Sokros

Solutions · Discovery

Everything in, from day one.

Sokros reads what your organisation already holds into typed, quoted fields: the archive on migration day, the connected systems every day after. Nobody re-keys anything.

The same reading stack runs live on CIPD assessment across 34 centres

01 · Two ways in

Migrate once. Ingest forever.

Every system swap threatens the same project: a year of re-typing the old estate into the new one. Discovery replaces it with two doors into one record.

The archive

A scoped migration, gated

The backlog is scoped like a pilot: a fixed slice first, priced upfront, read and checked against your own records. The gate decides when the next slice goes. Nothing is declared migrated because a progress bar filled.

The living estate

Ingestion on a schedule

Connected sources keep feeding the record: mailboxes on arrival, drives on a daily sweep, the database hourly. After migration day nobody runs an ingestion project. The schedule runs it.

Sources · Meridian estate
The archivescoped slices, checked at the gatemigrated
Document storepolicies, contracts, reportsdaily sweep
Shared drivesthe folders everyone means by “the system”daily sweep
Mailboxesattachments and the decisions buried around themon arrival
LMSsubmissions, understood by unit, level and cohort year06:00 sweep
Operational databasestructured rows, checked at the doorhourly
3 rows failed the date ruleReturned with reasons

02 · What it reads

Every carrier you actually hold.

An estate is never clean type. The same stack reads the fact wherever it sits, and lands it in the same typed field with the words, frame or cell that earned it.

Documents & PDF

Forty-one pages of contract become clause-level fields. Layout, footnotes and cross-references survive the read.

liability_cap£2,000,000

aggregate liability is capped at £2,000,000

MSA_Meridian.pdf · clause 14.2

Images & handwriting

Scans and photographs read like type, margin notes included. Nothing in the archive is set aside as too messy to migrate.

maintenance_flagtrue

forklift out of service since Tuesday

site_notes_scan.jpg · margin, p.2

Charts

Axes, series and the point that matters are read as values, not skimmed as decoration.

q3_utilisation96%

series B · September peak

capacity_chart.png

Tables

Cells become typed fields with their row and column held, ready to be checked as rules.

cert_expiry.forklift2026-03-01

forklift operator · certificate expires 2026-03-01

cert_register.xlsx · table 3

Video

A frame and its transcript are read together, and the finding carries both.

fire_door_compliantfalse

the second fire door was still propped open

site_walkthrough.mp4 · 03:12

Audio

Recordings are transcribed and quoted with their timestamp, so the claim points at the moment it was said.

agreed_cap_spoken£2,000,000

we agreed the cap at two million

briefing_call.m4a · 07:41

Web

The registers and portals your records already cite are read as sources, and the page is quoted like a page.

registered_statusactive

Meridian BioTech Limited · active

companies register · checked at read

LMS & file systems

Submissions and files are pulled from where they live. Unit, level and cohort year are understood on arrival, not assigned by hand.

submission.unit5CO01 · Level 5

cohort 2026-A · pulled on the 06:00 sweep

LMS · assignment queue

Email & attachments

A mailbox is a filing system nobody indexed. Attachments are split out, and the decisions written around them are read too.

supplier_mentionMeridian BioTech

Meridian (biotech)

mailbox · quote thread

Reading is multilingual, Arabic included. The quote stays in the language the author used, and the field arrives in yours.

03 · The connectors

A plugin, not a project.

No third software layer to manage, no file organisation, no copy-paste. A connector is pointed at your platform, the pull cadence set to your plan, and submissions arrive without a file changing hands.

Live connections to the platforms centres already run

The same pattern covers every source class: document stores, shared drives, mailboxes and operational databases connect through connectors authored as configuration, pulling on a schedule, each pull recorded like everything else. A new source is an afternoon of settings, not a quarter of integration work.

04 · What you bring

Day one asks for less than you think.

Calibration pack · what's inside
Assignment brief · 2026the case study and task wording
Rubric & grade descriptorsthe bands your assessors mark to
Feedback-sheet templatethe .docx your centre files
Human-marked scriptsdozens, across the grade range

What we need from you

One folder, honestly.

Onboarding asks for the things your centre already has: the assignment brief, the rubric and band descriptors, your official feedback-sheet template, and a calibration set of human-marked scripts. Dozens, not thousands. Sokros does not need your marking history, your markers' time, or a training period on live learners.

Request a demo
No training on live learnersNo marking history requiredGated go-live

05 · Accuracy you can check

Accuracy is a property you can open.

Not a percentage on a slide. Every migrated field carries its verbatim quote, its page and position, and an agreement count you can recompute yourself.

Audit checks are rules that run while the read happens, not a second model marking the first one's homework afterwards. Agreement is counted across independent runs, and the score is a fraction, never a feeling. A field that fails its checks goes to a person with its evidence attached, not quietly into the record.

Field record · liability_cap
value£2,000,000
source

aggregate liability is capped at £2,000,000

page 14 · clause 14.2 · lines 3–4

agreementindependent reading runs, identical value3 of 3
audit rule · ran mid-readcurrency and magnitude consistent with clause 14 context

06 · The boundary

Your archive stays yours.

Zero third-party AI in the marking path

Sokros' own models do the reading. No document in your archive is sent to a consumer AI service to be understood.

In-region or in-building

UK, EU and Gulf hosting, or fully on-premise on your own GPU floor. The read runs the same inside whichever boundary you draw.

Evidence, not training data

Documents are read, quoted and recorded. Nothing resets, nothing retrains, and your files never teach a model.

Where data goes, and doesn't: privacy & data usage →

Bring the archive nobody opens.

The shared drive from three offices ago, the leaver's mailbox, the scanned bundles. A 30-minute briefing, one slice read in front of you, every field quoted.

Migration is scoped and gated: a fixed slice first, checked against your own records, then the rest.