Per-span citations mean your document store's version history is now part of the audit trail

A char_range is an offset into one byte sequence. The guarantee only holds while the store can still serve manual_A2131 at doc_version 2024-11 exactly as it was chunked, which is the point of pairing doc_id with doc_version and a character range: the cited span can be re-fetched and re-read a year later, where a doc_id alone is useless once the manual has been revised ^[13]. So the ingestion pipeline becomes the artifact under audit. Re-extracting the PDF with a different parser, normalising whitespace differently, or moving chunk boundaries all repoint those offsets at other text, and nothing in the response JSON notices. Only a re-fetch does.

Which is why verification has to run as a job rather than sit in a policy document: re-fetch each cited span, confirm it still exists at that version, run an entailment check of the span against the sentence it supports, and let the failures pile up in a queue that is itself the evidence ^[16]. Budget one entailment call per span per answer. The checker is a model too, so the post's own supply-chain numbers apply to it: 56.67% license omission in adapter derivations, and 8.05% license drift in fine-tuned models ^[9].

The provisions the post cites do less work than its framing suggests. Article 19 asks for six-month log retention ^[4]. Article 99 sets the ceiling at EUR 15 million or 3% of global turnover ^[5]. The marking requirement wants machine-readable flags on generated content ^[11]. A well-built decision log satisfies all three. None of the named provisions requires support asserted at span level ^[22]. The claim that this is the artifact a regulator reaches for once the question turns from what the system did to whether any of it was true is the author's own ^[2], and it is a bet on supervisory practice rather than a reading of a clause. The engineering case is the stronger one: per-span support is what makes a hallucinated sentence unable to hide between two grounded ones, and what makes unsupported_spans computable at all ^[14].

Read the fairness comparison as a claim about someone else's dataset. In the 2026 clinical study the post cites, logistic regression reached 75.2% accuracy against a random forest at 70.1% ^[17], a gap of 5.1 points ^[20]. Equal opportunity difference ran 0.256 against 0.055 ^[17]. The post calls the second model roughly 57% less biased; 1 minus 0.055/0.256 is 0.785, so the printed numbers give about 79% ^[18]. For either figure to transfer to your model selection you would need the same protected attribute, the same base rates and the same decision threshold, and the write-up names none of them.

The cheap half of this design is deleting the citation string the prompt asks the model to append, because only a first-class field in the response contract is verifiable ^[21]. The expensive half is a document store that can still hand back bytes 1180 to 1372 of chunk 417 in a revision you shipped last year ^[12].