Roadmap audit 2026-08-15 — 27 changes applied, 7 calls for you, and 5 rules that fail your own test

[jwildfire]

Roadmap clean-up ran tonight under the new auto-accept grant. 27 changes applied without asking; 7 calls are yours. Full argument: The roadmap audit audits itself

The headline is your own aside. You wrote that if a finding needs your attention it probably isn't a good audit rule. Scored that way, 5 of the 22 rules fail — and 18 of the 24 findings that landed in your lap came from just two of them. The dividing line turned out to be grammar: rules that propose a specific change (set this milestone, move this stage) resolve themselves; rules that propose a question ("establish what is blocking it") queue for you forever. §3 of the page is that analysis and is the part worth reading.

Already applied, no answer needed: 11 hub milestones · backlog milestones created in obot.agent and safety.viz (neither had one, so your milestone-before-work rule was unsatisfiable there) · 4 board-stage corrections · 6 sub-issue links · 1 assignee · and one code fix: GOALLESS-REQUIREMENT now walks the whole ancestor chain, closing #133 — it was reporting 2 false positives out of 3 findings.

The seven calls:

• C1 — close 4 issues that are provably shipped (Requirement: Safety Histogram migration (pilot) #2, Requirement: obot GitHub App for agent and automation access #3, Requirement: safety.viz documentation site — example gallery, test evidence, and API reference #21, gsm.safety#44; evidence now recorded in-thread on each). Rec: close all four, and extend the grant to closes that clear a hard bar.
• O1 — 3 pieces of live scope stranded under closed parents (obot.agent#14, obot.agent#15, safety.viz#99). Rec: re-parent two, close the moot one, file a follow-up for the lab-family rollout.
• H1 — retire Morning review & release guide — 2026-07-25 (checkbox = go) #114, the 2026-07-25 review guide: 9 boxes checked, 11 open, three weeks cold, invisible to every surface. Rec: close it and re-file only its 4 live items.
• R1 — STALLED-IN-FLIGHT fired 13 times, all yours, and asks a question no data can answer. Rec: replace with an auto-park that moves 30-day-quiet items back to Backlog and says so.
• R2 — DESIGN-MISSING fires on work already built, which is documentation theatre. Rec: fire it on the Design → Development transition instead.
• R3 — GOAL-BOARD-INCONSISTENT has asked the same convention question for 20 straight nights. Rec: pick a default (goals are not board items) and let the rule enforce it.
• R4 — OPEN-IN-RELEASED can't tell "done and unclosed" from "phase 1 done, phase 2 pending", so Requirement: nepExplorer migration — KDIGO kidney-safety explorer for safety.viz #35 is a false positive. Rec: honour a phased marker; treat recent activity as a stage error, not a close candidate.

Adopting the recommendations takes the same roadmap from 24 findings needing you (53%) to about 6 (13%), and the six that remain are one-off scope judgements rather than nightly repeats.

Answering "adopt all" is a complete response.

One structural note: three of the five failing rules fail for the same missing capability — the audit sees only tonight's snapshot and has no memory of stage transitions. Adding a stage-history field is the single highest-leverage change to it.

---

This discussion was drafted by Claude Code using Opus 5 and reviewed by @jwildfire

[jwildfire]

Update — the apply batch itself turned out to be the most useful test.

Re-running the audit after applying produced 6 brand-new findings, which exposed three rule interactions that reading findings never would. Added as §3.5 of the artifact:

• UNTRACKED-TASK's fix manufactures UNSTAGED-BOARD-ITEM findings. Linking Add legend and hover detail to hep-waterfall box-and-whisker panels #83, Show release summaries on the roadmap page — recent releases render as an em-dash #94, Four documents disagree on whether an increment PR should be a GitHub draft #152 and demo-301's scheduled pipeline has failed silently since 2026-08-03 — and the repo belongs to no goal #153 under a parent silently added all four to the board with no Status, so accepting one finding immediately created another. Both rules are correct; neither knows the other exists. The fix is for UNTRACKED-TASK to set a Status in the same operation.
• SUBS-DONE-PARENT-OPEN mistakes an empty sub-issue list for finished scope. Correcting Requirement: open.csr submission outputs — RTF tables/listings, a values store, and seamless Reader editing #129's false Released stage made it propose Review — but Part C is real outstanding work that simply has no sub-issue yet. The rule is reasoning about the tracker, not the work.
• The escalation ladder works. Moving Requirement: abnormal-baseline DILI tools — migration Sankey + ALT waterfall (Amirzadegan 2025) #43 to Review immediately escalated it to a close proposal, so there is one more close waiting behind C1.

These had been latent since the audit shipped on 2026-07-24, because findings were being read rather than executed in bulk. General lesson worth adopting: a rule that reliably creates work for another rule is a defect in the first one, and the audit should re-run immediately after any apply batch.

Revised numbers: 32 changes applied (27 from the audit, 5 cleaning up after it), live findings 45 → 28. The seven calls above are unchanged.

---

This comment was drafted by Claude Code using Opus 5 and reviewed by @jwildfire