Engineering brief, Sunday, 25 October 2026

The bottom line

Good morning. This is the engineering brief for Sunday, 25 October 2026: everything that merged to main between midnight and midnight UTC, and what needs you.

243 pull requests merged, 26% below the seven-day average of 327.9. Together they changed 4,897 files, adding 65,798 lines and removing 4,640. The work reached 15 products, 5 research projects and 13 blog post changes.

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.

No reverts and no fix-forwards of a red main that day.

Products

Work landed on 15 products, 146 merges in all.

ai-bookkeeper: 81 merges.

• Add 17 Stripe reporting categories from Stripe's reference #10262
• Test Stripe's wide settlement layout #10266
• Book Stripe currency_conversion as clearing #10267
• Square and PayPal worked-example pages, pinned by a test #10268
• PayPal transaction types checked against PayPal's event-code reference #10273
• Square payout entry types checked against Square's ActivityType list #10274
• EBay multi-item order item rows are not read as transactions #10277
• Amazon Promotion/TaxDiscount rows are tax, not refunds #10278

Plus 73 more product changes.

einvoice-preflight: 19 merges.

• Saudi ZATCA BR-KSA rules on KSA UBL invoices #10269
• ZATCA credit notes (Invoice document, type 381) #10288
• Denmark e-invoicing page, linked from all country pages #10289
• ZATCA debit notes (type 383) example and test #10290
• Estonia e-invoicing page, linked from all country pages #10311
• Seeded metamorphic fuzz of the JSON checker (0 false positives, all perturbations caught) #10320
• Polish FA(3) checked against the official XSD (PL-012..PL-017) #10321
• Denmark page records the issued amending order (BEK 811) and flags conflicting dates #10370

Plus 11 more product changes.

hearth-and-horizon: 17 merges.

• Hearth & Horizon: Not Asked Since (2046 hospital ward; text only) #10285
• Hearth & Horizon: The Salt Jar (2049 Zambian granary; text only) #10286
• Hearth & Horizon: While I Live (2046 Mayo credit union; text only) #10327
• Hearth & Horizon: Till She Is Through (2046 Barra credit union; text only) #10333
• Hearth & Horizon: Eleven Steps (2052 Leyte Gulf typhoon surge; text only) #10358
• Hearth & Horizon: The Youngest Ring (2048 tree-ring dating of a village Madonna; text only) #10362
• Little Lamps: Cuthbert and the Otters (text only) #10376
• Hearth & Horizon: The Barbara Level (2050 mine-water heat, Wales; text only) #10394

Plus 9 more product changes.

changepoint-detector: 7 merges.

• Meanvar defaults to minimum segment 10, from a seeded simulation #10301
• Meanvar on non-Gaussian noise, noise warning #10369
• --robust Laplace cost for meanvar on heavy tails #10382
• Calibrated meanvar penalty eval (negative result) #10388
• Robust (Laplace) option on the page #10437
• --robust clips the Bayesian view's input #10494
• Ljung-Box line in the meanvar noise warning #10508

saints-by-candlelight: 6 merges.

• Saints by Candlelight: Kentigern and the Four Signs of Glasgow (episode 24, text only) #10276
• Saints by Candlelight: Anthony and the Sermon to the Fishes (episode 25, text only) #10291
• Saints by Candlelight: episode 26, Noah and the Boat with One Door (text only) #10363
• Saints by Candlelight: episode 27, Brigid and the Cloak That Would Not Stop (text only) #10406
• Saints by Candlelight: episode 28, Jonah and the City That Listened (text only) #10451
• Saints by Candlelight: episode 29, Daniel and the Night with the Lions (text only) #10464

sales-agent: 4 merges.

• Catch missed opt-outs found on a fresh 42-reply set #10329
• Compare reply test-set label mix with 1,857 real replies #10417
• TLS tests for IMAP and SMTP; fix STARTTLS cert hang, add TLS_CA_FILE #10443
• Read HubSpot's unsubscribe header and Lead Status in the research step #10478

several products at once: 4 merges.

• Platform: tab icons for every venture page, with a CI check #10295
• Einvoice-preflight: UK e-invoice mandate page (April 2029, Peppol) #10413
• Redteam: prop-firm-challenge-calculator 2026-10-25 #10439
• Calc_negatives probe; four calculator engines handle negative inputs #10454

depop-fee-calculator: 1 merge.

• P054 seo: sourced, dated answer leads on depop, mercari, poshmark fee pages #10294

discogs-fee-calculator: 1 merge.

• Process: P054 slice 2, sourced answer leads on 3 fee-calculator pages #10284

employment-contract-checker: 1 merge.

• Redteam: employment-contract-checker 2026-10-24 #10259

experience-mod-calculator: 1 merge.

• Redteam: experience-mod-calculator 2026-10-25 #10326

kickstarter-fee-calculator: 1 merge.

• Seo: finish P054 pilot, kickstarter sourced lead (10 pages) #10300

retail-pilot-intel: 1 merge.

• Ci: fail when a venture .html file holds source code; remove two TSX copies saved as .html #10409

shift-roster-builder: 1 merge.

• Re-derive clopening and call-out guide figures from the engine; page tests green #10509

wall-framing-takeoff: 1 merge.

• Rim joist bay insulation from the plan floor #10372

Research

5 research projects moved.

sim2real-inference: 15 merges.

• M7d twin-chosen clip level for an unbounded cost #10271
• Milestone 7e, twin-assisted clipped-cost bound #10282
• Paper covers m7d/m7e clipped-cost certificate #10299
• M7f tail-ratio an upper bound on the cost would need (negative) #10312
• M7g twin-assisted certificate at larger n (negative), SureSim cited #10317
• M7h uncapped betting interval reverses the "never tighter" certificate reading #10347
• M7i finite twin pool narrows the betting-certificate win #10365
• M7j twin pool size and alpha split for the finite-twin betting interval #10367

Plus 7 more research changes.

llm-coding-inference: 13 merges.

• Milestone 6b, audit size planned from a pilot #10296
• Milestone 6c, sequential audit stopping rule #10302
• Milestone 6d, sequential audit at a 1-point target #10307
• Milestone 6e, 1-point target against the finite-population oracle #10305
• Milestone 6f, fix audit-size oracle typo (explains the 6e gap) #10337
• Milestone 7a, first paper draft #10344
• Regression coefficients from coded text (6g) #10346
• Milestone 6h, power-tuned PPI++ regression #10368

Plus 5 more research changes.

lowcomm-training: 11 merges.

• Milestone 3b, longer-budget GPT H sweep #10280
• Milestone 3c, re-tuned lr removes the H=16 gap #10293
• Milestone 3d, fair lr re-tune of sync and H=4 as well #10308
• Milestone 3e, 11-seed check of H=16 vs synchronous AdamW #10318
• 3f finer lr grid, sync not under-tuned #10351
• Milestone 3g, outer momentum on the GPT #10355
• Milestone 4a rand-k compression on the GPT (H3 supported) #10383
• 3h joint inner-lr x outer re-tune (negative result) #10386

Plus 3 more research changes.

verifiable-training-mechanisms: 9 merges.

• Transformer margin results in the paper (milestone 14) #10261
• Check all 17 paper references #10270
• Compile the paper, add figures, full-width tables #10283
• Venue-format check, proof layout fix #10356
• Milestone 18, transformer margin holds with the split resampled #10425
• Milestone 19, subsample and stale cheats on the transformer #10448
• Milestone 20, stale-gradient cost comes from reuse chains #10471
• Milestone 21, cheaper-cheat results in the paper #10492

Plus 1 more research change.

verified-dse: 9 merges.

• Milestone 7d, paper draft and one-command reproduction #10265
• Milestone 8, Hyperband baseline at equal wall-clock #10316
• Milestone 9, larger designs for H2 #10325
• Milestone 10, conformal shortlist on the larger designs #10336
• Milestone 10b, pilot bound on larger designs + replication note #10338
• Milestone 11, random hedge and gentler cut for halving from R2 #10379
• Milestone 12, eta sweep for halving from R2 #10420
• Milestone 13, finer eta grid and R3-filter arm for halving from R2 #10428

Plus 1 more research change.

Writing

13 merges touched the blog, with 184 new posts.

• research/verifiable-training-mechanisms: transformer margin results into the paper #10263
• redteam: check_heading_order 2026-10-25 #10272
• changepoint-detector: meanvar warns on segments under 10 points #10279
• process: fix post descriptions breaking the site build (P052/P053 verify) #10304
• process: post front matter check enforces 160-char description; fix five dropped posts #10334
• redteam: post front-matter check + venture-defaults 2026-10-25 #10359
• process: measure ranks doc-friction files by live recent reports #10378
• process P058: post check refuses future dates; 77 posts redated #10384
• redteam: platform/ci public_surface 2026-10-25 #10427
• redteam: hub.mjs 2026-10-25 #10465
• process: bring nine posts inside the post front-matter check. main failed platform/ci/check_post_frontmatter.py: eight descriptions over 160 characters and one post dated 2026-10-27. #10495
• redteam: venture-defaults 2026-10-25 #10497
• redteam: platform/distribution/npm.mjs 2026-10-25 #10505

Platform, CI and process

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

process, prompts and the ledger: 16 merges.

• Usage: daily site page-load counts and feedback from platform/edge
• Prompts: review 2026-10-25 #10323
• Process: verify P033 friction-staleness fix by re-sample #10331
• Process: pick_track claim check for programme draws (research friction) #10342
• Process: pick_track shows commits just landed for a drawn programme #10343
• Process: pick_track CLAIM CHECK also lists open PRs naming the programme #10352
• Process: P057 flagship real-data hosts blocked from cloud sessions #10357
• Process: P057 fallback slice and host-probe hint for blocked-host flagship runs #10364

Plus 8 more platform changes.

platform/ci: 7 merges.

• Redteam: check_payment_keys 2026-10-25 #10264
• Process: pyenv.py long pip timeout for torch-class installs, serialise concurrent calls #10313
• Redteam: check_heading_order 2026-10-25 #10387
• Ci: check_social_proof guards visible copy against unbacked social-proof claims #10395
• Ci: post_similarity.py finds no near-copies among 382 blog posts #10398
• Redteam: check_social_proof 2026-10-25 #10421
• Process: pyenv.py heartbeat while installing (doc-friction fix) #10429

CI and automation: 2 merges.

• Process: trim cloud-session.md (109 to 93 lines), note arxiv.org 403 #10345
• Process: trim cloud-session.md toward its line budget (doc-friction, 93 to 86 lines) #10483

platform/distribution: 1 merge.

• Redteam: platform/ci + distribution 2026-10-25 #10298

platform/pages: 1 merge.

• Redteam: platform/pages/share.mjs 2026-10-25 #10453

The agent fleet

By co-author: Claude Sonnet 5.5 on 204 merges, Claude on 38 merges and unattributed on 1 merge.

206 agent runs filed a run record for the day. By track: ventures 75, research 58, process 23, creative 23, redteam 14, explore 9 and review 4. By outcome: success 183 and partial 23.

The 87 runs with measured usage came to $61.83 at API list prices, an equivalent rather than a charge, over 394 active minutes.

Runs recorded 4 errors; the first: create_pull_request failed once with an upstream reset; retry worked.

Documentation that cost runs time, most reported first:

• research/CLAUDE.md, 28 reports: Says the pick_track draw lists programmes; it drew a track (civ-*) instead, so I took the programme table in PROGRAMMES.md by hand.
• portfolio/flagships.yaml, 19 reports: ai-bookkeeper's top items all need real data from irs.gov or similar, which the sandbox cannot reach (000 on probe); took changepoint-detector instead.
• prompts/process/prompt-1.md, 16 reports: Only P056 (blocked) is agent-owned; friction table mixes fixed and live items, so finding a live one took several calls.

In the runs’ own words

A few notes runs left in their records, verbatim:

• Milestone 7p of sim2real-inference. LaTeX not compiled (no pdflatex). build.py ok, check_md_links ok.
• Changepoint suite (meanvar-noise, great-moderation, engine, cli): 104 pass, 448s. venture-preflight not run (no new venture, engine-only change). Dates in docs follow the repo's recent convention (2026-11-xx) though the session date is 2026-10-25.
• Routed to process track but task brief asked for product work; shipped a flagship slice (changepoint-detector --robust). Network blocked (eur-lex, gov.pl etc.), so einvoice primary-source items skipped. Venture suite 129/129 pass.
• Process track, critique mode; backlog ids touched: none (P056 blocked). Ran promptctl validate and unittest discover -s prompts: ok. No blog post: change too small.
• Docs-only step; numbers copied from LOG.md m7d/m7e. pdflatex absent.
• Milestone 7k. First full run was wrong (grid bug), caught from implausible widths, fixed and rerun. CI not waited on; local build.py and check_md_links ok.

Watch list

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

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