Asked, answered, on record.

No. Sokros reads each criterion into typed fields with verbatim quotes, checks the brief's hard requirements as binary present-or-not gates, judges quality against the unit's own rubric only after those requirements stand, and resolves everything through a written policy layer. Nothing probabilistic or generative decides an outcome, and every decision is stored with its evidence.

Yes — same script in, same grade and same reasoning out, whenever it is marked. Model versions are pinned, gates are declarative, and every release replays past scripts and compares output byte for byte. A release is blocked if a single calibration script flips between Refer and a clear Pass.

It can't. Core requirements are graded on extracted evidence — if the extraction finds no second theory, no fluency fills that field. The reverse also holds: a plainly-worded answer that meets the criteria extracts fully. An international learner's English is never the thing being graded.

Mathematically, from gate outcomes, judging margins and corroborating statistical signals — not from a model predicting its own reliability. A flag means a measurable margin was thin, and it routes the script to a human.

Yes. Every evidence mode — video, audio, chart, image, table — runs through the same extraction and gating as text, with modes preconfigured per criterion for supported qualifications, and output written into your centre's own feedback marksheet.

RTD returns ineligible work before any marking is spent on it — missing signature, wrong unit or brief, word count beyond tolerance, no referencing attempt where the level requires one, formative work in a summative area, or integrity results outside range. Checks run on arrival; depending on plan the return is instant (under two minutes) or in a daily sweep, always with the recorded reasons attached.

When ineligible work sits in a human queue, learners wait one to two weeks to be told about a trivial fix, then re-queue again — and centres pay markers to discover it. Automated RTD keeps the queue markable, which also makes workload forecasts and invoicing match reality instead of swinging with hidden RTD rates.

Every attempt has its own immutable record. A resubmission is marked with the prior referral in hand — the additions are checked against exactly what was referred — and awarding-body resubmission caps are applied automatically.

It pulls submissions on your schedule from the popular LMS platforms, understands the assessment's place on the site — unit, level, cohort year — and after marking completes the procedure on the LMS itself: attaches the document, selects grade buttons or criteria tables, writes general feedback on the assessment page, changes the workflow and marking state, and saves or releases.

No. Choose fully automated, or a dispatch queue where your team approves or amends on the Sokros Portal and one click releases on the LMS — in batches, not script by script.

Anywhere you choose: sample a percentage of marked work, hold referrals for review, or require sign-off on everything. Overrides are recorded like every other decision.

Every script's journey — inputs, extractions with quotes, gate outcomes, judgements with reasoning, policy resolutions, the rendered sheet — is stored and retrievable per learner and per cohort. “Why was this learner referred?” is answered from the record in seconds, and moderation-style sampling reports run across any cohort.

Never. Plagiarism, collusion and AI-usage results are signals with quoted evidence, delivered to an assessor to decide. Reference credibility is classified source by source the same way.

A new cohort year is a new configuration version alongside the old one. Both stay runnable and auditable; nothing resets and no accuracy clock restarts. Every behavioural change to the system itself is a numbered amendment with written rationale.

No. Proprietary models run on dedicated infrastructure under Sokros control — no third-party LLM APIs sit anywhere in the marking path, and UK, EU, Gulf and fully on-premise deployments are available.

Learner data is not used to train models. Sokros also doesn't need it to be: a unit is calibrated on dozens of human-marked scripts and is fully capable from the first assessment. Several grading tools' own privacy policies state the opposite — see our privacy comparison.

You do. The pipeline stores decisions and quoted evidence, not profiles; retention and deletion follow centre instruction, with encryption in transit and at rest.

It is accurate from the first script. There is no thousand-assessment warm-up and no period of watching your markers: a unit goes live after a gated calibration against dozens of your human-marked scripts, and an outcome does not depend on how many scripts came before it.

The brief, the rubric, your official feedback-sheet template and a calibration set of human-marked scripts — dozens, not thousands. Everything else is authored as configuration on our side.

A fixed-price pilot on your own scripts, per-script pricing for live centres, or a sovereign deployment on your infrastructure. RTD speed and LMS pull cadence scale with plan. Numbers after one call.