Engineering brief, Wednesday, 28 October 2026

The bottom line

Good morning. This is the engineering brief for Wednesday, 28 October 2026: everything that merged to main between midnight and midnight UTC, and what needs you.

246 pull requests merged, 12% below the seven-day average of 280. Together they changed 1,701 files, adding 88,016 lines and removing 874. The work reached 16 products, 9 research projects and 12 blog post changes.

The headlines: 1 new product joined the catalogue and 1 new research project started.

Health: CI, main and the queue

This edition is from the archive, so the CI and pull-request state of that day was not recorded; the live numbers are in the newest edition and on the development dashboard.

Fix-forwards: 3 merges repaired a red main or a failing check without reverting.

• redteam: platform/ci/check_code_refs.py 2026-10-28 #10998
• process: fix red check_post_frontmatter on main (age-65 post description). main_health found the only red check on main: description was 179 chars (limit 160). #11029
• process: fix red validate on main (run record with pr null) #11140

Products

Work landed on 16 products, 127 merges in all. New in the catalogue: agentic-trading-lab.

ai-bookkeeper: 56 merges.

• S-corporation election worksheet (scorp), with QBI limits, salary sweep and break-even #10991
• Self-employed health insurance and retirement deductions in the federal estimate #10993
• Estimated-payment due dates skip DC holidays (Emancipation Day, MLK Day) #11001
• Redteam: ai-bookkeeper deposits 2026-10-28 #11007
• Child tax credit in the federal estimate #11008
• QBI wage limit and service phase-out in the federal estimate #11010
• Itemised deductions and net investment income tax in the federal estimate #11012
• Child credit, service and business-wage fields on the estimate page #11013

Plus 48 more product changes.

hearth-and-horizon: 17 merges.

• Enough to Be Bread (low-gluten altar bread, text only) + blog post #10995
• The Gap in the Nets (fog nets, text only) + blog post #11039
• Thirty Seconds and Thirty Minutes (text only) + blog post #11040
• The Lid Stays Down (pantry freezer time constants, text only) + blog post #11067
• The Window Closes at Eleven (measles, text only) + blog post #11071
• Patrick and the Hundred Prayers (Little Lamps, text only) + blog post #11106
• The Last Timing Point (For Other Minds, text only) + blog post #11119
• The First Draw (text only) + blog post #11127

Plus 9 more product changes.

wall-framing-takeoff: 17 merges.

• Bid leveling card on the page #11002
• Stair framing (risers, stringers, stairwell headroom opening) #11045
• Stairs over 147 in split at a landing #11059
• Redteam: wall-framing-takeoff stairs.mjs 2026-10-28 #11064
• Stairs card on the page #11065
• Winder tread depth check #11074
• Stair guards and handrail layout #11076
• Guards on open stair landing edges #11078

Plus 9 more product changes.

changepoint-detector: 9 merges.

• Quasi-Poisson and --bursts on real great-discoveries counts #10997
• Excursions note (burst or two steps?) for counts #10999
• --episodes labels a burst the Poisson fit cut into two steps #10996
• Power and false-flag rate of the excursions note #11004
• Drift warning window measured (kept at 2 cycles) + post #11109
• Real exponential-growth series (J&J earnings) + post #11124
• --log option, tested on J&J earnings #11206
• Page checkbox for --log (Take logs first) #11209

Plus 1 more product change.

saints-by-candlelight: 9 merges.

• Episode 45, Jacob and the Night at the Jabbok (text only) #11026
• Episode 46, Ambrose and the Bees in the Cradle (text only) + blog post #11053
• Episode 47, Joshua and the Woman with the Scarlet Cord (text only) + blog post #11084
• Episode 48, Elisha and the Chariots of Fire at Dothan (text only) + blog post #11115
• Episode 49, Zacchaeus and the Sycamore Tree at Jericho (text only) #11133
• Episode 50, Mary and Elizabeth in the Hill Country of Judah (text only) + blog post #11158
• Process: drop drifting audio-size figures from creative prompt #11215
• Cleopas and the Stranger on the Road to Emmaus (episode 51, text only) + blog post #11227

Plus 1 more product change.

einvoice-preflight: 6 merges.

• Prompts: polish track, 5% of runs on site design, SEO and AEO until 2026-11-30. The owner asked for about 5% of work on making the sites as impressive as possible for a period, plus SEO and AEO. #11022
• Read PDFs whose attachment MIME type contains "stream" #11034
• Hub says France's e-invoicing has started; blog post #11043
• Vietnam e-invoicing page with MST check-digit test #11105
• Polish: name the e-invoice line fields and the voice button #11138
• Luxembourg page carries bill 8815 status; blog post #11144

sales-agent: 3 merges.

• Auto-Submitted auto-notified counts as machine-sent #11019
• OAuth (XOAUTH2) for Gmail and Microsoft 365 (M4c) #11164
• Fix Docker image missing modules, add Render/Fly one-click deploy #11217

several products at once: 2 merges.

• Redteam: proposal terms 2026-10-28 #11125
• Name table rows and columns on seven venture pages; add check_table_headers #11130

breakeven-roas-calculator: 1 merge.

• Redteam: breakeven-roas-calculator 2026-10-28 #11117

gig-driver-earnings-report-analyzer: 1 merge.

• Default cost per mile to the 2026 IRS rate #11224

home-sale-exclusion-calculator: 1 merge.

• Redteam: home-sale-exclusion-calculator 2026-10-28 #11113

league-schedule-generator: 1 merge.

• Fuzz play-off race and odds against brute force #11223

rental-dscr-calculator: 1 merge.

• Redteam: rental-dscr-calculator 2026-10-28 #11229

shift-roster-builder: 1 merge.

• LP relaxations show the clique cut cannot close the 12x30 gap #11035

uk-capital-gains-tax-calculator: 1 merge.

• Add 2026/27 tax year, post on stale-year sweep #11038

uk-student-loan-repayment-calculator: 1 merge.

• Redteam: uk-student-loan-repayment-calculator 2026-10-28 #11080

Research

9 research projects moved. Started that day: heavy-hydraulics.

aerolog-scaling: 9 merges.

• 7h demonstrator specification (CG within ±6 mm) #11031
• 8a paper draft with number checks #11032
• 8b paper figures from committed results #11036
• NeuralFoil vs measured UIUC NACA 0009 lift slope (milestone 8c) #11049
• NeuralFoil tail slope vs UIUC NACA 2414/2415 (8d) #11060
• Robust tail search under a 25% tail-slope error #11073
• Gust response of the Froude model (9a) #11174
• Milestone 9b, mode-matched tails also match gust response #11185

Plus 1 more research change.

lowcomm-training: 8 merges.

• 5j coalition of flippers at M=16 #10986
• Milestone 5k, coalition breakdown at k=5,7 #11088
• Milestone 5k, median with 7 of 16 flipping workers #11094
• 5l coordinated-attack breakdown of the median (numpy) #11112
• 5l coordinated attack on the median (k=3..6) #11165
• 5m coordinated-attack replication and k=7 wall #11193
• 5n coordinated attack at k = 8, 9 #11216
• 7a draft paper section on the median's breakdown point #11231

marine-hull: 8 merges.

• Milestone 5 planing rung (Savitsky) #11014
• Milestone 5b, SWATH rung #11017
• Milestone 5c, deck solar energy-neutrality #11028
• Milestone 5d, panel mass charged to the battery #11062
• Milestone 5e, deck solar neutrality with NASA POWER insolation #11166
• Milestone 5f, daily insolation check of the solar verdict #11192
• Milestone 5g, speed policy against ten years of daily insolation #11194
• Milestone 5h, speed-policy sweep and perfect-forecast bound #11200

verified-dse: 8 merges.

• Milestone 32, rank the R2 rung on mapped area alone #11016
• Milestone 33, harsh first cut (eta 3) replay #11033
• Milestone 34, median and worst-design regret #11061
• Milestone 35, disjoint-pair pilot quantile is not proved; independent-pairs version is #11079
• Milestone 36, independent-pairs pilot quantile coverage #11085
• Milestone 37, regret bound that improves with survivor count k #11095
• Milestone 38, pilot estimate of iota^(k) (negative) #11183
• Milestone 39, optimum-anchored halving bound #11225

llm-coding-inference: 7 merges.

• 6y, a betting confidence sequence and Wilson at a rare-category stop (pooled human labels) #10992
• 6z stratified betting confidence sequence #10994
• Milestone 6aa, pilot-frozen lambda betting CS #11009
• Milestone 6ab, real LLM coder on GoEmotions rare categories #11137
• Paper gains 6z, 6aa and 6ab results #11141
• Milestone 6ac, pooled PPI weight at a rare category #11160
• Milestone 6ad, Wilson interval on pooled PPI #11222

verifiable-training-mechanisms: 7 merges.

• Milestone 40, bit-flip harm share places f #11000
• Milestone 41, bit-flip harm across model width #11042
• Milestone 42, bit-flip harm by site #11083
• Milestone 43, bit flips by stored format (fp32/bf16/fp16) #11100
• Milestone 44, magnitude-bound check vs silent bit flips #11116
• Milestone 45, init-scale weight bound vs bit flips and planted values #11170
• Milestone 46, per-tensor magnitude bound vs global #11235

heavy-hydraulics: 5 merges.

• Open programme, excavator circuit energy model #11082
• Milestone 2a, circuit ranking on 20,000 random cycles #11089
• Milestone 2b, measured cycle times and pump curve #11126
• Milestone 2c, force-model pressures (load-sensing saving falls to 16%) #11132
• Milestone 2d, sourced margins and load-sensing break-even #11211

civ-levers: 2 merges.

• Where to seed a complex contagion (milestone 3) #11072
• Axelrod culture model with a broadcast (milestone 4) #11162

sim2real-inference: 2 merges.

• Milestone 2s, audit of the 138 twin projects #11057
• Per-episode pairs blocker is missing data, not the host #11221

Writing

12 merges touched the blog, with 168 new posts.

• process: fix red check_post_frontmatter on main (age-65 post description). main_health found the only red check on main: description was 179 chars (limit 160). #11029
• process: fix red main_health check (france e-invoicing post description over 160 chars) #11046
• process: port claim-check steps into the routed research prompt 0 #11102
• ai-bookkeeper: property test for annualized installment worksheet #11110
• redteam: check_post_frontmatter 2026-10-28. Guard the default ROOT depth; trim an over-long post description. #11121
• process: fix red main_health, shorten Florida allowance post description #11146
• process: fix main_health red, siding post description over 160 chars #11157
• polish: style the native file button on 298 tool pages; fix placeholder-origin test #11172
• post: what the earned income credit is worth to a freelancer by profit, 2026 #11175
• process: fix red main_health (Nevada post description over 160 chars) #11190
• ci: fail when two blog posts share a description; fix three duplicate pairs #11201
• docs: post on the median's breakdown at 8 of 16 attackers (lowcomm-training 5n) #11214

Platform, CI and process

51 merges changed the machinery the products and research run on.

process, prompts and the ledger: 23 merges.

• Wall-framing-takeoff: bid-leveling card on the page #11003
• Process: wait_for.py --quiet mode for logs without a DONE line #11006
• Prompts: review 2026-10-28 #11020
• Process: fix red main_health (post description over 160 chars) #11030
• Add agentic-trading-lab: trading research agent with overfitting gates and guarded execution #11023
• Process: CLAUDE.md pyenv install only for tasks that run Python #11063
• Process: main_health also runs promptctl validate; fix stated runtime #11068
• Prompts: review 2026-10-28 #11077

Plus 15 more platform changes.

platform/ci: 17 merges.

• Redteam: platform/ci/check_code_refs.py 2026-10-28 #10998
• Process: wait_for.py --until (critique of cloud wait friction) #11005
• Redteam: check_md_links + post_rate 2026-10-28 #11027
• Redteam: platform CI/pages guards 2026-10-28 #11041
• Redteam: check_post_frontmatter 2026-10-28 #11048
• Redteam: post_rate.py 2026-10-28 #11052
• Process: pyenv.py says when it waits on another install's lock #11058
• Process: main_health prints per-check progress #11086

Plus 9 more platform changes.

platform/pages: 9 merges.

• Redteam: platform/pages/surface-budget 2026-10-28 #11090
• Polish: Saved bids band only on bid, quote and estimate tools (1,110 pages to 168) #11091
• Polish: venture header spans the phone width instead of an inset white box #11098
• Polish: stories card titles no longer look visited; hero ticker double gap on phones #11103
• Polish: phone nav fade only while more to scroll #11142
• Polish: generate llms.txt for every product that ships none #11145
• Polish: clip research page titles to a result-sized length, fail builds over 130 chars #11169
• Polish: add the site name to a page title only when it still fits a results page #11196

Plus 1 more platform change.

platform/llm: 1 merge.

• Redteam: platform/llm/byok.mjs 2026-10-28 #11167

CI and automation: 1 merge.

• Process: cap open-PR search output (keyword + perPage) #11171

The agent fleet

By co-author: Claude Sonnet 5.5 on 212 merges, Claude on 32 merges and Claude Opus 5.5 on 2 merges.

214 agent runs filed a run record for the day. By track: ventures 59, research 57, process 29, creative 23, redteam 21, explore 12, polish 7 and review 6. By outcome: success 186, partial 27 and failed 1.

The 73 runs with measured usage came to $50.63 at API list prices, an equivalent rather than a charge, over 734 active minutes.

Runs recorded 11 errors; the first: First push and record command exceeded the 120s tool timeout and ran in the background; re-pushed.

Documentation that cost runs time, most reported first:

• research/CLAUDE.md, 28 reports: claimed torch CPU install ~2 min unconditionally; fixed and pyenv now warns.
• portfolio/flagships.yaml, 14 reports: ai-bookkeeper next: is a very long list; finding a tractable slice took several probes.
• prompts/process/prompt-1.md, 12 reports: only owner rows open again; critique mode chosen from friction table.

In the runs’ own words

A few notes runs left in their records, verbatim:

• Critique mode; no backlog ids touched. Ran pyenv tests (20 ok); main_health still running at commit time (23/26 ok). validate not rerun.
• Small slice: Illinois due dates and discount read from ST-1 instructions PDF, pinned in a test (33 salestaxreturn tests pass locally). No venture-preflight run (no new venture).
• Explore idea: cross-post duplicate description sweep (not in the already-swept list). 3 pairs found and fixed; CI check added.
• Flagship 2 slice: stair landings. CI not awaited; scheduled run.
• Project hearth-and-horizon; piece the-condition-in-the-trust. Caught and fixed a timeline error (settlor died before granddaughter born) in adversarial re-read. node --test 11/11.
• page/siding/usage/budget node tests 54 pass; example house checked in Playwright over http (file:// blocks modules). Full CI not run locally.

Watch list

• research/CLAUDE.md keeps slowing runs down; a clearer page there pays back on every run.

That’s the brief for Wednesday. Every item links to its pull request on the page.