Deterministic judgement infrastructure
The deterministic decision layer, for when predictive is not policy.
Live on CIPD today · 34 centres · 20 markets · zero third-party AI in the marking path
01 · The layer
Sokros' own deep-learning stack does the reading: text, video, audio, charts and tables into typed, quoted evidence. Written rules do the judging, quantitative and qualitative. Nothing predictive sits between the evidence and the outcome.
The models read. The rules decide. The record remembers. How the engine works →
01 · Extract
A contract page, a site-visit recording and a certificate register read by the same extraction stack. No field is filled without the words, frame or cell that earned it.
MSA_Meridian.pdf · clause 14.2
“aggregate liability is capped at £2,000,000
site_walkthrough.mp4 · 03:12
“the second fire door was still propped open
cert_register.xlsx · table 3
“forklift operator · certificate expires 2026-03-01
02 · Structure
Evidence lands as typed fields in the schema you define, each value carrying its source. Structured output for your system of record, not a paragraph to re-read.
28 fields on this review, every one quoted. Exported to your schema.
03 · Judge
Quantitative limits run as present-or-not rules against the extracted values. Qualitative quality is banded against your own descriptors, with the reasoning written down. In that order, every time.
04 · Trace
Verdict, rule, quote and model version are stored as one record. Re-run it next year and it lands identically, which is what an auditor, a regulator or an appeal panel actually needs from you.
02 · The doctrine
Ask a generative model the same question twice and you get two answers. An institution cannot certify a coin flip. Sokros checks requirements as present or not against quoted evidence, bands quality only after they stand, and returns the same record every run.
The determinism, replay proofs and calibration underneath: how the engine works →
A distribution moves with the batch. The answer you get starts to depend on what else came in that week.
Fluent prose earns marks that plain English is refused. The bias sits in the phrasing, not the evidence.
Same script in, same grade out, replay-proven on every release. That is the whole doctrine.
03 · Use cases
The same layer, pointed at different rooms: admissions desks, deal rooms, model-release boards. If a decision must survive an audit, an appeal or a headline, it runs here.
Use case 01 · identification
Enrolment records, ID pages and observed footage read into one rule trail: name match, document consistency, photo match, submission behaviour. Four rules, four recorded answers, and a decision a caseworker can reopen a year later, not a similarity score no one can defend.
Use case 02 · business & legal documents
A 41-page agreement becomes typed fields: parties, term, liability cap, governing law, each pinned to the clause that says so. Hard limits check as rules, drafting is banded against your playbook, and the output lands in your schema, not in a summary.
Use case 03 · AI safety & regulation
Sign-off on a model needs an evaluation that holds still: policy criteria as binary checks per output, severity banded with written reasoning, every verdict quoted. Re-run the suite tomorrow and it returns the same result, which is the property a regulator actually asks for.
See your own material judged: a contract, an ID pack, or one unit.
Book a demo04 · Qualifications
Education is where the layer runs live today: 34 centres, 10,000+ assessments a year, every grade replayable for moderation. Everything body-specific is authored configuration, and the full case lives behind your door.
Delivering something else regulated? Bespoke and regulated work →
05 · Partner networks
Built with the researchers who stress-test the layer, the governments regulating AI and education in the UK and the Gulf, and the bodies whose names are on the certificate.
Research
evaluation methods & extraction science
Government & regulators
education and AI oversight, UK · GCC
National programmes
AI & education initiatives, KSA · UAE
Professional bodies
qualifications marked on the record
Centres, awarding bodies and platforms each have a track: how partnering works →
06 · Active work
Live deployments, research lines and programmes being scoped, each with its honest status. Where the work is early, the ledger says so.
PRG-01
CIPD units at Levels 3, 5 and 7, marked in production across 20 markets.
PRG-02
With Google DeepMind: how far a judgement pipeline can be verified with no probabilistic step in the loop.
PRG-03
Assessment rails scoped for NEOM's education programmes: marking, verification and audit as one layer.
PRG-04
With MBZUAI: the extraction stack is multilingual today; this makes Arabic a first-class marking language, not a translation.
PRG-05
Alignment work with SDAIA and UAE education authorities: one deterministic record beneath national qualification programmes.
PRG-06
Scoped with UK public bodies: licensing and casework queues judged against published criteria, with the full rule trail on every decision.
07 · Reach
34 centres mark on Sokros across 20 markets, 10,000+ assessments a year. Different time zones, one deterministic record, the same grade wherever the qualification runs.
London · Dublin · Paris · Berlin · Lisbon · Amsterdam · Bern · Valletta · Ankara · Cairo · Riyadh · Manama · Abu Dhabi · Muscat · Abuja · New Delhi · Kuala Lumpur · Singapore · Jakarta · Ottawa
Sovereign by design
No learner work touches a consumer AI service, and marking never trains a model. No external APIs, no LLM anywhere in the marking path.
UK, EU and Gulf hosting, or fully on-premise on your own GPU floor. The boundary is the institution's to draw, and the pipeline runs the same inside it.
Any decision re-runs byte for byte on demand: for an EQA, an appeal, or an auditor holding a year-old question.
The pipeline stores decisions and quoted evidence, not profiles. Deletion follows the centre's instruction.
Where learner data goes, and doesn't: privacy & data usage →
A demo is your own material judged in front of you: a contract, an ID pack, or a unit with a handful of human-marked scripts. Calibration is measured in dozens, not thousands.
On the call: 30 minutes, one workload scoped, a gated pilot plan. Nothing integrates before you've watched it judge.