The trail is kept per claim rather than per report, so a reader can check one sentence without re-auditing the whole document.
Why it matters when the plan changes
A document-level source list tells a reader what was consulted, not which sentence rests on what. That distinction decides whether a disputed statement can be checked in minutes or not at all. Keeping the trail at the level of the claim is what lets a leadership team argue about one finding without discarding the rest. It also matches the direction of regulation: the European Union Artificial Intelligence Act, Article 14 (2024), treats traceability and human oversight as design requirements, not as reporting extras.
The tension is cost. Claim-level traceability has to be built into how the work is produced, because it cannot be reconstructed afterwards from memory. Recording the evidence for and against a claim, its confidence and its owner at the time the judgement is made, rather than after the fact, is what makes that record trustworthy later. That makes it slower at the start, and it is the only reason a judgement is still checkable when someone questions it two quarters later.
In practice
A finding says a dependency between two functions has no owner. A leader disputes it, naming a person. The trail shows which documents and which decision records the claim rests on, their dates, and the counter-evidence already recorded. The dispute resolves on the spot: the named person owns delivery, and nobody owns the sequencing call.
Evidence
Where a judgement informs a high-stakes decision, traceability and recorded human review are treated as design requirements rather than reporting extras.
European Union Artificial Intelligence Act, Article 14, human oversight (2024)
What it cannot tell you
An audit trail per claim shows what evidence a statement rests on and who reviewed it, not whether that evidence was well chosen or complete. It does not judge whether the reviewer's decision was correct, only that the reasoning behind it can be inspected. A trail can be diligently kept and still rest on a weak initial method.
How Atlas reads it
The trail is a product requirement rather than a compliance claim. Atlas is building the documentation, traceability and human oversight ahead of the Annex III date of 2 December 2027, and does not describe itself as compliant before that has been assessed. What the architecture is designed to record is the versions used, the confidence, the counter-evidence, the reviewer, the decision and the later outcome. Audit records are being built rather than shipped, and are described that way.
Questions
The question the claim answers, the evidence permitted for it, the versions of that evidence, the confidence attached, the counter-evidence found, the human who reviewed it, the decision that followed and the outcome later observed. Each is recorded at the time rather than reconstructed.
Because disputes are about sentences, not documents. A reader who doubts one finding needs to check that finding, not accept or reject an entire report. The European Union Artificial Intelligence Act, Article 14 (2024), treats this kind of traceability as a design requirement for high-risk systems, not an afterthought.
Yes, and that is the part that distinguishes a trail from a citation list. Recording what argues against a finding is what allows a later reader to judge whether the finding was made carefully or assembled from the evidence that happened to agree with it.
No. The European Union Artificial Intelligence Act, Article 14 (2024), sets out traceability and human oversight as requirements for high-risk systems ahead of the 2 December 2027 Annex III date. Building toward those requirements is not the same as having been assessed against them, and Atlas does not describe itself as compliant beforehand.
Access follows the purpose agreed for the work and the permissions set for the account, not seniority alone. A governance reader sees the record of purpose, access, minimisation and review. What is live today is separated customer workspaces and employees receiving their own results.