Two questions from @jwildfire: what do the other products actually do, and how should medical writers change TLFs from inside the document — proposing in plain language, an agent drafting the new display version via the ARD approach, a biostatistician signing off? Part 1 answers the first with an 18-product deep-dive. Part 2 answers the second with a five-pillar framework and five shippable increments.
The seam held, and got sharper. Eighteen products examined: none tracks a TLF change request as an object, none re-flags prose when a table's numbers change, and the vendors say so themselves — Narrativa's automation “shortens response times to medical writer change requests” (the request is still an email to programming), TriloDocs works “much like a photocopier,” Merck's internal system consumes “immutable TLFs.” The bottleneck the whole 2025–26 wave ignores: most of a 12-week CSR cycle is cross-functional review, not drafting (J&J, AMWA 2025). One new competitor is assembling the pieces on the TLF side — Clymb Clinical — and shell versioning is literally its Enterprise paywall. Meanwhile the substrate moved our way: {cards} 0.8.0 shipped a purpose-built ARD diff, and CDISC ARS was verified to have no lifecycle model at all.
The framework in one sentence: the writer asks in prose, the agent answers in spec, the pipeline answers in numbers, and the statistician signs the diff, not the anecdote — with the writer signing the symmetric half, a report of every sentence whose number moved.
Every funded entrant of 2025–26 — Saama, Weave, AlphaLife, TrialAssure, plus the incumbents — accelerates the first draft. The practitioner literature and J&J's own head of regulatory writing put the cost in the review loop: hundreds of RTF files, comments re-keyed by hand into trackers, a 100%-manual number check after every regeneration. open.csr's differentiator is aimed at the expensive half. (Part 1 §1, §4)
The industry already runs two-tier change control: analysis changes ride signed SAP amendments; shell/layout changes are low-ceremony because shells are typically unsigned. And its de-facto change counter is a column called Round Number in an Excel tracker. The framework doesn't invent governance — it makes the existing asymmetry mechanical, with the ARD hash as the proof of which tier a change belongs to. (Part 1 §4, Part 2 §2)
cards::compare_ard() (0.8.0) is the value-level diff engine the loop needs, purpose-built, upstream. cards::mock_*() makes a table shell an empty ARD in the results schema — one diff engine for shell review and results review. ARS has no lifecycle fields at all (verified in the model source; the gap is a three-year-old open issue), which makes open.csr's iteration ledger the reference implementation of a documented hole, not a deviation. (Part 1 §5)
No published precedent exists — the industry vocabulary is whole-artifact sign-off. The strongest measured result in the space says specification quality, not model sophistication, governs safety (20%→90% correctness with behavior-driven specs). The framework's regulatory placement is deliberate: the agent never generates evidence — numbers come from the deterministic pipeline — keeping it outside FDA's draft credibility framework while dual human sign-off implements EMA's supervision language as workflow. (Part 1 §5, Part 2 §5)
ARS-native shell authoring, generated programs, the only dedicated TLF review product (with e-signature), and a CSR builder — announced as a unified hub in February 2026. What they haven't shipped is exactly open.csr's thesis: the reverse edge from a reviewer's comment to a spec edit, and prose that knows its numbers changed. Their published research even validates the ARD-first bet (3–5× over RTF parsing). Watch them; move first. (Part 1 §2)