A free, browser-based rig-tune and sail-trim trainer for one-design sportboats — the J/70 first, the Melges 24 second. No backend, no account, works offline on a phone after the first visit.
▶ Try it: rjdscott.github.io/sailflow
The Simulator at 10 kt, close-hauled. The right half of the instrument band is the world — wind speed, the rose you drag for angle, sea state, crew, sail set — and every value in it is the control that sets it; the left half is what the boat does about it. The Rig panel carries the forecast the tune is bet on and the three shroud turns. Every number wears its confidence tier (A/B/C).
Sailflow answers two questions a one-design crew asks every race day:
- The rig — what shroud, rake and forestay setting do I commit to for today's forecast, and what does it cost me at each end of the range?
- The sails — with the rig I have, where do the sheets, traveller, backstay, vang, outhaul, cunningham, leads and kite controls go for the wind in front of me, and what does each move do to the sail?
Both are answered on one page, the Simulator at
#/sim — rig and sails on the
same screen, so a shroud turn and its effect on the jib are visible together
(ADR 0021).
Links written against the old #/race and #/dock URLs still open.
It solves a steady-state velocity-prediction model (VPP) in a Web Worker, draws the flying shape of every sail in 3D and plan view, compares its answer against the North and Quantum tuning guides, logs what you tried, and drills you on the decisions with fault-injection exercises. Trim states are shareable as links; a pinned trim shows as a ghost outline with deltas.
390 px wide, the top 1180 px of the page: the conditions are the first thing you touch, the boat's numbers are under them, then the picture. The whole instrument band is inside an 844 px phone screen.
Most trim advice is a table in a PDF. Sailflow is a model you can argue with:
- Every number carries a confidence tier. A = a number you can use; B = a direction and a band; C = a direction only. Tiers are computed, not asserted — an output is demoted the moment its inputs leave the calibrated envelope.
- Provenance is enforced. Every literal in the physics core has a
prov:tag or a row inASSUMPTIONS.md; a script fails CI if one is missing. Third-party data is committed one file per source with its citation inPROVENANCE.md. - Disagreement is shown, never resolved silently. When the model and a tuning guide differ, you see both and the delta.
- Validation is public.
validation/report.mdreplays the ORC J/70 polar through the solver with two wind speeds held out of the fit. CI regenerates it on every push and puts the held-out tables and the gate verdict in the job summary, so a regression is visible in the PR. The current verdict, and the rows it fails, are below. - Deterministic core. Same inputs, same outputs. No randomness, no clock,
no framework in
src/core. A golden corpus of solved cases per boat catches any drift.
Complete and live: one Simulator page with the 3D hero (three.js,
lazy-loaded, first-frame budget) — the conditions are the editable right half
of the instrument band, the Rig panel carries the forecast, the expected regret
of committing once, the three shroud turns and the commit lock, and the
mainsail, headsail/gennaker and helm panels sit around the boat (ADR 0021).
Plus: disagreement panel driven by whatever guides are in data/tuning/,
tuning log with export/import, drills v2, PWA offline, dark/light themes,
desktop-first layout down to phone. Phase two added shareable trim links and
pin-and-compare (ADR 0019), a first-run tour and a schematic explainer for
every control, a downwind model that soaks (ADR 0018), a boat picker with the
Melges 24 as a second class (ADR 0020), and a phone that loads the 3D chunk
only when asked.
Where it stands, stated plainly. The polar hold-out gate (ADR 0012: every row at TWS 8 and 14 kt withheld from calibration) reads PASS — 10 of 10 rows on the J/70: boat speed within 3 % (VMG rows) and 5 % (60/90/120° rows), and every VMG row within 1 % of its own best VMG when sailed at the polar's printed angle. Two things closed it, and neither is a loosened tolerance:
- Heel now costs published drag (ADR 0022). The polar's upwind speed plateaus at 5.89–5.95 kt from 14 to 20 kt and the model used to plateau at 6.23–6.34, ~6 % fast on the held-out row and on the fitted 16 and 20 kt rows alike. The heel term was anchored on viscous resistance and could not reach the penalty the plateau needs; on the published Delft heel law it can. Held-out TWS 14 upwind is now 0.9 % fast.
- The VMG-angle criterion measures knots, not degrees (ADR 0023). The last failing row was 3.3° from a 2° tolerance on a VMG curve flat to 0.11 % over 168–172° — sailed at the polar's own 172° the model gives up 0.06 kt of 6.27. Each VMG row is now solved a second time at the polar's printed angle and must keep 99 % of its own best VMG.
What is still wrong is on the record rather than in the gate: the model's
downwind optimum is compressed into 168–169° from 14 kt up while the polar's
spans 141.9–174.0°, and closing that needs a second mechanism, not a better
number. Written up in
ASSUMPTIONS.md.
The Melges 24 gate is 8 of 10, on two boat-speed rows
(validation/report-m24.md).
Downwind boat speed stays tier B and downwind optimum angle tier C: the two gated downwind rows now pass on boat speed (0.1 % and 1.9 %), but they pass at an angle the model still picks 3° hot, which is not enough to promote a tier. The app says so on screen. Phase 01 of the phase-two plan carries the detail.
The aero layer is the documented ORC VPP parametric model (2023 edition,
PROVENANCE.md) with an explicitly invented layer on top
that maps rig bend, headstay sag and sheet positions to sail draft, draft
position and twist, and those to ORC's flat, reef and effective twist.
There is no public evidence base for that layer on a J/70; it is calibrated
against the ORC polar and labelled tier B or C throughout. Hydro is ORC's
canoe-body and appendage resistance with added resistance in waves.
ADR 0006
records the split; ASSUMPTIONS.md "Where the model is honestly weak" lists
what a sailor should not trust.
Parametric VPPs under-predict boat speed relative to CFD tools. This is a decision-rehearsal tool for tuning and trim choices, not a wind tunnel.
make setup # pnpm install --frozen-lockfile
make dev # vite dev server on http://localhost:5173
make check # docs-check + lint + typecheck + 1100+ unit tests
make validate # polar hold-out gate, regenerates validation/report.md
make help # every target
pnpm exec playwright install chromium # once, before the first test:ui run
pnpm test:ui # Playwright: layout, a11y, 3D smoke, screenshotsNode 20, pnpm 9, and uv — make docs-check
runs the documentation tests through uvx pytest, so make check needs it:
curl -LsSf https://astral.sh/uv/install.sh | shPython 3.12 is used only by that docs tooling. Playwright screenshot baselines
are generated inside mcr.microsoft.com/playwright:v1.62.1-noble so they match
CI — see docs/runbooks/.
src/core pure physics — geometry, rig, shape, aero (ORC), hydro, solve
no DOM, no framework, no Math.random, no Date
src/worker the typed request/response protocol; the UI's only door to core
src/ui Svelte 5 — simulator cockpit, log, drills, three.js hero
data/ boats, tuning guides, ORC polar: one JSON per source
calibration/ fits the invented-layer knobs to the polar; writes the golden corpus
validation/ replays the polar through the solver; hold-out gate; report
The core/UI boundary is enforced by ESLint (no-restricted-imports): the UI
may import only types from src/core/types
(ADR 0003).
Solves run in a Web Worker with a fixed Newton budget and a fixed seed table,
so a solve is reproducible and never blocks the frame.
Stack: Svelte 5 runes, Vite, TypeScript strict, Vitest, Playwright, three.js (one lazy chunk, gated by a measured first-frame budget), vite-plugin-pwa, GitHub Pages. Two boat classes, 21 ADRs.
The repo is run on a documentation pipeline that is machine-checked
(make docs-check): research → decision → plan → audit, plus runbooks.
| Where | What |
|---|---|
docs/adr/ |
21 decisions, Nygard format, immutable once accepted |
docs/plans/ |
Phase plans with status tables; resumable by a stranger |
docs/audits/ |
Point-in-time sweeps (UX ×4, docs consistency, first impressions) with evidence and punchlists |
docs/research/ |
Cockpit UX, spinnaker aerodynamics and trim, simulator landscape |
docs/runbooks/ |
Nine operational how-tos: deploy, release and cache-bust, recalibrate, add a boat, add a drill, add a tuning guide, export the log, reshoot the screenshots, start a new project from the template |
docs/initial-prompt.md |
The original build brief and acceptance criteria |
CHANGELOG.md |
Per-release, from squash-merge titles |
Start with docs/plans/README.md for what is done and
what is next, and CLAUDE.md for the engineering
rules.
One thing that corpus will not explain about itself: Sailflow was built autonomously by Claude Code agents against a written brief, with the owner reviewing and merging. That is why the plan progress logs and audit method lines name models — Fable, Opus, Sonnet — as the party that did a piece of work. Those logs are dated, append-only records of what happened; they are kept as written rather than tidied up, for the same reason the polar gate is published while it still reads FAIL.
Phase two and the Simulator merge are closed
(docs/plans/2026-08-26-phase-two/,
docs/plans/2026-08-28-simulator/).
What they leave open, each with the reason written where it lives: the
downwind optimum angle (3° hot; needs a second mechanism, ADR 0018), a
helm/rudder-angle readout (needs a yaw balance in the core), a third J/70
tuning guide and any Melges 24 guide (none publicly fetchable), and the
Melges 24's own gate (7 of 10 rows; six knobs unfitted, ±11 % polar spread).
Out of scope, by the brief: multiplayer, race simulation, tactics, starts, accounts, any backend.
Issues and PRs welcome, especially from J/70 and Melges 24 sailors with
measured numbers.
The rules are short: branch + PR + squash, make check green, every literal
carries provenance, every model output carries a tier, tests land with the
change. Details in CLAUDE.md — the sections above
that heading are generic template conventions, not Sailflow's. To add a boat
class, follow
docs/runbooks/add-a-boat-class.md.
MIT. See LICENSE. Tuning-guide numbers are transcribed from
publicly available guides with attribution
(ADR 0008);
if you own one of those guides and object, open an issue.