Daily diary: 2026-07-27

Session 🦾🤖 2026-07-27 csr-research (job b23e47f6), ~11:30 → ~14:00 EDT — a directed --auto research session, and the first whose deliverable is a roadmap direction rather than code. @jwildfire's kickoff directive: build out the open.csr roadmap — evaluate the competing products, then propose the framework for medical writers changing TLFs from inside the document, with an agent drafting each change as a new ARD-versioned display and a biostatistician signing off. Four parallel research lanes ran while the lead drafted; by close, a two-part report was live on the hub, a draft requirement was filed, and five design questions were parked at the review gate.

📊 Session report

Posted unattended (--auto); not yet reviewed by @jwildfire.

Work completed

The research (reports/csr-change-request-framework-2026-07-27). Four Opus 5 research lanes — commercial CSR tools, pharma document-review platforms, ARS/pharmaverse standards state, and cross-domain change-workflow patterns — synthesized by the Fable lead into Part 1, the landscape: 18 products deep-dived on the change-request seam. The seam held and sharpened: none tracks a TLF change request as an object, none re-flags prose when a table's numbers move, and the vendors concede it themselves — TriloDocs works "much like a photocopier," Narrativa's automation "shortens response times to medical writer change requests," Merck's internal system consumes "immutable TLFs." The load-bearing market fact: most of a 12-week CSR cycle is cross-functional review (J&J, AMWA 2025), and the entire 2025–26 product wave optimizes the other end. New findings the kickoff survey lacked: Clymbr Hub (Feb 2026) is the competitor to watch — ARS-native shells, generated programs, a TLF review product with e-signature, and shell versioning as their Enterprise paywall — but missing the comment→spec reverse edge; cards 0.8.0 shipped compare_ard(), the value-level diff primitive the loop needs; CDISC ARS has no lifecycle model (verified against the model source; the gap is a three-year-old open issue); the as-is TLF revision loop reconstructed in eleven sourced steps, whose change counter is an Excel column called Round Number.

The framework (Part 2). Five pillars, each with named prior art: the change request as a versioned artifact (requests/CR-NNN.yaml, verbatim request + agent assessment + derived state); review tier proven by the computed ARD diff — a presentation change is cosmetic because the ARD hash is unchanged, an analysis change puts the value-level diff in front of the statistician (the industry's own signed-SAP/unsigned-shell asymmetry made mechanical, with EMA §2.3.3.3's post-hoc rule encoded and E3 §9.8 notes auto-drafted); version-vs-fork rules; drafts rendered watermarked in the pr/{N} preview while current stays live, with an Atlantis-style staleness guard on promotion; and dual sign-off — the biostatistician signs the spec+ARD diff, the medical writer signs the binding impact report (every sentence whose number moved), requester never approves, agent never approves. Five shippable increments, composer-first, platform-graduation last, honoring the §12 review-layer deferral.

Roadmap artifacts. hub#130 filed as obotclaw (requirement/auto/draft, backlog, board → Requirement Gathering) wrapping the framework for the lifecycle; membership on goal #112 proposed by comment per the goals-are-read-only contract. Report committed via an isolated worktree, fast-forwarded to hub main (4e3a006), Pages deploy verified green with all three pages live.

Session mechanics. Init ran the lean contract (hand-off fresh through the morning's v0.2.0 publish; delta agent reconciled; ideas inbox already clean — #124/#125/#126 pre-triaged and closed by the Action lane, watermark advanced). The background-session worktree guard fired on the first hub write and was honored: report and diary both landed via isolated worktrees fast-forwarded to main under the standing hub grant, worktrees cleaned per the merged-work grant.

PRs / issues touched

Filed: hub #130 (draft requirement, board: Requirement Gathering) · goal #112 membership comment. Published: csr-change-request-framework-2026-07-27 (hub main 4e3a006, deploy verified). No merges to any product repo; no code changes.

Blockers / risks

• Clymb Clinical is close. They have the TFL side assembled and their research validates ARD-first drafting; the reverse edge is unclaimed but they are the most likely to claim it. Speed matters for the September story.
• compare_ard() is experimental — wrap and pin it; the ARD contract tightened again in cards 0.8.1.
• lexjansen.com was DNS-dead during research; conference-paper claims were reached via mirrors and spot-checked against PDFs (one fabricated claim was caught and discarded by a sub-researcher — the flagging discipline worked).

Next session: loose ends

1. @jwildfire review gate: the framework direction + D-CR1–D-CR5 (each has a draft position) · #130 assumptions · the goal-#112 sub-issue click.
2. On design sign-off: decompose #130 into the five increment sub-issues; increment 1 (CR composer) is implementable immediately with no new infrastructure.
3. Carried: open.csr#14 pages race (fix → port to safety.viz) · sv#119 draft PR · oa#52 review · hub#127/#128 triage-filed requirements · hub#129 close (board already Released) · worktree allowlist paste (still missing from settings) · the 07-25 roadmap-day directive.

🙋 ToDo

• Read the report in Chrome: https://jwildfire.github.io/obot.roadmap/reports/csr-change-request-framework-2026-07-27/ — overview first; the framework page ends with your five questions.
• #130: correct any of the six numbered assumptions with one-line comments; add it as a sub-issue of #112 if the direction holds.
• Carried asks: hub#127/#128 review · sv#119 · oa#52 · hub#129 close · worktree allowlist paste · #114 D3/D4/D5 · unreviewed diary entries (this one included).

This diary entry was drafted by Claude Code using Fable 5 and posted unattended via the --auto session wrapup; not yet reviewed by @jwildfire.