Board PackBF-NORTHWIN-20260925
Northwind Engineering
Last 180 days · 2026-03-29 to 2026-09-25 UTC
431 commits77 merged PRs9 people1 repo
GradeC63/100 ▲ 1
The verdict
Win, On Record: 9 weeks kept in a row. Working, proven, on record. Take the credit - we brought receipts.
Aisha Patel, Diego Fernandez and Priya Sharma hold 3 areas alone. Rebuilding that knowledge costs €8,930-€26,790.
The Slop Treadmill: 21.6% of AI-delivered code comes back as a bug fix - vs 6.3% for everyone else.
AI-attributed commits37.6%6.5% of merged PRs
AI PR cycle time vs the rest6.5% faster144.0h vs 154.0h medianlow sample: 5 of 20 PRs
AI rework vs human1.00×89.1% vs 88.7% re-edited within 14 dayslow sample: 5 of 20 PRs
Danger-zone areas1of 8 active areas
Protect list, rebuild cost€8,930-€26,790120-360 hours, 3 areas, 3 people
Decisions for the board
- Fund 120-360 hours of handover this quarter, starting with payments from Priya Sharma.
- Hold AI spend flat until 20 AI PRs have merged, then judge it.
- Put this quarter's time behind risk 1: The Slop Treadmill.
The three risks that matter
- 1The Slop Treadmill
21.6% of AI-delivered code comes back as a bug fix - vs 6.3% for everyone else.
21.6% ai work back as bugs6.3% other work back as bugs0.15 the gap- 01kxt2q “feat(api): fix flaky test”docs/architecture.md · re-opened after 11.375d
- 01kxt2q “fix(payments): fix flaky test”payments/ledger.ts · re-opened after 8.0417d
- 01kxt2q “fix(infra): tighten validation”infra/queue.ts · re-opened after 6.5833d
Fix Add a self-review step before AI diffs open a change, and refactor the top bounced paths once - the receipts name exactly which deliveries came back.
- 2Bus Factor One
“payments” has exactly one living maintainer. One goodbye email away from archaeology.
1 people who know it0.15 review knowledge spread3 files only one reviewsFix Put a second pair of hands on “payments” this quarter: pair rotation, review-buddy, and one recorded walkthrough.
- 3Unpaid Glue
One node bridges whole territories of the collaboration graph - nobody wrote that into a role.
16.41 bridge score (z)7 collaboration links- Ben Carter5 shared files
- Priya Sharma5 reviews
- Tom Reilly5 reviews
Fix Decide on purpose: promote the glue into an actual role, or de-load it deliberately. Either move beats waiting for the bridge to close for maintenance.
A
AI adoption vs outcomes
Last 180 daysNot enough AI-attributed delivery to judge the tradeoff yet - 5 AI-attributed PRs against a 20-per-cohort floor.
37.6%of commits (162 of 431)
6.5%of merged PRs (5 of 77)
| Cohort | PRs | Cycle p50 | Review wait p50 | Size p50 | Re-edited |
|---|---|---|---|---|---|
| AI-attributedlow sample | 5 | 144.0h | 74.0h | 241 lines | 89.1% |
| Everything else | 72 | 154.0h | 101.0h | 275 lines | 88.7% |
| Tool | Commits | PRs | Cycle p50 | Re-edited |
|---|---|---|---|---|
| Claude Coden=5 | 54 | 5 | 144.0h | 89.0% |
| Cursor | 33 | 0 | - | 85.7% |
| Devin | 11 | 0 | - | 100.0% |
| GitHub Copilot | 64 | 0 | - | 89.4% |
The treadmill: where AI lines get rewritten
- api/orders.ts508 AI lines · 41 touches · 01kxt2q refactor(api): address review nit
- api/users.ts383 AI lines · 39 touches · 01kxt2q feat(api): improve error message
- api/auth.ts406 AI lines · 36 touches · 01kxt2q test(api): handle edge case
- payments/ledger.ts435 AI lines · 30 touches · 01kxt2q fix(payments): tighten validation
- infra/config.ts402 AI lines · 29 touches · 01kxt2q test(infra): clean up types
B
Who understands the code
8 active areas2 Review is not catching it1 The argument zone1 The danger zone1 Unguarded
payments · atlasbus factor 115.0% of change32.3% AI touches100.0% reviewedno doc
Who knows it: Priya Sharma 98.7%
The danger zoneTop-band change (15.0% of shipped work), no doc at all, and a bus factor of 1.Write the doc before the date, not after it. This is the area where one person not showing up costs you the most, and it is the cheapest thing on this list to fix.
infra · atlasbus factor 216.1% of change42.0% AI touches77.8% reviewedno doc
Who knows it: Marcus Webb 63.6%, Ben Carter 35.8%
UnguardedTop-band change with only 77.8% of merged PRs reviewed.Add a second pair of eyes here before the date. One reviewer on the busiest area is the cheapest insurance you can buy.
The argument zoneTop-band change with 0.57 changes-requested round(s) and 0.57 human review comment(s) per reviewed PR.Settle the ownership or the design question before you hand this over. A handover into an unresolved argument fails, and the receiver inherits the argument, not the code.
web · atlasbus factor 214.5% of change38.9% AI touches100.0% reviewedno doc
Who knows it: Elena Kovac 72.5%, Tom Reilly 26.4%
Review is not catching itTop-band rework (92.2% of adding touches re-edited) alongside top-band review volume (3.00 reviews per merged PR).The process is running and not catching. Ask what reviews here actually check - and be ready for the answer to be "the wrong things". A handover here inherits the defect rate.
notifications · atlasbus factor 18.4% of change42.3% AI touches100.0% reviewedno doc
Who knows it: Diego Fernandez 100.0%
Review is not catching itTop-band rework (96.2% of adding touches re-edited) alongside top-band review volume (2.90 reviews per merged PR).The process is running and not catching. Ask what reviews here actually check - and be ready for the answer to be "the wrong things". A handover here inherits the defect rate.
Leave alone api (bus factor 2, 88.2% reviewed)
C
Who you cannot lose
3 people · €8,930-€26,790 to rebuildAisha PatelSole owner: tests
Diego FernandezSole owner: notifications
Priya SharmaSole owner: payments
What breaks without them
Priya Sharma€2,980-€8,930
1 area orphaned, 2 more at risk.
- payments hand to Marcus Webb
Aisha Patel€2,980-€8,930
1 area orphaned, 1 more at risk.
- tests hand to Yuki Tanaka
Diego Fernandez€2,980-€8,930
1 area orphaned, 1 more at risk.
- notifications nobody to hand to
D
The moves
- The Slop Treadmill: Add a self-review step before AI diffs open a change, and refactor the top bounced paths once - the receipts name exactly which deliveries came back.
- Bus Factor One: Put a second pair of hands on “payments” this quarter: pair rotation, review-buddy, and one recorded walkthrough.
- Unpaid Glue: Decide on purpose: promote the glue into an actual role, or de-load it deliberately. Either move beats waiting for the bridge to close for maintenance.
- payments: Write the doc before the date, not after it. This is the area where one person not showing up costs you the most, and it is the cheapest thing on this list to fix.
- Hand payments from Priya Sharma to Marcus Webb before either takes leave.
- Hand tests from Aisha Patel to Yuki Tanaka before either takes leave.
E
Method and seal
- AI-attributed
- A commit or PR carrying a co-author trailer, bot identity, agent branch or PR marker. Untagged AI use is not counted, so the share is a floor.
- Rework
- A line-adding change re-edited by a later commit within 14 days.
- Bus factor
- The fewest people who together hold 80% of an area's knowledge (authored and reviewed files).
- Protect list
- Everyone who is the sole owner or the only approver of an area.
- Rebuild cost
- 40-120 hours per orphaned area at the org's loaded cost.
- Seal
- sha256 over every byte of this document. Same data in, same hash out.
sha256
dee2a69c20f154a5e1fd21b8a083d0d9573bd28801e1983e6f167fde9b5d3b10inputs
ec64fca0a131846e7972094f7e69797cdcbff88c9e9011ccd17aa7581b5d296fengines
aiRuleset ai-tools@4 · pack audit_pack_v1 · recoveryEstimator knowledge_recovery@40-120h · restructure restructure_v1 · typology area_typology_v1as of
2026-09-25 UTCHow to read these numbers (3)
- Dependency fan-in reads 0: no import graph yet. Run a codegraph pass to unlock the quiet-landmine check.
- Lockfiles and build output count toward ownership, so a lockfile-heavy area reads as more owned than it is.
- "Only approver" means the only person who approved in this window, not the only one permitted to.