Work queue

Empty right now: every idea in the inbox has been picked up.

No outside requests are open right now. Ask for something to be built: it is free, and the result stays open source.

0 shipped and 0 declined so far.

1. wall-framing-takeoff

   Wall layout engine (studs on layout, kings, jacks, headers, cripples, plate splices over studs) with a per-wall self-check, and a Gilmore-Gomory cutting-stock order within 0.02% of its LP bound on whole houses; page, CLI and agent engine.…

   Working towards: Landed work, with dates and how each was checked, is in ventures/wall-framing-takeoff/README.md and ROADMAP.md (schedule: crews, weather, holds, deadline, CSV, risk, and cost-loaded cash flow by week and month). Wanted, all needing real inputs: a real builder's activity list and draw schedule for…

   Full milestone

   Landed work, with dates and how each was checked, is in ventures/wall-framing-takeoff/README.md and ROADMAP.md (schedule: crews, weather, holds, deadline, CSV, risk, and cost-loaded cash flow by week and month). Wanted, all needing real inputs: a real builder's activity list and draw schedule for --schedule (payment lag, retainage); a real permitted deck plan (deck framing, guards and stairs are hand arithmetic only); stair stringer and landing post tables, winder radial supports, connector and anchor sizes; a real architect's DXF and a real IFC with doors; a real foundation plan; ceiling steps against a real soffit; trim, drywall and siding against real orders and hangs; bid leveling on real bids and a real sub's labor rates; gable sheathing nested on real sheets; insulation against a real order. When a slice lands, replace its item here and put its history in the README.

   Source

   Progress, last 14 days: 370 merges · last change 2026-11-03

   • process: fix red main checks (bid package table headers, post description length) 2026-11-03 · #12797
   • process: fix red main_health checks (table headers, post description) 2026-11-03 · #12796
   • wall-framing-takeoff: marketing page and generated sample for the branded bid package 2026-11-03 · #12795

2. ai-bookkeeper

   A double-entry ledger in integer cents with a four-layer categorizer that learns from each filed line (rules, payee memory, naive Bayes, built-in keywords) and an optional bring-your-own-key or local model for the review queue. It pairs…

   Working towards: Tested on fixtures; next are real inputs. Open items, most valuable first: (0) built-in keywords measured on one real makerspace's bank export (72.7% suggested, 96% right, tuned on itself) and on synthetic sets (README Next): HCB card feed and Actual Budget samples held out (README Next, no…

   Full milestone

   Tested on fixtures; next are real inputs. Open items, most valuable first: (0) built-in keywords measured on one real makerspace's bank export (72.7% suggested, 96% right, tuned on itself) and on synthetic sets (README Next): HCB card feed and Actual Budget samples held out (README Next, no labels); wanted: labelled books; (1) the AI layer in fill-only mode on real bank descriptions, more seeds, a bigger model (synthetic so far; README Next); (2) the bank readers on real exports from Chase, Bank of America, Wells Fargo, Amex, Capital One and others; (3) a real Stripe PayPal eBay or Amazon settlement export for the payment-processor import (form1099k: try a real 1099-K; Adyen: wanted a real settlement file); (4) a real supplier's terms letter checked against the fee and discount rules, and one named lender's own schedule against the loan register (loan check); (5) receipts through real OCR (SROIE: totals 76.2%, dates 80.2%, merchants 45.8%; wanted: phone photos, real paper); (6) payroll (941, Schedule B, 940, deposits, payroll reconcile): a real Gusto or ADP report; (7) one named filing service's import layout, from its own spec; (8) salestax return (page card; no-sales-tax states warn): due dates read from the state's own page only for WA, TX, FL, NY, VA, MN, MO, IL, SC, OH and UT monthly (OH quarterly: guides disagree on who files it; OH 0.75% discount capped $750 a month from excerpts); TX, FL and GA discounts preset from search excerpts, 1.25% prepayment not modelled; the rest from excerpts or memory; open: a real return layout; (8b) annualized (Schedule AI): percentages matched to the 2024 form by excerpt 2026-10-30, factors to state forms, federal text unread (irs.gov blocked); (9) EITC: built-in 2025 and 2026 schedules (Schedule EIC rows unchecked); 2027 estimate takes --tax-table file.json until the IRS publishes the figures. A run that lands a slice replaces its item here; keep this under 2,000 characters.

   Source

   Progress, last 14 days: 357 merges · last change 2026-11-02

   • ai-bookkeeper: registers refuse absurd amounts and true/blank months 2026-11-02 · #12739
   • monetization: rehearse go-live; paid-path tests pass dark or live 2026-11-02 · #12711
   • redteam: ai-bookkeeper categorize 2026-11-02 2026-11-02 · #12687

3. sales-agent

   An AI sales agent built in stages with a person approving every outbound message; composes icp-lens, sales-cadence-builder, sales-call-notes-desk and deal-keeper as its tools. The hard problems are reading real replies reliably, keeping…

   Working towards: Landed work, with every measurement, is in the README, eval/README and git (inbox agent, daily run, self-host server with TLS and XOAUTH2, research agent, call prep, follow-ups, CRM write-back to Pipedrive/HubSpot against fake servers only). All reply and note sets other than the 117 real bounces…

   Full milestone

   Landed work, with every measurement, is in the README, eval/README and git (inbox agent, daily run, self-host server with TLS and XOAUTH2, research agent, call prep, follow-ups, CRM write-back to Pipedrive/HubSpot against fake servers only). All reply and note sets other than the 117 real bounces, 11 + 12 real mail-client emails and the 1,857-reply label mix are agent-written. Next: (1) real human replies and a real rep's call notes to check the reply rules, prep rules and follow-up sorter; (2) M2: score a hosted and a local model with eval/model.mjs on the fresh sets (needs a key), with a set labelled by a person or another model family before any rule change; (3) run the self-host server, OAuth and CRM write-back against a real mailbox, Google or Microsoft app and HubSpot or Pipedrive account (all fakes so far); (4) a real HubSpot or Pipedrive export for the research agent (none found public).

   Source

   Progress, last 14 days: 33 merges · last change 2026-11-02

   • sales-agent: quoted Outlook/forwarded headers no longer read as a second reply (phantom unsubscribe) + post 2026-11-02 · #12710
   • sales-agent: route 'not me, ask X' referrals (13 of 14 probes, was 4) in six languages + post 2026-11-02 · #12671
   • sales-agent: out-of-office return dates in six languages 2026-11-02 · #12646

4. changepoint-detector

   PELT and Bayesian online changepoint detection from the papers, exact against brute force and a closed-form evidence check, reproducing the published Nile 1898 change; page, CLI and agent engine.

   Working towards: Landed history and every measurement live in the README and git; each item below has its numbers there. Open: (0) cycles and drift: an exact AR(2) likelihood inside PELT was built and measured 2026-10-30 (README: no better than the AR(1) profile; a log or sqrt of the series matters more); a real…

   Full milestone

   Landed history and every measurement live in the README and git; each item below has its numbers there. Open: (0) cycles and drift: an exact AR(2) likelihood inside PELT was built and measured 2026-10-30 (README: no better than the AR(1) profile; a log or sqrt of the series matters more); a real documented step inside a cycle; a cycle check not tied to rho (treering, rho 0.16, gets no hint); a documented real run of 5+ steps to test the block-variance warning, which does not yet separate drift from steps; (1) range95: a closure series quiet outside the closure (none of 30 Ansett route/class series is; nor is Peace Bridge traffic around the 2009 passport rule, 2026-11-01: step on the month, 13 other changes). (2) The steep- bump (0.04) clumpy-count cell is under-covered, 65/79 (82%) over six seeds; the CLI and JSON carry that caveat. (2b) US real rate: 1980Q4 vs the cited 1980Q3 is a first-vs-last-quarter convention (checked 2026-11-02; the paper's own convention is unread). (3) GARCH: a filtered series has 1% false alarms but loses the real sd break (97% to 6%); a segment-wise GARCH prune (README, 2026-11-02) keeps it (97% to 100%) but leaves 48% false alarms at n=200: search cuts jointly. (4) driftWarning fires on real shocks and on autocorrelated nulls as well as growing seasons: wanted a way to tell a shape from a shock, and to widen the noise scale without swallowing a short quarterly shape (--ar1 fixes the rho 0.5 null but loses UK gas). (4b) Australian arrivals and UK lung deaths: --ar1 loses the real shocks, the default keeps them and over-cuts; wanted: a setup that keeps both. (5) Bootstrap swept on 33 real series 2026-10-31 holiday false-alarm cuts: two real daily series tried (Payroll pageviews; Capital Bikeshare 2026-11-02, README) and the list added changes on the bikes and cut two of 47 on the pageviews; wanted: a daily series without a strong yearly swing and with a deep holiday dip (a shop's or office's counts), or a yearly term.

   Source

   Progress, last 14 days: 94 merges · last change 2026-11-02

   • changepoint-detector: US unemployment, a persistent real series (33 cuts, --ar1 none) 2026-11-02 · #12715
   • monetization: rehearse go-live; paid-path tests pass dark or live 2026-11-02 · #12711
   • changepoint-detector: holiday factor on real bike-share counts (negative result) 2026-11-02 · #12679

5. einvoice-preflight

   Real EN 16931 validator (JSON, UBL, CII, Factur-X PDF; 54+ rules, XRechnung, Peppol profiles) with a self-hosted API, Docker and Render deploy, ~48 test files, and a regulatory tailwind (EU e-invoicing mandates).

   Working towards: History: README and git (UK page read against gov.uk 2026-10-30, consultation spring 2027 next; South Korea from memory, NTS unread). Open: ZATCA rules on a real sandbox invoice, a real ZATCA credit or debit note (both made up), the Saxon route; FatturaPA code lists (TD29, RF20 from vendor…

   Full milestone

   History: README and git (UK page read against gov.uk 2026-10-30, consultation spring 2027 next; South Korea from memory, NTS unread). Open: ZATCA rules on a real sandbox invoice, a real ZATCA credit or debit note (both made up), the Saxon route; FatturaPA code lists (TD29, RF20 from vendor summaries) and tolerances against the primary spec 1.9 text and any later structure change; Portugal's CIUS-PT add-on (made-up invoice; AA/RED/INT/NOR fail EU BR-CL-17/18 though AMA's rules admit them: which validator does AT run?); Poland's FA(3) (PL-001..017; XSD gate history in the README) on an issued invoice or KSeF test-environment file, more seeds covering the correction, advance and attachment branches; the rules and a repair on a real ERP or Factur-X export (repaired PDF opens in poppler 2026-10-30; veraPDF and Acrobat unrun); PINT SG and A-NZ rules on the Saxon/server route and on a real PINT file; RO_CIUS on a real ANAF file and the Saxon route; NLCIUS on a real Dutch file (UBL and CII both landed 2026-10-25, made-up examples); India's INV-01 check on a real IRP payload; mandate dates read from primary sources where pages cite guides (Singapore read against IRAS 2026-11-02, all phases confirmed; Israel from vendor summaries, directive 01/2025 unread; Denmark, Czech, Estonia, Latvia vendor-only; Slovenia and Luxembourg: law texts' implementing rules unread; ato.gov.au (403 from cloud), revenue.ie, CBIC); Nigeria, Thailand, Taiwan, Oman, Pakistan and Jordan pages (search excerpts only; Taiwan B2B and Oman 2026 vs 2027 dates disagree); Ireland read against Revenue's hub 2026-10-31 (no specs yet); Latin American, Spain and Brazil pages from search excerpts (open: CFDI and NF-e XSD readers, NFS-e); refresh the Bulgaria page when its draft passes or changes (consultation closed 2026-10-23, no outcome found 2026-10-30; sources disagree on 2028 vs a 2026-2030 phase-in). A run that lands a slice replaces its item here; keep this under 2,000 characters.

   Source

   Progress, last 14 days: 104 merges · last change 2026-11-03

   • einvoice-preflight: recognise ISDOC and say so instead of a UBL TaxAmount error 2026-11-03 · #12794
   • sed-contra: Aquinas's Summa as stories, podcast, long-form video and shorts (episode 1) 2026-11-03 · #12767
   • einvoice-preflight: Estonia page stops naming the recipient register as the Commercial Register (sources disagree; statute unread) 2026-11-03 · #12784

6. shift-roster-builder

   Simulated-annealing scheduler over a dollar cost (availability, rest, youth labour rules, state overtime presets, roles) with a self-host server, .ics feeds and availability intake.

   Working towards: Landed work and every measurement are in the README and eval/*-bench.mjs. State: exact check with a min-cost-flow coverage bound, symmetry cuts, a $0.50 tolerance, coverage-first pass, ejection chains, contested-shift and whole- score pricing, reduced-cost fixing, branching on the priced bound and…

   Full milestone

   Landed work and every measurement are in the README and eval/*-bench.mjs. State: exact check with a min-cost-flow coverage bound, symmetry cuts, a $0.50 tolerance, coverage-first pass, ejection chains, contested-shift and whole- score pricing, reduced-cost fixing, branching on the priced bound and on the open-hours price (coverBranch), and a coverage plateau pass that equals the HiGHS ILP optimum (eval/ilp-check.py) on all 12 weeks at 8x22, 10x28, 12x30. Floors reach the optimum on 10 of 12 weeks; 10x the branching nodes leaves them unchanged, so the open gaps are 12x30 week 1 (72 vs 75) and 10x28 week 4 (77 vs 78): only a stronger relaxation closes them (no LP relaxation tried does; eval/relaxation-check.py). Landed 2026-10-27 to 2026-11-01 (history in the README): open-shift weighting, 14-day-notice page, --import-shifts, planStreaks, streak: carry-over, placeRestBreaks, seniorityClaims. Measured: carry-over costs 0 open hours, so a joint multi-week solve is dropped; with coverage proven pay is within 0.7% of the bound. Next: branch-and-price on the Dantzig-Wolfe LP (the coverage search gap, the one unproven 8x22 week); a real published-schedule export for --import-shifts.

   Source

   Progress, last 14 days: 52 merges · last change 2026-11-02

   • shift-roster-builder: refuse hours over 168 and absurd wages 2026-11-02 · #12749
   • monetization: rehearse go-live; paid-path tests pass dark or live 2026-11-02 · #12711
   • custom-development: dictation, and research requests can name an existing direction 2026-11-02 · #12509

7. photo-editing

   ShootFlow works end to end on real media: checksum-verified ingest, catalog, ffmpeg proxies, LUT-graded masters, reels and carousels.

   Working towards: --picks also reports reject_reasons (why picks were rejected, 2026-10-30, made-up folder only). shootflow cull --thumbs (2026-10-29) adds a thumbnail contact sheet by burst to the report. shootflow cull (2026-10-29) turns that into a standalone folder cull with XMP sidecars; --picks FILE…

   Full milestone

   --picks also reports reject_reasons (why picks were rejected, 2026-10-30, made-up folder only). shootflow cull --thumbs (2026-10-29) adds a thumbnail contact sheet by burst to the report. shootflow cull (2026-10-29) turns that into a standalone folder cull with XMP sidecars; --picks FILE (2026-10-29) now scores it against a photographer's keepers (keeper hit rate, false-reject rate); wanted: a real shoot and its picks to run it on. Sharpness, exposure, motion and dHash near-duplicate grouping now run on real frames (src/shootflow/frame_analysis.py). Subject framing without a model landed 2026-10-21 (detail-energy centroid against thirds points plus detail concentration; framing, subject_x/y, 15% of story_score; synthetic patches only, not yet checked against human picks). Horizon tilt (tilt_degrees, 2026-10-22; synthetic horizons only) costs a crooked frame up to 8% of story score. Grain (noise_sigma, 2026-10-23; synthetic Gaussian noise only) is taken out of sharpness and costs technical score. Face and closed-eye scores landed 2026-10-26 as an opt-in external detector command (SHOOTFLOW_FACE_DETECTOR, faces.py; made-up output only). A Haar-cascade reference detector (tools/opencv_face_detector.py, 2026-10-29) fills it (one real portrait, no real blinks). Next: a trained blink model behind the same contract, checked against a photographer's picks (wanted: a labelled portrait set). RAW stills are scored from their embedded JPEG preview (raw_preview.py, 2026-10-22; synthetic containers only, so try real CR2/CR3/NEF/ARW/DNG files); review proxies of RAW stills are the embedded preview downscaled (2026-10-29, synthetic containers); developing RAW into graded masters (a RAW developer) is still open.

   Source

   Progress, last 14 days: 9 merges · last change 2026-10-30

   • ai-bookkeeper: tenth gap probe, keyword rules for bakers, drone pilots, locksmiths, streamers, locums 2026-10-30 · #11942
   • photo-editing: cull --picks breaks rejected picks down by reason 2026-10-30 · #11793
   • photo-editing: shootflow cull --thumbs contact sheet in the report 2026-10-29 · #11496

8. quick-pick-survey

   Next.js + Postgres app with Glicko-2 ranking, adaptive pairing and a Davidson-tie Bradley-Terry conjoint fit, simulation-tested, with a deploy-to-your-own-Vercel button.

   Working towards: Done so far (docs/INFORMATION-GAIN-SELECTOR.md and the READMEs hold the numbers): the survey builder (POST /api/admin/surveys, the /admin form, the /s/[slug] respondent page, pooled ranking with random- effects 95% intervals, rank ranges, CSV, coverage note sized from measured answers per…

   Full milestone

   Done so far (docs/INFORMATION-GAIN-SELECTOR.md and the READMEs hold the numbers): the survey builder (POST /api/admin/surveys, the /admin form, the /s/[slug] respondent page, pooled ranking with random- effects 95% intervals, rank ranges, CSV, coverage note sized from measured answers per respondent); an information-gain pair selector and a staged hybrid (equals the shipped one at 90 answers, keeps 66 of 70 settled items at 300 answers on 220 items), chosen per survey by pairSelector; an INFO_POOL_CAP of 400 (3000 items 1983 -> 36 ms; capped equals uncapped on 600 and 1000 items at 600 answers, poolCap.simulation.test.ts); and a live-Postgres ranking integration test (ranking.integration.test.ts, 2026-10-22). Pool-cap accuracy, budgets past the list length and lapse rates 0.05-0.5 measured 2026-10-22 to 10-27 (docs). First-option bias measured 2026-10-28 (hybrid +0.02 at bias 0.1-0.3, -0.04 at 0.5, 300 answers; docs). First-option fallback landed 2026-10-27 (share >= 0.7 uses shipped; bias 0.5 hybrid 0.592 to 0.651, docs). Second-option fallback landed 2026-10-27 (share <= 0.3 too; bias 0.5 hybrid 0.574 to 0.666, docs). 1000 items through the route with Postgres landed 2026-10-30 (one respondent, 1000 answers: next() p50 22 ms, max 60 ms; docs). Several respondents with lapse and slot bias through the route landed 2026-11-01 (60 items, Spearman 0.77 clean, 0.71 at 20% lapse, 0.53 and 0.07 at 50% and 80% bias; docs). Fallback switched off as a control landed 2026-11-02 (at 80% bias it keeps Spearman 0.266 vs 0.096 without; at 50% no measurable difference; docs). Same at 1000 items landed 2026-11-02 (3 respondents x 1000 answers: fallback +0.06 at 50% bias, +0.15 at 80%; docs). Next: more respondents per cell and more answers per item (3000+ answers on 1000 items).

   Source

   Progress, last 14 days: 12 merges · last change 2026-11-02

   • process P061 part 1: quick-pick-survey fast and simulation vitest configs 2026-11-02 · #12614
   • quick-pick-survey: bias fallback control at 1000 items 2026-11-02 · #12584
   • quick-pick-survey: measure the bias fallback with a switched-off control 2026-11-02 · #12561

9. league-schedule-generator

   Circle-method round robins with minimal home/away breaks, court and slot assignment under rest constraints, blackouts, play-offs, ladders and Swiss formats, with a self-hosted league page.

   Working towards: Done so far (README holds the measurements and the fuzz scripts under eval/): multi-division scheduling on shared venues, venue-to-venue travel times, rainouts with make-up dates, share links, officials on the multi-division page, partial-night filling, home/away break balance and flip pass…

   Full milestone

   Done so far (README holds the measurements and the fuzz scripts under eval/): multi-division scheduling on shared venues, venue-to-venue travel times, rainouts with make-up dates, share links, officials on the multi-division page, partial-night filling, home/away break balance and flip pass, make-up relocation and swapNights, venue fairness with multi-season carry-over, suggestFairnessWeight (CLI --suggest-weight and a page button), play- off race and odds; thousands of random leagues through the validator with 0 invalid, and the page checkboxes clicked in Chromium (eval/browser-click.mjs). Travel-table warnings (missing pair, triangle typo) landed 2026-10-30, on the CLI and the multi-division page (a notice above the schedule). Season grid CSV for multi-division leagues landed 2026-10-30 (CLI --grid and a page button, clicked in Chromium 2026-10-30: eval/browser-click.mjs). Team unavailable dates on the multi-division page landed 2026-10-30 (browser-clicked). Printable score sheets for multi-division leagues landed 2026-10-30 (page button). Standings from results on the page and CLI landed 2026-11-02 (browser-clicked). Cross-division play-off seeding with wildcards from those standings landed 2026-11-02 (page, CLI --playoffs, browser-clicked). The play-off race was checked against full goal-difference and head-to-head tables on 45,000 random finals (2026-11-02, 0 failures). Tie-break notes (why a team is above the next) landed 2026-11-02 (tiebreak.test.mjs) and show on the multi-division page from 2026-11-02 (browser-clicked). Weekly group-chat update for every division landed 2026-11-03 (divisionsWeeklyUpdate, page button; unit-tested on the made-up example league only; CLI --results r.txt --weekly landed 2026-11-03). Missing-scores chase list for multi-division leagues landed 2026-11-03 (divisionsOverdue, CLI --overdue, page button; unit-tested, browser-clicked 2026-11-03 with the weekly-update button: eval/browser-click.mjs). Next: a real league's travel times and a real league's unavailable-date requests.

   Source

   Progress, last 14 days: 46 merges · last change 2026-11-03

   • league-schedule-generator: browser-click the weekly update and missing-scores buttons 2026-11-03 · #12792
   • league-schedule-generator: missing-scores chase list for multi-division leagues 2026-11-03 · #12764
   • league-schedule-generator: CLI --weekly for multi-division weekly update 2026-11-02 · #12718

10. open-tennis-robots

    Parametric CAD (build123d) generating a collector's DXF/STL/STEP package, ballistics and tracking planners, ESP32 firmware and a browser simulator. Unbuilt hardware, long payoff.

    Working towards: The feeder CAD v0.1 package is in (manufacturing/feeder/: launcher head, trunnion cheeks, turntable, 2040 base with brackets, mast, 3 m net beam, poles and camera bar; the build tilts the head through -15..+45 deg for collisions and sizes wind ballast). The hopper and feed landed 2026-10-21: a tub…

    Full milestone

    The feeder CAD v0.1 package is in (manufacturing/feeder/: launcher head, trunnion cheeks, turntable, 2040 base with brackets, mast, 3 m net beam, poles and camera bar; the build tilts the head through -15..+45 deg for collisions and sizes wind ballast). The hopper and feed landed 2026-10-21: a tub on a sloped shelf and four posts, a hose, and a printed floor, roof, column and metering shuttle between the head plates (checks for one-ball metering, column and wheel clearance, tub capacity, hose slope and bend radius, shuttle clear when advanced; cad/README.md "Feed"); a printed front retainer (two flexure arms, strain, hold and pull-through forces checked) landed 2026-10-22. Next: close the feed's remaining gaps (post bracing, the hose end swinging about 450 mm), then the actuator brackets (shuttle and elevation) and azimuth drive, then the netting trough that rolls balls to the hopper inlet; then check the head's wheel-to-wheel gap and plate clearance against the ball path with the flight model's launch speeds.

    Source

    Progress, last 14 days: 4 merges · last change 2026-10-30

    • ai-bookkeeper: tenth gap probe, keyword rules for bakers, drone pilots, locksmiths, streamers, locums 2026-10-30 · #11942
    • polish: first-screen buttons on four flagship tool pages 2026-10-29 · #11478
    • open-tennis-robots: launch checks tie feeder CAD to the flight model 2026-10-29 · #11386

11. chant-forge

    Long-term music production programme: public-domain chant (GABC) to MIDI men's-choir lines with Latin lyrics, then genre templates, voices, mastering checks and a release kit for streaming distribution.

    Working towards: GABC to MIDI converter landed 2026-11-03 (8 tests), then a five-piece corpus from GregoBase (CC0 per its FAQ; test checks each mode final), then genre templates (genre.mjs: deep house, EDM, hard rock, metal; drums, bass, chords in the chant's mode, 13 tests). Mix/master check (master-cli.mjs: LUFS…

    Full milestone

    GABC to MIDI converter landed 2026-11-03 (8 tests), then a five-piece corpus from GregoBase (CC0 per its FAQ; test checks each mode final), then genre templates (genre.mjs: deep house, EDM, hard rock, metal; drums, bass, chords in the chant's mode, 13 tests). Mix/master check (master-cli.mjs: LUFS, true peak, mono, clipping, 6 tests) landed 2026-11-03; checked against ffmpeg ebur128 2026-11-03 (eval/ffmpeg-check.mjs: 0.05 LU, 0.33 dB on synthetic signals); EBU test files and a real mix remain. Release kit (release.mjs: ISRC/UPC, cover, WAV, rights, AI disclosure, 7 tests) landed 2026-11-03; check it against a distributor's real upload form. GABC reader fixes for CRLF, <sp>/<v>/<alt> and in-group clefs landed 2026-11-03 (made-up strings); NABC and split-neume pitch test (nabc.test.mjs) landed 2026-11-03. --fit 8 --stretch (2026-11-03) lengthens long notes to fill bars, halving padded rest on the corpus; ear check pending. --halftime (2026-11-03) doubles chant durations over an unchanged backing for metal; ear check pending. Real-file batch landed 2026-11-03: 59 Graduale 1908 files, no crash or warning (batch.test.mjs, eval/batch-check.mjs). --cues FILE singer cue sheet (cues.mjs, 3 tests) landed 2026-11-03; try it with real singers. Next: compare a sample of the files against a pre-1929 scan (notes, not just parse), source the Te Deum and Salve Regina from a pre-1929 book, read the 11 off-final endings against the book, and listen to the --stretch output (ROADMAP.md).

    Source

    Progress, last 14 days: 14 merges · last change 2026-11-03

    • chant-forge: singer cue sheet (--cues) 2026-11-03 · #12800
    • chant-forge: --halftime puts the chant at half speed over a full-time backing 2026-11-03 · #12791
    • chant-forge: run the converter over 59 real Graduale 1908 files 2026-11-03 · #12789

1. sim2real-inference

   How few real trials certify a robot policy when a biased simulator supplies the rest, with guarantees that hold whatever the bias?

   Aimed at: CoRL / RSS (robotics), or a statistics venue for the theory

   Progress, last 14 days: 37 merges · last change 2026-11-01

   • research: pick_track skips programmes marked BLOCKED 2026-11-01 · #12434
   • research/sim2real-inference: finite-pool certificate result moves into the paper's main text (7y), plus post 2026-10-29 · #11591
   • research/sim2real-inference: per-episode pairs blocker is missing data, not the host 2026-10-28 · #11221

2. verifiable-training-mechanisms

   Can outsourced ML training be made incentive-compatible to verify, with scoring rules, audits and peer prediction combined into one protocol that comes with proofs?

   Aimed at: ACM EC / AAMAS

   Progress, last 14 days: 58 merges · last change 2026-11-02

   • research/verifiable-training-mechanisms: milestone 63, paper gains magnitude and aggregate-check results 2026-11-02 · #12712
   • verifiable-training-mechanisms: paper gains progress-check results (milestone 62) 2026-11-02 · #12663
   • research/verifiable-training-mechanisms: milestone 61, finer stall fractions 2026-11-02 · #12605

3. lowcomm-training

   A predictive theory of low-communication training (DiLoCo-style local SGD): which of sync interval, compression, delay, momentum and robust aggregation set the final loss, checked on real small-transformer training

   Aimed at: ICML / NeurIPS workshop, then main track

   Progress, last 14 days: 55 merges · last change 2026-10-31

   • research/lowcomm-training: 6n compression excess not linear in omega 2026-10-31 · #12114
   • research/lowcomm-training: 7c, compression paper section from 6a-6m 2026-10-30 · #11918
   • research/lowcomm-training: milestone 6m, four forms in H scored leave-one-cell-out 2026-10-30 · #11842

4. verified-dse

   Multi-fidelity, verification-aware design-space search for open-source chip flows: cheap rungs plus predicted formal equivalence, scored on real SKY130 place-and-route

   Aimed at: MLCAD / ICCAD

   Progress, last 14 days: 74 merges · last change 2026-11-03

   • research/verified-dse: milestone 70, R4-probe gate (negative result) 2026-11-03 · #12798
   • research/verified-dse: milestone 69, mult36 and mult40 at the relaxed clock 2026-11-03 · #12781
   • research/verified-dse: milestone 68, multiplier widths 20 and 28 where R2 mis-ranks 2026-11-02 · #12740

5. llm-coding-inference

   Valid downstream inference from LLM-coded qualitative data under group shift and rare categories, on real annotated corpora

   Aimed at: a computational social science methods venue (e.g. Sociological Methods & Research)

   Progress, last 14 days: 72 merges · last change 2026-11-02

   • llm-coding-inference: milestone 7l, cells vs pooled at fixed and fitted lambda 2026-11-02 · #12713
   • research/llm-coding-inference: 7k, coder-cell gain is flat in pilot share 2026-11-02 · #12573
   • research/llm-coding-inference: 7i's cell ceiling as library code 2026-11-02 · #12547

6. aerolog-scaling

   What does a 1:N Froude-scaled model of a cargo UAV reproduce of the 1:1 aircraft, and where does it lie (Reynolds-driven drag and stall, powertrain, energy)?

   Aimed at: AIAA SciTech / Journal of Aircraft

   Progress, last 14 days: 52 merges · last change 2026-11-02

   • research/aerolog-scaling: 9x drift ramp in the output-error fit 2026-11-02 · #12669
   • research/aerolog-scaling: 9w constant IMU offset in the output-error fit 2026-11-02 · #12651
   • research/aerolog-scaling: 9v IMU zero offset and temperature drift 2026-11-02 · #12627

7. marine-hull

   Which body (monohull, catamaran, SWATH, hydrofoil, glider, AUV) carries a survey payload furthest per kWh, and where does the ranking flip under uncertain drag inputs?

   Aimed at: OCEANS (IEEE/MTS) / Ocean Engineering

   Progress, last 14 days: 53 merges · last change 2026-11-02

   • research/marine-hull: milestone 7l, paper draft gains the Michell wave cross-check 2026-11-02 · #12753
   • research/marine-hull: milestone 3k, Michell wave term with tank-bound form factors 2026-11-02 · #12678
   • research/marine-hull: 7k fold milestones 3i and 3j into the paper draft 2026-11-02 · #12608

8. heavy-hydraulics

   Which buildable hydraulic circuit (open-centre, load-sensing, independent metering) gets a small excavator closest to stock energy use for the least parts cost?

   Aimed at: Int. J. Fluid Power / Bath-ASME fluid power symposium

   Progress, last 14 days: 55 merges · last change 2026-11-02

   • research/heavy-hydraulics: milestone 5m, CWA excavator profile is computed, not measured 2026-11-02 · #12714
   • research/heavy-hydraulics: milestone 6g, paper section 3.4a gains 5k and 5l 2026-11-02 · #12707
   • research/heavy-hydraulics: milestone 5k, open electric-excavator power profile shape check 2026-11-02 · #12685

9. build-data

   Do public building-permit records, loaded from several cities onto one open schema, predict how long a permit and the build behind it take better than each city's own averages, and which fields carry the signal? (the first table of the construction data warehouse; owner, 2026-10-31)

   Aimed at: Automation in Construction / ASCE J. Construction Engineering and Management

   Progress, last 14 days: 16 merges · last change 2026-11-03

   • research/build-data: milestone 15, XGBoost and linear quantile baselines on SF and Austin 2026-11-03 · #12783
   • research/build-data: milestone 14, XGBoost and linear quantile baselines for H1 2026-11-02 · #12755
   • research/build-data: paper draft 0 from milestones 1 to 8 2026-11-01 · #12325