How We Translate

This reading is augmented by AI, not produced by it. The work is not a matter of asking for a translation and copying it down: every word choice goes through a contradictory debate between two AI systems (Claude and Gemini), which alternate between the roles of drafter and merciless critic. One proposes a reading argued from the root and internal cross-reference; the other must tear it down if it does not hold — a philological blind spot, an inconsistency with another verse, a register that betrays the meaning. The debate continues until agreement, never before.

This agreement between AI systems is never final. It is only a pre-filter: AbuIlyass reviews every proposal, settles any remaining disagreement, and has the final word on every translation choice — including overturning a point the two AIs had already agreed on.

When a word or reading departs from the most widespread exegesis, the small ⓘ icon next to the French or English text opens a note explaining the choice: the root relied upon, the verse it was checked against for consistency, and why the common reading was not followed. Nothing is hidden — every departure is justified and available to read.

The concrete workflow, verse by verse

What these figures and this mechanism don't show is the sheer materiality of the work: every surah, from the start to the end of the corpus, was reviewed and corrected over several distinct passes, not just once. A first draft establishes a baseline translation, consistent with terms already fixed elsewhere in the corpus. This first draft is then reviewed verse by verse by AbuIlyass, who corrects reading angles, flags formulations that ring false or betray the text's register, and often proposes a rereading grounded in his own knowledge of the text and of the Arabic roots.

When such a rereading touches a term that structures meaning across several surahs at once — not a mere local rewording — it triggers a doctoral debate between the two AI systems: a first complete argument is produced, then submitted to contradiction, sometimes over several successive rounds, until the proposed reading withstands scrutiny or is abandoned. AbuIlyass oversees this debate at every stage: he can step in directly to settle a specific point, redirect the line of argument, or demand further verification before a term is considered fixed. Nothing is validated by tacit agreement — every system term ("obey" rather than "worship", "anchor point" rather than "direction", to name only two examples) goes through this filter before being applied consistently to every one of its occurrences across the entire corpus.

Once a choice is validated, it is delivered as a patch — never applied blindly: every proposed correction is first tested on a working copy of the database, to check that it breaks nothing and that the result truly matches what was decided, before being handed to AbuIlyass for final application. This working pattern — review, angle correction, supervised contradictory debate, tested patch — repeated surah after surah until the end of the corpus. No surah was validated on a single, definitive reading: each one was reviewed several times, sometimes weeks later, in light of terms fixed in the meantime elsewhere in the text, along with the retroactive corrections that this implies.