Engineering brief, Monday, 26 October 2026

The bottom line

Good morning. This is the engineering brief for Monday, 26 October 2026: everything that merged to main between midnight and midnight UTC, and what needs you.

248 pull requests merged, 21% below the seven-day average of 315.4. Together they changed 1,605 files, adding 97,233 lines and removing 864. The work reached 12 products, 7 research projects and 22 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: 6 merges repaired a red main or a failing check without reverting.

• process: fix post front matter red on main (shift-roster-ilp-check) #10558
• process: fix 13 blog front-matter faults on main #10567
• process: fix front matter on 8 blog posts (check_post_frontmatter was red on main). check_post_frontmatter.py reported 12 faults on main; the future-dated posts topped the feed and long descriptions break the Pages build. #10586
• process: fix post front matter red on main (check_post_frontmatter) #10588
• process: fix front matter on two posts failing check_post_frontmatter on main #10628
• process: shorten two post descriptions over 160 chars (main red on post front matter check) #10643

Products

Work landed on 12 products, 142 merges in all. New in the catalogue: vsme-report-builder.

ai-bookkeeper: 58 merges.

• Read Triodos-style '>NN' MT940 subfields; Dutch MT940 sample sweep #10517
• Read D/C flag columns and Label descriptions #10519
• Sweep OCA xlsx/xls bank fixtures (no fault), pin two shapes #10520
• Redteam: ai-bookkeeper plaid reader 2026-10-26 #10521
• Note when every date in a file reads either way; sweep OCA xlsx/xls fixtures #10522
• Keep Narration and Umsatztext beside the payee #10529
• Redteam: ai-bookkeeper/estimatedtax 2026-10-26 #10531
• Put the OFX ledger balance on the right line of a newest-first file #10532

Plus 50 more product changes.

wall-framing-takeoff: 17 merges.

• Drywall end joints snapped to studs (jointsOnStuds) #10585
• Page control for drywall jointsOnStuds #10592
• Browser click-through for the Wall end joints select #10594
• Ceiling drywall joints snap to joists #10599
• Page control for ceiling joist spacing #10604
• Insulation option on the page #10631
• Insulation bags in the budget #10636
• Drywall ceiling joints read the roof's ceiling joists (ceilingJoistsFromRoof) #10658

Plus 9 more product changes.

changepoint-detector: 14 merges.

• --dispersion penalty for meanvar on heavy tails and clustered volatility #10511
• Growth-change range on clumpy counts re-checked on fresh seeds #10584
• Real count series with a documented break (UK seat-belt law) #10623
• Pin Nile 1898 dam break under four settings #10633
• Check plain Poisson on coal-mining disasters #10637
• A second dispersion round does not cut GARCH false alarms #10653
• GARCH-filter experiment for meanvar (negative result) #10677
• --bootstrap penalty for meanvar #10678

Plus 6 more product changes.

einvoice-preflight: 13 merges.

• Luxembourg bill 8815, transitional dates, thresholds #10552
• Malta e-invoice rules page (B2B voluntary, Peppol for public buyers) #10582
• Bulgaria page adds reported scope and penalty start (draft, no consultation outcome yet) #10593
• Escape output in the India GST page #10651
• Austria mandate page (B2G only, B2B voluntary) #10673
• Finland mandate page #10674
• Sweden mandate page #10679
• Turkey e-Fatura/e-Arşiv mandate page #10683

Plus 5 more product changes.

hearth-and-horizon: 13 merges.

• Hearth & Horizon: The Folded Paper (text only) #10523
• Hearth & Horizon: The Book Says Twenty-One (For Other Minds, text only) #10535
• Hearth & Horizon: The Four Minutes (text only) #10541
• Hearth & Horizon: The Thirty-Minute Cycle (text only) #10548
• Hearth & Horizon: The Last Barge (text only) #10564
• Hearth & Horizon: The Eighteenth Place (text only) #10572
• Hearth & Horizon: The Clear Reading (text only) #10607
• Hearth & Horizon: Three Plain and a Bob (text only) #10615

Plus 5 more product changes.

shift-roster-builder: 11 merges.

• Independent ILP cross-check of open hours (35 h, not 36) #10557
• Coverage plateau pass closes the last open hour on 12 of 12 seeded weeks #10580
• Branch on the open-hours price; floor equals ILP optimum on 10 of 12 weeks #10589
• 10x coverBranch search leaves the last floor gaps unchanged #10598
• More coverage plateau walks close a 5 h search gap #10602
• Cover-bench and a measured negative on three-person chains #10627
• Exact large-neighbourhood repair of open hours (10x28 week 7 and 12x30 weeks reach the ILP optimum) #10645
• CoverRepair ends on a node budget (deterministic) #10657

Plus 3 more product changes.

saints-by-candlelight: 8 merges.

• Saints by Candlelight: Martin of Tours and the Half a Cloak (text only) #10530
• Saints by Candlelight: Seraphim of Sarov and the Bear in the Clearing (text only) #10539
• Saints by Candlelight: Samuel and the Voice in the Night (text only) #10555
• Saints by Candlelight: Rebekah and the Camels at the Well (text only) #10568
• Saints by Candlelight: Jacob and the Ladder at Bethel (text only) #10613
• Saints by Candlelight: Tobias and the Angel on the Road to Media (text only) #10620
• Episode 36, Naaman and the Seven Dips in the Jordan #10713
• Episode 37, Moses and the Bush That Was Not Consumed (text only) #10749

quick-pick-survey: 3 merges.

• Selectors measured at budgets up to 40x the list length #10583
• Pool cap costs no rank accuracy at deep budgets #10587
• Hybrid selector holds up against an inconsistent respondent #10722

sales-agent: 2 merges.

• Recognise DE/FR/ES/IT/PT/NL auto-replies and opt-outs in reply triage #10563
• Strip non-English 'wrote:' quote headers (mailgun/talon real emails, 11 of 12 clean before) #10603

job-application-tracker: 1 merge.

• Redteam: job-application-tracker 2026-10-26 #10666

photo-editing: 1 merge.

• Opt-in closed-eye and face scoring via an external detector command #10543

several products at once: 1 merge.

• Ventures: fix 15 suites left red by the inline favicon sweep. Tests allow the SVG xmlns and skip the icon tag; build scripts for estimate-maker, sales-cadence-builder and self-hosted-saas-stack emit it. #10527

Research

7 research projects moved. Started that day: codesign-aero-cant-poshold.

llm-coding-inference: 18 merges.

• Milestone 6m, HC3-type leverage for PPI++ at small audits #10513
• Milestone 6n, width of valid PPI++ intervals #10550
• Milestone 6n, interval width of valid PPI++ #10551
• Milestone 5, first real LLM coder #10561
• Paper draft 1 (regression, small audits, real LLM) #10629
• Milestone 5b, LLM coder size vs PPI++ width #10639
• Milestone 5c, calibrated cut on charged labels #10640
• Milestone 5d, break-even accuracy of a synthetic coder #10648

Plus 10 more research changes.

sim2real-inference: 12 merges.

• M4c cluster bootstrap of SIMPLER intervals #10512
• M4c, check SIMPLER eggplant values against the PDF #10514
• M4e, SIMPLER calibration with a held-out task or policy #10516
• M4f, real trials per policy to rank as well as the SIMPLER twin #10524
• Sync paper abstract to SIMPLER 4c/4e results #10525
• M7q, our betting construction in SureSim's stated setting #10597
• M7s, SureSim's public repo holds no data or code #10610
• Paper draft builds (m7t) #10638

Plus 4 more research changes.

verified-dse: 11 merges.

• Milestone 15, larger non-arithmetic designs #10526
• Milestone 16, pilot gate on 20 designs #10560
• Milestone 17, pilot size and held-out threshold #10577
• Milestone 18, which factor of the R2 score mis-ranks #10606
• Milestone 19, learned score vs hand-made ADP #10622
• Milestone 20, multi-fidelity regression surrogate #10624
• Milestone 21, exploration term vs learned surrogate #10630
• Sync paper and reproduce.sh with milestones 17 to 21 #10652

Plus 3 more research changes.

verifiable-training-mechanisms: 9 merges.

• Milestone 23, paper compiles with FL citations #10518
• Milestone 24, path-length predictor for H6 #10528
• Milestone 25, fitted correction to the path-length predictor does not transfer #10545
• Milestone 26, path-length predictor C does not carry to subsample and stale cheats (H6) #10573
• Milestone 27, cheap-gradient cheat damage is set by error direction, not size #10576
• Milestone 28, subsample damage lies in the orthogonal error component #10591
• Milestone 29, curvature of the subsample cheat's error (H6) #10611
• Milestone 30, paper H6 section gains milestones 27-29 #10709

Plus 1 more research change.

lowcomm-training: 8 merges.

• Milestone 4d, knee of error-feedback top-k (draft) #10609
• Milestone 4e, quantised worker deltas #10664
• Milestone 4f, 3-bit and un-fed-back scaled sign #10680
• 4f, 2-bit error-feedback residual norm #10682
• Milestone 5a, median aggregation under an attacker #10671
• Milestone 4e, stochastic quantisation and error feedback #10687
• Milestone 5b, strong attacker vs mean and median aggregation #10685
• Milestone 5c, the median's sign-flip cost is a sign effect #10758

civ-levers: 1 merge.

• Milestone 2, contact that widens tolerance #10670

codesign-aero-cant-poshold: 1 merge.

• Cant buys yaw, not position; the lateral force breaks designs that ignore it #10575

Writing

22 merges touched the blog, with 177 new posts.

• vsme-report-builder: VSME Basic Module draft report builder #10510
• process: bring 16 blog posts inside the front-matter check (P058 follow-up) #10547
• process: run post front-matter check on post-only PRs; fix five posts #10556
• process: fix post front matter red on main (shift-roster-ilp-check) #10558
• process: fix 13 blog front-matter faults on main #10567
• redteam: platform/ci checkers 2026-10-26 #10574
• ai-bookkeeper: fixture sweeps saturated, steer next runs to real exports #10579
• process: fix front matter on 8 blog posts (check_post_frontmatter was red on main). check_post_frontmatter.py reported 12 faults on main; the future-dated posts topped the feed and long descriptions break the Pages build. #10586
• process: fix post front matter red on main (check_post_frontmatter) #10588
• process: main_health.py lists red CI checks on main in one command #10596
• redteam: platform/pages/writing.mjs 2026-10-26 #10608
• redteam: platform/ci/slug_dupe.py 2026-10-26 #10617
• process: fix front matter on two posts failing check_post_frontmatter on main #10628
• process: shorten two post descriptions over 160 chars (main red on post front matter check) #10643
• platform/ci: page_js_errors.mjs, headless-browser load sweep of all venture pages (0 of 1,326 with JS errors) #10655
• posts: trim three descriptions over 160 chars (main_health red) #10656
• process: point prompt-1 at main_health.py (critique: doc friction) #10669
• process: fix red check_post_frontmatter (seven post descriptions over 160 chars) #10694
• process: fix red check_post_frontmatter (changepoint FX post description) #10698
• redteam: sync_tree_counts 2026-10-26 #10704
• redteam: platform/ci post_similarity + publish_surface 2026-10-26 #10718
• process: validate runs the post front-matter check; fix four red posts #10742

Platform, CI and process

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

process, prompts and the ledger: 9 merges.

• Prompts: review 2026-10-26 #10540
• Usage: daily site page-load counts and feedback from platform/edge
• Redteam: check_input_bounds + check_social_proof 2026-10-26 (16/16 caught) #10578
• Process: promptctl backlog marks blocked items, points to Critique mode #10595
• Process: trim three flagship next: texts under the 2,000-char cap #10635
• Prompts: review 2026-10-26 #10650
• Prompts: review 2026-10-26 (E93 cadence check) #10719
• Prompts: review 2026-10-26 (run 21) #10720

Plus 1 more platform change.

platform/ci: 7 merges.

• Explore: post_rate.py prints blog posts per day #10618
• Ci: check in-page anchors and aria references resolve #10632
• Redteam: platform/ci/check_post_frontmatter.py 2026-10-26 #10665
• Redteam: platform/ci/unlanded_work.py 2026-10-26 #10681
• Redteam: check_storage_guard + check_duplicate_ids 2026-10-26 #10690
• Redteam: post_similarity 2026-10-26 #10724
• Redteam: platform/ci 2026-10-26 #10756

platform/pages: 4 merges.

• Redteam: platform/pages/share.mjs 2026-10-26 #10549
• Prompts: review 2026-10-26 #10590
• Redteam: platform/pages/share.mjs 2026-10-26 #10642
• Redteam: platform/pages/check.mjs 2026-10-26 #10689

CI and automation: 2 merges.

• Process: fix null pr in run record that broke validate; pyenv torch note #10619
• Process: post front-matter check in the verify checklist #10644

platform/deploy: 1 merge.

• Redteam: platform/deploy/sealedbox 2026-10-26. open() returns null on too-short input. #10621

platform/webcheck: 1 merge.

• Redteam: platform/webcheck/page-integrity.mjs 2026-10-26 #10757

The agent fleet

By co-author: Claude Sonnet 5.5 on 210 merges, Claude on 37 merges and unattributed on 1 merge.

208 agent runs filed a run record for the day. By track: ventures 78, research 58, redteam 21, creative 20, process 18, explore 7 and review 6. By outcome: success 184 and partial 24.

The 75 runs with measured usage came to $56.70 at API list prices, an equivalent rather than a charge, over 329 active minutes.

Runs recorded 7 errors; the first: HVP call passed per-parameter grad_outputs for a flat output (RuntimeError); fused CPU attention has no double backward (fixed with the math SDPA backend).

Documentation that cost runs time, most reported first:

• portfolio/flagships.yaml, 22 reports: ai-bookkeeper and wall-framing next: items nearly all need real outside data; took quick-pick-survey instead, whose items are simulations.
• research/CLAUDE.md, 21 reports: pyenv.py torch install took several minutes and printed nothing until done; the stop hook nagged about uncommitted run.out during the 20+ minute experiment.
• prompts/process/prompt-1.md, 14 reports: Only owner rows open besides blocked P056; friction table mixes fixed items (cite_check eprint/DOI) with live ones, so finding the live fault (16 posts failing check_post_frontmatter) took several cal.

In the runs’ own words

A few notes runs left in their records, verbatim:

• Explore idea: post_rate.py, posts per day and tag mix from docs/posts. 543 posts, 9 days.
• Found a harness defect: redteam.py fell back to unittest for pytest-style suites (zero tests run, all mutants escaped). Fixed with a guard and test; reran campaigns with pytest; added 4 tests in post_similarity and publish_surface suites. First campaign on…
• Project: Hearth & Horizon (For Other Minds shelf); piece slug the-book-says-twenty-one. Text only.
• Four campaigns x 8 mutants on platform/ci checkers (check_input_bounds, check_form_defaults, check_shared_link_clears, shorten_descriptions). 28 of 32 caught; the 4 escapes were each triaged equivalent with a reason. No real defects found, so no code or test changes.
• m4f: twin ranks policy pairs at 0.887; real trials match at m=11 per policy, agreeing with n*~10. Tests and homepage build ok locally.
• Did not rewrite flagships.yaml next: (item 2 sweep, no change to open items). Cloning ledger fixtures via blobless sparse checkout was slow (>2 min).

Watch list

• portfolio/flagships.yaml keeps slowing runs down; a clearer page there pays back on every run.

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