Today the first CIPD cohort marked on Sokros receives its results. Six centres, three levels, and just under a thousand scripts, marked against the same written rules, with every grade carrying the quotes that earned it. No announcement we have made matters more than this one, because none of the others came with a register of learners at the end of it.
This post is the introduction we owe the sector: what the capability is, how a unit goes live on it, what the first month actually looked like in numbers, and what assessors keep hold of throughout.
What the capability is
CIPD marking on Sokros covers Levels 3, 5 and 7: the full spread from foundation assignments to the strategic writing Level 7 demands. A centre's units run on the centre's own briefs and CIPD's own descriptors. The extraction stack reads each submission and quotes every field it fills; written rules check each assessment criterion against that quoted evidence, present or not; band judgements are made against the descriptors with the reasoning recorded; and the whole decision is stored so it replays, byte for byte, for a moderator, an EQA or an appeal panel a year from now.
Two properties matter more than any feature list. There is zero third-party AI in the marking path, so learner work never touches a consumer AI service and never trains a model. And calibration is measured in dozens of human-marked scripts per unit, not thousands: a centre does not need a labelled archive to start, it needs a handful of scripts its own assessors have already marked.
How a unit goes live
- 01 · Brief and schemethe centre's unit, CIPD's descriptors
- 02 · Calibrationdozens of human-marked scripts
- 03 · Live markingrules first, quality banded after
- 04 · Moderation-readyevery grade carries its quotes
The first month in numbers
Launch numbers should be read as a starting line, not a victory lap. Agreement with assessor grades on sampled scripts opened at 94.8%: strong for a first window, and below where we expect it to settle as calibration beds in. Turnaround on instant plans landed under two minutes from brief open to judgement recorded. And every one of the first month's scripts closed with a complete evidence trail, which is the figure the moderation season ahead will actually test.
First cohort · September 2025
Centres live at launch, UK and Gulf
6
12 more onboarding for January
Scripts marked across Levels 3, 5 and 7 in the first month
940
First-pass agreement with assessor grades on sampled scripts
94.8%
opening baseline, first window
Median turnaround per submission on instant plans
1 min 52 s
What assessors keep
The capability is built around a simple division of labour, and the launch cohort held us to it from day one. The pipeline reads, quotes, checks and drafts. The assessor reviews the evidence and owns the grade. Nothing about that second part is decorative: every grade a learner received this month is one a named assessor signed, and every integrity flag landed with a person as a signal to investigate, never as a penalty applied by software.
The division of labour, as launched
The pipeline
- Reads every submission and quotes each field it fills
- Checks written criteria against quoted evidence, present or not
- Bands quality against CIPD descriptors, reasoning recorded
- Drafts feedback that quotes the learner's own work
The assessor
- Reviews the evidence trail instead of reading scripts cold
- Owns and signs every grade a learner receives
- Investigates integrity signals with the quotes attached
- Handles appeals, boundaries and reasonable adjustments
I stopped reading scripts and started reading evidence. The judging is still mine; the archaeology is gone.
What comes next
January brings the second cohort: more centres, more units, and the first full moderation cycle over Sokros-marked grades, which is the examination this capability was really built to sit. We will report the numbers the same way every quarter: what sped up, what held, and what we changed because moderation told us to.