Expert-at-the-Core.
Our governance framework for AI-assisted language work — published, versioned, and specific enough to be audited.
Version 1.0 · 26 August 2026 · reviewed annually at Language_Tech Patras
Talk to us about your content →
Why this page exists
Every language company in this market now says it keeps humans in the loop. The claim costs nothing to make and nothing to break, which is why all of them make it. It has stopped being information.
What follows is our version, written down at the level of detail your compliance officer can check: which content may be touched by a machine, which may not, who signs, what is kept, and what we refuse. We publish it deliberately. A framework that cannot be inspected is a slogan with better manners.
We are not asking to be believed. We are describing what we do precisely enough that you can hold us to it.
At the core, not in the loop
The distinction is not a rebranding.
In-the-loop places the person somewhere inside a machine process — a checkpoint on a pipeline the machine owns, inspecting output it has already produced. The loop is the system; the person is a component of it, and components get sampled, batched, thinned and eventually removed as confidence rises. That is the direction of travel across most of the industry, and it is honest about what it is: a cost curve.
At-the-core inverts the ownership. The specialist holds the ground truth — the verified human reference every machine output is measured against — and the machine runs underneath as infrastructure. The expert does not inspect the pipeline’s output; the expert defines what correct means, and the pipeline is measured against that. Remove the machine and the work still happens, more slowly. Remove the expert and there is no standard left to measure anything by.
The assistance ladder
Five levels. Your contract names the level for each class of your content. The level is a ceiling, not a target.
| Level | What it permits | What it requires |
|---|---|---|
| A0 — Human-only, closed | No machine translation, no LLM assistance, no external service call. Isolated environment where your posture requires it. | Named specialist translator; named independent reviser; full bilingual review. |
| A1 — Human-only, assisted environment | No machine-generated target text. Machine assistance confined to memory retrieval, terminology recognition, consistency and numeric checking, format repair. | Named specialist translator; named independent reviser; full bilingual review. |
| A2 — Machine-drafted, full post-edit | A machine may produce a draft. The draft has no standing — it is raw material. A specialist with domain depth revises to publication standard against source. | Named post-editor with domain qualification; independent revision; MQM scoring of raw against final. |
| A3 — Machine-drafted, light post-edit | As A2, revised to a defined fitness standard rather than publication standard. Accuracy and safety complete; style not optimised. | Your written acceptance of the standard, recorded in the contract — never assumed. |
| A4 — Machine output, human sampling | Output released against statistical sampling rather than full review. | Available only for low-consequence content, only on your instruction, with the sampling rate and acceptance threshold written in. |
The level is declared, never assumed. If nobody has told you which level your work runs at, you are being sold an assurance that has not been specified. Where we have not written the level into the engagement, the default is A2 and nothing higher may be applied.
And it only moves one way on its own. Any linguist on any job may move your text down the ladder — towards more human involvement — without asking permission, and records why. Moving it up requires a written variation from you. That asymmetry is the mechanism that makes everything above it survive a deadline.
Consequence tiers
We classify your content by what an error would do in the world, not by how hard it is or what it costs.
| Tier | An error would… | Ceiling |
|---|---|---|
| C4 — Critical | Endanger a person, or bind them in law without recourse — clinical documentation, safety instructions, court filings, certified documents, regulatory submissions | A1 |
| C3 — High | Create legal, financial or regulatory exposure — financial disclosures, tenders, technical specifications, compliance documents, EU institutional text | A2 |
| C2 — Standard | Damage reputation, mislead, or force costly rework — corporate communications, marketing, websites, training material, subtitles, scholarly text | A2, or A3 by agreement |
| C1 — Low | Cause inconvenience — internal information, gisting, commodity product content, archive material | A3, or A4 on instruction |
The tier travels with the text, not with the account. A critical document arriving inside a routine batch is treated as critical. That is the failure most likely to bite in high-volume work, and catching it is somebody’s named job here, not a hope.
No commercial override. A deadline does not lower a tier. A budget does not lower a tier. If the work cannot be done properly in the time available, we will tell you what is achievable and let you decide — we decline the deadline, not the standard.
What you receive
Alongside the translation:
- The assistance level and consequence tier applied, per file.
- Who translated it, who revised it, who ruled on terminology — by name.
- An MQM / DQF score report: per-category failure counts, severity ratings, examples. Not a pass or a fail.
- A raw-versus-final comparison, so you can see what the machine actually contributed and decide where to invest — in better prompts, or in more human work.
- Terminology release notes: what was ruled, on whose authority, what changed.
- Any point at which we moved your text down the ladder, and why.
That last set is the part of this framework that costs us something. It is also the part nobody can claim without producing it.
What we will not do
The refusals are part of the specification. A framework with no refusals in it is a description of whatever the client last asked for.
- We will not release machine output as human translation. Where the assurance is wanted without the work, we decline the engagement.
- We will not staff an engagement anonymously. You will know who did the work.
- We will not lower a consequence tier to meet a deadline or a budget.
- We will not apply sampling-only review to anything above low-consequence content, on any commercial terms.
- We will not certify or attest a translation we did not govern. Rubber-stamping third-party output is not a service we sell.
- We will not train models on your material without your written consent, nor permit a subprocessor to.
- We will not accept a brief whose accuracy requirement we cannot staff. Where the expertise is not in our network we will say so, and where we can, tell you who has it.
Where it comes from
The framework is not a position paper. It rests on infrastructure built over twenty years: 146 termbases and 197 translation memories under active governance, three EU-co-funded R&D programmes, research partnerships with the University of Patras and the University of Piraeus, and our own translation management system, R.I.M.A., in which these rules are enforced by the software rather than remembered by people.
It is developed inside RIMA Language_Tech_Culture, the research and technology hub, and lived with by Rhyme&Reason, the studio inside it, on real client work. That is an unusual shape for a language company, and it is the shape this framework needs.
It is reviewed once a year and opened to challenge at Language_Tech Patras, the forum we convene each November in Patras. The version number and date are on every release. A framework that visibly gets revised is a framework in use.
Tell us what your content has to survive.
Send us the material, or the problem. We will tell you which tier it falls in, which level it should run at, and what that costs — before you commit to anything.