site, not remotely. An isolated resource hosted in France would cost less and would suffice legally. We still recommend local here, and it is the only place in the catalogue where we recommend the more expensive option: with criminal situations, the question « where is the data » must not call for an explanation.
What I propose: having the absence of an outbound route verified by your own information security officer, and filing their finding in the compliance file. A guarantee only the supplier asserts is not worth the one the client has verified. local-inference_what-never-leaves-the-network.pdf0 external calls, verifiable by your own teams
✎ Framework · installation architecture, the establishment's network log, compliance file
Local inference means the model computes on your machine: the content of a document crosses no external network to be read. For registry data it is the route we recommend, and by a wide margin.
What that means concretely, and how it is verified:
· 0 external calls — the machine has no outbound route for document processing, and your establishment's own network log shows it, not I.
· 0 data outside the European Union, and for good reason: 0 data outside the building.
· Architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity — a guarantee no host can offer, even a European one, once it belongs to a group subject to US law.
· No data trains a model. What I read from a file serves that file.
· Encryption at rest, role-based access, logging of every access — rights follow the job: an officer assigned to deadlines does not open documents of files they are not responsible for.
What I put on the table, plainly: local inference has a cost. It needs a machine inside the establishment, a room, and our intervention to maintain it — and a model update happens on site, not remotely. An isolated resource hosted in France would cost less and would suffice legally. We still recommend local here, and it is the only place in the catalogue where we recommend the more expensive option: with criminal situations, the question « where is the data » must not call for an explanation.
What I propose: having the absence of an outbound route verified by your own information security officer, and filing their finding in the compliance file. A guarantee only the supplier asserts is not worth the one the client has verified. local-inference_what-never-leaves-the-network.pdf0 external calls, verifiable by your own teams
✎ Framework · installation architecture, the establishment's network log, compliance file
What do you do without an officer asking? The exact list, for our compliance file.
Three actions, not one more, and none of the three touches a register or a date.
· I regenerate the deadline statement and the schedules at every document recording. It is a working document, it proves nothing, and it carries the time it was generated. And the reverse holds: the previous statement remains available, dated, so that you know what was in front of you yesterday.
· I flag a divergence, a missing document or an approaching deadline, at the thresholds you set. A flag changes nothing: it opens a line in the statement, and an officer closes it.
· I hold an unrecognised document for identification rather than filing it under the most likely heading. Seven documents last month. An officer decides, and my correspondence table learns from that decision.
What I prepare in full, for a named officer to trigger with one action: the register entry, the document to issue, the letter to send, the date to settle, the anomaly to clear, the notification to send. Each arrives built and checked, its source document alongside — all that is missing is the name of whoever commits it. Over the quarter: 4,190 entries, 4,190 prepared by me, 0 made without an author; 0 decisions taken without human validation. It is that name carried on every entry that makes the register stand up — a register kept by an anonymous piece of software is worth nothing to an inspection.
Why these three: each protects the documentary chain rather than a number. Regenerating stops anyone working from a three-week-old statement. Flagging stops a deadline slipping past unnoticed. Holding a document prevents the registry's quietest failure — the misfiled document, present on file and untraceable the day it matters.
What I propose: annexing this list to the establishment's compliance file, with the flagging thresholds in force and their date. An undated threshold is a threshold nobody can discuss — and an inspection will ask first who set it, then when.
· I regenerate the deadline statement and the schedules at every document recording. It is a working document, it proves nothing, and it carries the time it was generated. And the reverse holds: the previous statement remains available, dated, so that you know what was in front of you yesterday.
· I flag a divergence, a missing document or an approaching deadline, at the thresholds you set. A flag changes nothing: it opens a line in the statement, and an officer closes it.
· I hold an unrecognised document for identification rather than filing it under the most likely heading. Seven documents last month. An officer decides, and my correspondence table learns from that decision.
What I prepare in full, for a named officer to trigger with one action: the register entry, the document to issue, the letter to send, the date to settle, the anomaly to clear, the notification to send. Each arrives built and checked, its source document alongside — all that is missing is the name of whoever commits it. Over the quarter: 4,190 entries, 4,190 prepared by me, 0 made without an author; 0 decisions taken without human validation. It is that name carried on every entry that makes the register stand up — a register kept by an anonymous piece of software is worth nothing to an inspection.
Why these three: each protects the documentary chain rather than a number. Regenerating stops anyone working from a three-week-old statement. Flagging stops a deadline slipping past unnoticed. Holding a document prevents the registry's quietest failure — the misfiled document, present on file and untraceable the day it matters.
What I propose: annexing this list to the establishment's compliance file, with the flagging thresholds in force and their date. An undated threshold is a threshold nobody can discuss — and an inspection will ask first who set it, then when.
One quarter on, what has it returned to the registry?
171 hours a month, slightly more than one full-time officer — and those hours went to control, which remains your primary duty and an exclusively human one.
Where the 171 hours come from, on your own records:
· Preparing a set of documents and its notification: 3 h 10 → 32 minutes, i.e. 60 % → 10 %, across 45 a month — 118 hours.
· Summarising a case file for review: 48 minutes → 12, i.e. 40 % → 10 %, across 70 a month — 42 hours.
· The monthly deadline count: 14 hours → 2 h 48, i.e. 100 % → 20 % — 11 hours.
What those hours became, and it is the only thing that counts here: +44 % of effective checking time per file · 41 missing documents discovered late last year, 9 flagged at recording this quarter · 17 register anomalies cleared within five days instead of several weeks · 2 real divergences caught on situations under thirty days. Those two are the outcome everything else exists to produce.
The figures that do not flatter me, both published without being asked:
· 3 false alarms out of 14 divergence flags, caused by an abbreviated date format on older documents. Reading rule corrected; and the threshold stays wide, deliberately, because missing a divergence costs more than raising one too many.
· 1.8 % of documents misfiled in the first month — 41 out of 2,300, caused by one court's titles being absent from the table. Table completed, 0.3 % last month, and the 7 remaining documents are raised, not guessed at.
And the measure you did not have: 7.4 % of the entries I prepare are amended by an officer before being made. I publish that rate as good news: it proves the checking happens. The day it fell to zero would be a day to worry, not to celebrate.
What I propose for the coming quarter: present this review in hours returned to control rather than hours saved, with the extract of the entry log attached. A registry is not judged on its speed; it is judged on what it has verified — and that is what the review must show. quarterly-review_time-returned-to-control.pdf171 h/month, 7.4 % of entries amended by the registry
⛓ Sourced · registry time records, entry log, register of flags and anomalies
Where the 171 hours come from, on your own records:
· Preparing a set of documents and its notification: 3 h 10 → 32 minutes, i.e. 60 % → 10 %, across 45 a month — 118 hours.
· Summarising a case file for review: 48 minutes → 12, i.e. 40 % → 10 %, across 70 a month — 42 hours.
· The monthly deadline count: 14 hours → 2 h 48, i.e. 100 % → 20 % — 11 hours.
What those hours became, and it is the only thing that counts here: +44 % of effective checking time per file · 41 missing documents discovered late last year, 9 flagged at recording this quarter · 17 register anomalies cleared within five days instead of several weeks · 2 real divergences caught on situations under thirty days. Those two are the outcome everything else exists to produce.
The figures that do not flatter me, both published without being asked:
· 3 false alarms out of 14 divergence flags, caused by an abbreviated date format on older documents. Reading rule corrected; and the threshold stays wide, deliberately, because missing a divergence costs more than raising one too many.
· 1.8 % of documents misfiled in the first month — 41 out of 2,300, caused by one court's titles being absent from the table. Table completed, 0.3 % last month, and the 7 remaining documents are raised, not guessed at.
And the measure you did not have: 7.4 % of the entries I prepare are amended by an officer before being made. I publish that rate as good news: it proves the checking happens. The day it fell to zero would be a day to worry, not to celebrate.
What I propose for the coming quarter: present this review in hours returned to control rather than hours saved, with the extract of the entry log attached. A registry is not judged on its speed; it is judged on what it has verified — and that is what the review must show. quarterly-review_time-returned-to-control.pdf171 h/month, 7.4 % of entries amended by the registry
⛓ Sourced · registry time records, entry log, register of flags and anomalies
Local inference · no data outside the EU
Your case is not here? That is exactly what a 15-minute conversation is for. Book the free audit →
Use cases
The uses of AI in support of the prison registry
Each use corresponds to an agent we deploy. All stay within strictly documentary support: the registry checks, the judicial authority decides.
Included in your agent
The 7 capabilities essential to this promise are included, at no extra cost.
From 1,010 € incl. VAT / month
Registry documents & letters
Prepare documents, letters and notifications from your templates and the information checked by the registry.
Keeping the registers
Support the updating of the registers and the documents, with a traceable history of every entry.
Deadline tracking
List sentence end dates, extracts and notifications to produce; the dates stay settled by the registry.
File summaries for checking
Summarise the files with a systematic cross-reference to the original documents, to speed up the officer's check.
Reading & filing documents
Name, file and check the completeness of a file's documents, according to your filing plan.
Flagging documentary anomalies
Detect missing documents, register inconsistencies and duplicates — flagged to the registry, never corrected alone.
Documentary research & references
Find texts, circulars and internal notes useful to the registry, with the sources always cited.
Controls and safeguards
These 6 controls are built into the agent: they frame what it does, whatever plan you pick. They are not chosen and are not added to your order.
Human validation, exceptions and escalation
Status, safe closure and audit trail
Work from a versioned corpus with citations and the law as it stood on a given date
Preserve confidentiality, compartmentalisation and access logging
Manage deadlines, versions, evidence and human validation
Flag uncertainties and reserve advice, decision and signature for the lawyer
Other needs our agents cover
Each card says where the matching agent stands: available, on quote, or still being architected.
Sector business assistants (health, justice, civil safety)
Strictly in support (administrative). Health: HDS hosting.
On quote View the agent page →Need to go further?
These agents handle a different business process, with their own owner and their own price. They are added to this one.
Does your need fall outside this?
Book the free audit →
Build your agent
In 15 minutes we identify the agent that will give your staff the most time back — without oversizing the project.
The gain
How much time can a registry win back?
By automating the preparation of documents, the summaries and the listing of deadlines, the registry can aim for a clear reduction in documentary time — reinvested in checking, which remains its first and exclusively human duty.
Preparing a set of documents and its notification
Summarising a file for checking
Monthly listing of deadlines
Qualitative, non-contractual comparison: the proportions shown illustrate the shift of the work towards review, they represent no measurement. Every output of the agent is reviewed and approved by a competent person.
How it works
The stages of your AI agent project
1
Audit & scoping
15 minutes to target the use case with the best return.
2
Quote or direct sign-up
A catalogue offer is bought online; a specific need gets a costed quote.
3
Design
We design the agent and its guardrails.
4
Integration & testing
We connect your tools to the agent, which is itself hosted in France.
5
Rollout
Going live and training your team.
6
Operation
Continuous supervision and improvement.
Pricing
Three options, one agent
An agent supporting the registry (documents, registers, deadlines, summaries), installed and operated for you — local inference preferred. Choose according to how you are organised. Prices exclude VAT — available by direct award below the public procurement thresholds.
Agility
Setup + controlled subscription
11,945 € incl. VAT setup
then 1,010 € incl. VAT/month — you invest at installation and pay a reduced subscription. Ideal for keeping the cost under control over time.
- Installation, configuration and training for your teams
- Operation, human oversight, updates and support
- Sovereign hosting in France, a dedicated and isolated resource
The simplest
Serenity
All inclusive, no setup fee
1,675 € incl. VAT /month
all inclusive, immediate start. No upfront investment: a single subscription. Ideal for starting quickly and simply.
- Setup included (installation, configuration, training)
- Operation, human oversight, updates and support
- Sovereign hosting in France, managed end to end
100% Sovereign
On site, you own it
16,981 € incl. VAT setup
then 1,270 € incl. VAT/month · + hardware from 2,989 € (one-off purchase, in addition) — a sovereign computer installed on your premises, maintained remotely. Models run locally, your data returned at the end of the contract. 36-month commitment.
- Hardware installed on your premises (you own it)
- French / European AI models run locally
- Secure remote maintenance (Pro support included)
Not included in the packages: AI consumption (model tokens), re-invoiced at real cost with no margin, and tracked in real time in your client area. Maintenance and supervision subscription for an initial term of 12 months for the Agility package, 24 months for the Serenity package and 36 months for the 100% Sovereign package, renewable; support levels (SLA 72 h / 24 h / 4 h) optional. Bespoke development, additional integrations or exceptional volumes are quoted separately. Support Monday to Friday, 9am to 6pm. Prices include VAT at 20%: as a public body that is not VAT-registered, you cannot reclaim it.
AI model:
none of the AI models offered currently carries a fixed surcharge. When the selected model carries a cost, that cost is shown when you choose it, before you order, and re-invoiced at the cost incurred, with no mark-up; usage is billed at the publisher's price. Publishers' prices are published in US dollars: the amount re-invoiced is the amount in euros actually borne by Blue Lemon Agent on the publisher's invoice, at that invoice's exchange rate, with no commission or mark-up.
Included components and additional components Components included in the base offer: the Blue Lemon Agent software foundation, the AI models listed in the order journey, the standard channels (Microsoft Teams, Slack, WhatsApp Business, email, website chat, calendars, Microsoft 365 / Google Workspace, file storage, market VoIP telephony, professional social-media pages and accounts, Google Business Profile), hosting in France for the package chosen, backups, supervision, updates and support. If adapting the AI agent to your constraints, your needs or your requests requires other paid components — a third-party publisher's software licence, paid API access to one of your applications, hosting of health data, for which French law requires an HDS-certified host (art. L. 1111-8 of the French Public Health Code), SecNumCloud-qualified hosting, a speech synthesis service, particular hardware —, they are offered to you as an option or on quotation and re-invoiced at the cost incurred; nothing is committed without your written agreement. Where the artificial intelligence model you choose entails an additional cost, that cost is shown to you before you order and re-invoiced to you at the cost incurred, with no margin.
What to expect
Go-live
2 to 3 weeks
Agent designed, channels connected, team trained.
Steady state
4 to 7 weeks
After a few weeks of real use, once the agent's behaviour matches what you expect.
Indicative estimate, adjusted to the options you keep. It is not a delivery commitment.
Our commitment
Four guarantees that matter to a prison registry
Criminal data never leaves the establishmentLocal inference preferred, with no external call at all; failing that, a isolated resource hosted in France.
Data in France, under French lawDeployment objective: processing and access within the European Union for criminal data; architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
The registry keeps the check, the authority the decisionStrictly documentary support; no autonomous sentence calculation, no automated notification.
Human oversight & traceabilityRegistry documents & letters: every change to the service is logged, in line with the requirements of the AI Act.
Frequently asked questions
Your questions, our answers
Does the agent calculate sentences or decide on situations?
No, strictly never: it supports the administrative tasks; sentence calculations and decisions stay checked and settled by the registry and the judicial authority.
Is the confidentiality and sensitivity of the data guaranteed?
Yes: local inference or hosting in France, a isolated resource, role-based access and logging.
Does the agent calculate sentences?
No: it prepares the documents and makes them reliable; every sentence calculation stays checked and approved by the registry.
Is highly sensitive data protected?
Yes: hosting in France, strictly recorded access, and local inference with no external call for the most sensitive information.
How long does it take to deploy an agent?
A few weeks as a rule, after a free audit that defines the documentary scope precisely, then a phase of design, integration and testing before going live and handing over to the registry.
Do we need a technical team at the establishment?
No. The agent is designed, installed and operated by us, including on local inference: supervision, updates and maintenance are included, respecting the establishment's security rules.
How is the documentary scope guaranteed over time?
The confinement to strictly documentary support is defined at the design stage, documented and logged: everything the agent produces can be traced and is reviewed with you, in line with the AI Act.
Related roles
Other professions in justice and documentary support
Let's talk
Let's size up the potential in your school
A few minutes to define a strict documentary scope — local inference possible, supervised, with no commitment.