A browser-based, physically-laid-out 3D simulation of a fusee movement with a Swiss lever escapement: mainspring drum → chain → fusee & great wheel → center → third → fourth wheel → escape wheel ⇄ pallet fork ⇄ balance wheel + hairspring, plus an alarm complication with its own barrel and gong, and a dial with correctly driven hour/minute hands, a small-seconds sub-dial, a power-reserve sub-dial and a 12 h alarm-setting ring.

Tornado layout — the movement is composed as a face design on a flat construction: crown and barrel exit at ~1:50, the fourth wheel sits exactly at 6 o'clock so its arbor carries the small-seconds display directly (no fake linkage), the escapement trails to ~6:25 with the balance at 8, and the power-reserve sub-dial answers at 12. The going train no longer runs a uniform stride: the z-stack is solved bottom-up (src/layout.js), the escape wheel drops below the train while its pinion stays up in the fourth wheel's plane, and the fusee cone is squashed to ~3 units — the height four turns of true-scale chain actually need — which is what lets the three-quarter plate floor come down. The whole assembly, hands to alarm barrel, measures ~26.4 units front-to-back; at the scale pinned in §39 (0.379 mm/unit, derived from real fusee-chain pitch) that is ≈10 mm deep on a 32.5 mm plate.

Use the dev server, which also serves the /__state endpoint the save/load buttons persist to and sends Cache-Control: no-store so edited modules aren't served stale:

python3 dev_server.py        # :8347, loopback only
python3 dev_server.py --host 0.0.0.0   # also reachable from the LAN (trusted networks only)

then open http://localhost:8347/ — no build step, no network access needed (Three.js 0.165 is vendored in vendor/). A plain static server (python3 -m http.server 8347) also works; state then falls back to localStorage.

Three of them, published to GitHub Pages by .github/workflows/pages.yml (§88). GitHub Pages gives a repository one site, so they are three paths under it rather than three sites — which works only because release URLs are relative (see §28 in docs/BUILT.md), and so the app never needs to know how deep under the origin it is being served.

So the ladder is merge → cut a release → promote, and every pointer is a git ref: main, the newest major.minor.patch tag, and a production branch the promote step moves. Nothing about which environment serves what lives outside the repository.

The Pages workflow itself only ever runs from main — each environment's tree comes from its own ref, but the tooling that stamps them comes from the checkout, so pinning that to one place is what makes a rebuild reproducible.

Each carries the version it was built with, readable without a fetch:

curl -s https://kelaiem.github.io/timesim/development/version.json
# {"version":"2.1.9-28-g4b64e7d","environment":"development"}

Development's version is git describe — literally "28 commits past 2.1.9", which is what the tip of main is. Testing and production name their release tag. Only production is indexable; the other two are served noindex.

Every deploy — Pages and the SFTP release alike — ships the payload defined by tools/payload.sh: the app, vendor/, and the licences, and nothing else. Repo documentation (every *.md, including docs/), dev_server.py and the git hooks describe the project to people working on it and are not site content, so they stay in the repository. Both workflows assert both halves — no doc leaks in, no licence goes missing.

Two words carry precise, different claims here, and the honesty ledger (TODO.md) trades on the difference:

• The model is the description: parametric geometry generated from its own dimensions and constraints, constants derived from stated relationships, and declared structure (MECH_GRAPH — what supports what, what drives what). The model exists at boot, motionless. Saying a thing is modelled claims only that it is described — a spring can be modelled as matter yet have no force law.
• The simulation is the model advanced through time by causality: step(dt) carrying force from its sources (mainspring, crown) through drive edges to every display. Saying a thing is simulated is the stronger claim — its behavior is driven, not posed. The negative space has names of its own: a part that animates with no force path is a simulation fiction (the graph check's term), and a run whose members are posed from the output rather than driven from the input (TODO 20) is modelled truthfully but not yet simulated.

"The sim" is shorthand for the running artifact — model, simulation, rendering and panel together. The schematic view's motto ("draw the model, not the metal") uses model in a narrower, idealized sense: the pitch circles, ratios and laws the solids embody.

• Balance: 2.5 Hz torsional oscillator (18,000 bph, 5 beats/s).
• Escapement: per-beat lock → unlock (with draw recoil) → impulse → drop state machine; the escape wheel advances exactly 12° (half a tooth pitch of the 15-tooth club-tooth wheel) per beat with an eased impulse ramp.
• Train: all wheel angles are closed-form functions of the escape-wheel angle — barrel 80T→10 center pinion, center 75T→10, third 80T→10, fourth 80T→8 escape pinion; fourth wheel = 1 rev/min, center = 1 rev/h, barrel = 1 rev per 120/7 h (§124). No drift.
• Mainspring: visible coil in the barrel cutaway relaxes over ~30 simulated hours. Winding goes through the keyless chain, whether you drag the crown or press Wind (which turns it for you); a one-way click means only forward turns bank reserve, and at full wind the cone stops however hard you crank.
• Maintaining power: drive runs cone → base ratchet → pawls on the maintaining wheel → maintaining spring → great wheel, so the train keeps going while you wind instead of stopping dead — the standard fusee answer to losing power at the one moment you are adding it.
• Power reserve (functional): the movement runs on its own "movement time" that stalls when the spring is spent — balance amplitude sags as tension drops (like a real watch), and at zero the balance stops dead-centre with the train locked and hands frozen until the next wind. The indicator is driven off the state of wind through a visible reduction train (below) and reads on a sub-dial at 12 o'clock — a 150° arc graduated 0 → 30 h, figured 0/12/24 in Arabic — plus an hours readout in the panel.
• Small seconds: the seconds hand rides a sub-dial at 6 o'clock centred exactly on the fourth wheel's axis — its display arbor (the heart cam's slip-coupled arbor, extended as a real rod through the wheel's bore and the plate) is coaxial with the actual 1-rev/min fourth wheel.
• Fusee & chain: the movement is a fusee movement — the mainspring lives in a plain drum off to the side, and a chain runs to a helically-grooved cone on the great-wheel arbor. Fully wound, the chain pulls at the cone's small radius; as the spring weakens it pays off toward the large radius, so torque delivered to the train stays level (see the Spring/Train torque bars in the panel — the cone profile satisfies S(t)·r_f(t) = const). The chain visibly migrates between drum and cone as the reserve changes.
• Fast-forward: a ~5400× mode that rips through the whole 30 h reserve in seconds so you can watch the chain pay off, the reserve hand fall, and the movement run flat; auto-disengages at zero. Winding restarts the balance.
• Alarm: a second, independent complication with its own crown, barrel (1.75 turns of wind), striking wheel, hammer and coiled gong. The setting disc is read against a 12 h ring on the dial at quarter-hour marks; a release feeler drops into the disc's notch at the set time, unlocking the striking train for ≈28 strikes (~12 s) on one wind. The arming run is forward-driven end to end: a press advances the column wheel (its parity IS the on/off), the beak rides castellations cut from the same function the tick reads, the rod stands between its two contacts, and a crank pin rides a forked tab's groove to drive the selector ring positively both ways — the alarmHandoffs check measures all six hand-offs closed, zero waivers. (The one filed gap: the pawl's index stroke is a transient the pose-based instrument cannot reach; TODO 20's closing status carries it.)
• Keyless works: knurled crown → stem → winding pinion → crown wheel → ratchet + click; the winding train turns with true tooth ratios when you wind, and the fusee chain migrates as the cone takes it up.
• Setting-lever linkage (visible hacking actuation): the stem carries a grooved collar pair; the setting lever's pin rides in it, so pulling the crown rotates the lever. Its tall tail post presses the hack spring — a long blued blade reaching across the movement whose ruby pad lands exactly on the balance rim — and drives the reset-hammer rod; a separate yoke tracks the sliding pinion's hub between the winding and setting meshes.
• Power-reserve gear train: a visible two-mesh reduction across three arbors (8T/28T × 10T/12T = 1/4.2) whose first pinion sits slip-coupled on the barrel arbor and whose last wheel shares the indicator hand's arbor. The ratio is derived, not chosen: the arbor makes 1.75 turns over the 30 h reserve (§124's first-stage re-gear) and the hand sweeps 150°, so R = 630/150 = 4.2 — and a build-time assert holds the graduation, the hand's travel and the tooth counts to each other, because for a while they disagreed (TODO 18).

Pause/play, time-scale (a log slider, 0.02×–1×; starts at 1× — around 0.15× is where the unlock-impulse-drop sequence becomes followable by eye), Wind, fast-forward, sync-to-wall-clock, sound, camera presets (plus orbit/zoom), exploded view (whole-movement slider or one assembly at a time), part labels, x-ray, power-flow highlighting, a measurement overlay with an in-scene mm ruler, a guided tour and an inspection route, a control HUD (which carries the clock and alarm-time readouts with it, so the alarm can be set without the panel), a reconfigure mode (§33: drag either crown, the pusher (which carries the whole alarm module with it), the barrel, escapement or balance to propose a new arrangement — dragging the crown re-solves the keyless cluster around a decoupled stem, idler and all; the pure solvers shadow-solve each candidate live, warnings and all, specs apply at reload and are refused with reasons where they cannot work, a Trial boot button reads a candidate's full boot-assert verdict from a hidden throwaway boot without committing the view, and named variants with history-based undo make the spec a document), save/load of the whole scene state, and the beat counter, simulated clock, reserve and alarm readouts. The alarm has its own crown and pusher. ?inspect=1 and ?cycle=1 deep-link into the inspection and alarm-cycler routes; an Advanced panel exposes the finish parameters from src/aesthetics.json.

The chrome is localised — English, German, French, Japanese, Chinese in both scripts, and Arabic — selectable in the panel or by ?lang=fr / ?lang=zh-Hant (a script subtag, so zh-TW, zh-HK and zh-MO all resolve to Traditional) / ?lang=ar (the one right-to-left locale: the document's direction follows it, the drawings and the panels' corners do not). primer.html and explain.html, the mechanism explainer linked from the HUD, are translated in all seven. A released build also loads with the network gone: a service worker precaches the release, so a page that has been visited online once boots offline, deep links included. Neither applies to a source tree, which registers no worker at all so that an edit is never shadowed by a cache.

• src/layout.js — the layout contract: tooth counts, the z-stack, the one clearance margin, the unit→mm pin, and solveLayout (the planar solve as a pure function of its spec).
• src/geometry.js — parametric part builders (gears with crescent crossings, club-tooth escape wheel, pallet fork with ruby stones, balance with timing screws, Archimedean hairspring, cutaway barrel, plates, dial, hands).
• src/materials.js — shared PBR materials (brass, steel, blued steel, ruby…).
• src/main.js — scene, studio lighting + procedural PMREM environment, movement assembly (mesh distances from pitch radii), escapement kinematics, UI.
• src/inspect.js — the realism inspector (below).
• src/state.js — state persistence via the dev server's /__state, falling back to localStorage.
• src/aesthetics.js, src/aesthetics.json — finish parameters; see AESTHETICS.md for the reasoning.
• src/i18n.js — the UI localisation, keyed by the English source string. src/page-i18n.js is the shared engine for the two static pages, and src/explain-i18n*.js / src/primer-i18n*.js are their per-locale tables.
• explain.html — the mechanism explainer: a plate per mechanism, quoting the real source constants, linked from the HUD and styled as it.
• sw.js, manifest.webmanifest, favicon.svg — the offline worker (inert in a source tree; baked at release time), the web-app manifest, and the icon, which tools/make-favicon.mjs generates from the same monogram the winding crown carries.
• dev_server.py — the static server plus /__state and no-store.
• test-geometry.html — standalone visual smoke-test page for every part builder.
• SPEC.md (architecture contract), docs/BUILT.md (how each shipped feature was designed, numbered §n and cited from source comments), docs/MODELING.md (geometry conventions), TODO.md (mechanical-realism debt), CLAUDE.md (working rules).

window.__clock.step(dt) in the console single-steps the simulation deterministically (useful because background tabs throttle requestAnimationFrame).

src/inspect.js sweeps the mechanism deterministically through fourteen phase axes — beat, crown, reserve, wind, arrest, stemSlip, train, jumperEngage, handSet, alarm, alarmStrike, alarmWind, alarmToggle, alarmPress — via __clock.setPose() and reports every pair of functional units whose meshes intersect (exact triangle tests via the vendored three-mesh-bvh). Pairs with intended mechanical contact (gear meshes, pallet lock, chain-on-cone…) are classified EXPECTED and reported separately; everything else that touches is FORBIDDEN — a defect.

Run it from the console with start()/status(), never by awaiting a sweep directly: full runs take 100 s+ and will blow a browser-eval timeout.

document.getElementById('btn-pause').click();
const I = await import('./src/inspect.js');
I.start(__clock, 'inspection', { includeExcluded: true });  // then poll I.status()
I.start(__clock, 'support');                                // 0 failures
I.start(__clock, 'clearances');                             // 0 violations
window.__inspect.show('<pair>', '<axis>');                  // jump camera to a hit pose

Beyond overlap the module carries checkMechanicalGraph (is every part grounded and driven), checkPenetrationBudgets (how deep an intended contact goes — being on EXPECTED_PAIRS proves contact was intended, not that its depth is reasonable; budgeted pairs run from the ruby stones to the alarm linkage), checkAlarmHandoffs (does each claimed contact of the §35 arming run actually close — signed gap-or-burial at both toggle parities, with out-of-band rows carried as waived debt citing their TODO items), checkSweptOverlap against a swept-volume registry, checkStockFloor and slenderness (a part can be thick enough and still be a noodle — measured per FREE length, so a shaft that declares its bearings is judged on the spans between them and on its overhangs, not on its stock), checkLowCorridor, auditOscillators, stockCensus and fingerprint for determinism, plus the three that answer the blind spots below — intraUnit (all three pair classes inside a unit — movers vs fixtures, fixture pairs, movers across rigid frames), expectedContacts (per-contact clearance floors across declared pairs) and restoring (§48's no-spring audit: anything that reciprocates has a restoring element, is driven both ways, or is waived citing its TODO). focusedCheck(clock, names) runs the same budgets scoped to the parts you just moved, in seconds rather than minutes.

CI runs the whole bar on every PR: .github/workflows/battery.yml drives tools/ci-battery.mjs under headless Chromium, plus a boot-silence check and a double-boot fingerprint comparison. node tools/ci-battery.mjs runs the same gate locally (needs npm ci in tools/ and a Playwright Chromium).

Two things the pair sweep structurally cannot see, both written up in TODO.md (items 5 and 6): a part colliding with another part of the same unit, and any second overlap between a pair that already has one declared contact. checkAlarmHandoffs closed both for the one run that was hiding in them — its rod⇄tail and rod⇄crank rows are intra-unit contacts the pair sweep can never enumerate — and each now has a general instrument as well (intraUnit, expectedContacts), each with known residue: intraUnit's fixture-pair and cross-frame mover tiers gate only INTRA_TIER_SCOPE (the alarm complex) and REPORT everywhere else, same-frame mover splits gate only ASSEMBLY_SCOPE, and an EXPECTED pair with no declared floors row still gets the blanket excuse. If you are checking one of those cases, read the reported rows — or measure it yourself.

The finishing and layout follow the Glashütte school of watchmaking: a three-quarter plate, a separate balance cock screwed to it, Glashütte striping across that plate and the escape bridge, perlage on the base plate, screwed gold chatons over every upper pivot jewel this plate carries — five of them, on the centre, third, fourth and escape wheels and on the alarm's striking arbor, each dropped into its own counterbore and held by blued screws whose heads lap its rim (two at the escape wheel, where the plate's escapement opening leaves room for no more) — and blued steel screws and hands. The flush rubbed-in stone, which is the older bearing and also a Glashütte one, survives on the cocks and on every lower pivot.

The dial no longer speaks German. It used to: the power reserve was graduated AUF/AB — "up/down", the conventional marking for a German reserve indicator. That scale is now figured 0 → 24 in Arabic numerals, and both sub-dials are captioned in English (POWER RESERVE, SECONDS), so the words that named the school on the front of the watch are gone. Removing them is not a claim to have left the school. The movement behind the dial is where the vocabulary actually lives, and every item in the list above is still there — the three-quarter plate and its cock are the layout, not a decoration applied to it.

These are traditional techniques and conventions, in general use across German horology for well over a century. They belong to the craft, not to any one maker. Nothing in the list is exclusive to one house, and none of it is borrowed from a specific caliber. That is a statement about the FINISHING, and it stays true: three specific watches are acknowledged under Sources below, but for ideas in the complications — an alarm's architecture, a dial's smoke, a hand that hides — never for the finish, and never as geometry.

"Swiss lever" names the escapement's design, not its passport. The term means the straight-line layout (escape-wheel centre, pallet pivot and balance centre collinear — this build measures exactly 180.0° at the fork, from the layout solve's own positions) with a club-tooth escape wheel splitting the impulse between tooth face and pallet stone — as opposed to the English lever's right-angle layout and pointed teeth. A German-styled movement carrying a Swiss lever is the industry's norm, Glashütte's included; the historical Glashütte lever escapement (Glashütter Ankerhemmung) is a distinct variant of the lever family and is not what is built here.

The movement itself is an original parametric model: every part is generated from its own dimensions and constraints in src/geometry.js, not traced or measured from any manufacturer's caliber. No maker's name, logo, or trademark appears in this project or on the dial.

This project is not affiliated with, endorsed by, or a product of any watch manufacturer. If the result puts you in mind of a particular Glashütte house, that is the shared vocabulary of the school, borrowed admiringly.

An acknowledged influence. Bartosz Ciechanowski's Mechanical Watch (https://ciechanow.ski/mechanical-watch/) is an inspiration for this project's explanatory side — the conviction that a mechanism is best explained by letting someone move it, and that the explanation deserves the same care as the thing explained. explain.html, primer.html and the schematic view all exist because of that conviction. Nothing here is copied from it: no geometry, no code, no text, and the movement is a different one (a fusee-and-chain caliber with an alarm). The debt is one of ambition, and it is gladly acknowledged.

The influence above is one of ambition. This section is the other kind of debt, and it has two halves. First the descriptions of how these mechanisms actually work, which the movement was built from. Then, at the end, the handful of specific WATCHES this movement takes an idea from — a different sort of citation, and the one that has to be most careful about what it is claiming. Nothing here was traced, measured or copied: these are the accounts that were read, the watches that were admired, and the ones to check this movement against.

Fusee & chain. SJX Watches is the reference for the mechanism and for what it is for — that a mainspring's torque is usable only across its flat middle, and that the cone trades radius against that decay so the train sees a level input. Which is exactly the claim equalisation gates and Fusee & chain plate 2 draws.

• SJX, Insight: The Chain and Fusée
• SJX, Explaining the Ups and Downs of Constant Force Mechanisms
• SJX, Explaining the Microscopic Chain Inside the A. Lange & Söhne Pour le Mérite — the reason the chain is built here as riveted inner/outer plate pairs rather than a cord, and the reason CHAIN_PITCH_MM = 0.72 is pinned to a manufactured standard instead of chosen (it is what sets UNIT_MM, so the whole movement is dimensioned off a real chain).
• A. Lange & Söhne, The fusée-and-chain transmission

Maintaining power. A fusee loses its drive while being wound — the arbor turns the wrong way — so the great wheel runs loose and is fed by a spring held against a detent for the duration. That is John Harrison's arrangement (1730s), standard on fusee marine chronometers; Lange solves the same problem with planetary gearing under the ratchet wheel instead. It is why this movement has a maintaining detent at all, and why the escapement cannot run backwards during a wind.

The lever escapement. George Daniels, Watchmaking (rev. ed., Philip Wilson) is the standard treatment of the phases Swiss lever escapement names — lock, draw, impulse and drop — and of why the locking face's lean is what holds the fork on its banking. The build asserts the sign of that draw torque at boot; Daniels is where the requirement comes from. Donald de Carle, Practical Watch Repairing, is the everyday reference for the keyless works and the motion works.

Tooth form. Watch and clock wheels are cut cycloidal, not involute — Ciechanowski's gear-train section is the accessible statement of why, and Daniels the practical one. This movement does not cut either: see TODO.md item 61, which says so at length rather than letting the citation imply otherwise.

The gong. The partial ratios in The gong's voice are the Euler–Bernoulli clamped-free beam's, not a horological source: the mode constants (βₙL)² = 3.516, 22.03 are standard, and the ~6.3× second partial they give is why a struck bar clangs where a string sings.

Three, and each is an IDEA rather than a design: an arrangement, or a trick of display, that this movement adopts and then builds from its own dimensions. No geometry, no dimension and no part of any of these calibers is reproduced here, and where this movement does the thing differently — or does not do it at all — that is said in the entry rather than left for the reader to discover. The finishing note above is unaffected: the Glashütte vocabulary it lists stays school-level and unattributed, and these three are complications, not finish.

The alarm — Jaeger-LeCoultre Memovox. 1950, cal. 489, manually wound. The idea taken is the ARCHITECTURE, which is the one every later alarm wristwatch argues with: a second mainspring barrel for the alarm standing beside the going barrel, its own crown to wind and set it, and a hammer let go against a sounding part when the set time arrives. That is what this movement is — the alarm barrel, the alarm crown, and the striking works equalisation holds to their own set-up and gap. Three differences worth stating. The Memovox's hammer strikes a post on the caseback; this one strikes a wire GONG carried round the movement, whose voice is derived from the clamped-free beam rather than from any watch. A Memovox reads its alarm time off an inner rotating disc with a small triangle — where this movement uses a central hand, which is the next entry. And the alarm's ON/OFF switch is not the reference's at all: this one steps a chronograph's COLUMN WHEEL half a pitch per press, and the lock lever's tail beak rides its castellations — beak on a column and the brake cannot lift, beak over a gap and the lever answers to the release. That is a mechanism borrowed from a different complication to do a job the reference does another way, and it is the largest departure on this list: the gong and the hand are substitutions for parts a Memovox has, where the column wheel is the movement's own. It earns the space — the wheel reads to FOUR riders (the brake beak, the click, the §35 link beak and the pusher's pawl), which is why §35 had to grow it, and why explain.html gives it a plate of its own.

• Jaeger-LeCoultre Memovox, Grail Watch Wiki — where the caliber history above comes from. A collectors' reference, not a factory document, and the dates should be read as such.

The dial's smoke — A. Lange & Söhne "Lumen". Zeitwerk "Luminous" (2010) was the first, then Grand Lange 1 "Lumen" (2013), Grand Lange 1 Moon Phase "Lumen" (2016) and Datograph Up/Down "Lumen" (2018); the coating is patented. It is a semi-transparent coating on a COLOURLESS sapphire dial: it attenuates the visible spectrum so the dial still reads with contrast, while passing the ultraviolet that charges luminous material — including material under the dial, on date discs, plates and wheels, which is why a Lumen shows its own works. Two things are taken. The MODEL: the smoke is a coating on clear sapphire, not a tinted crystal, which is what §220's smokedGlass(T) builds and what the entry argues at length (a bulk tint would be thickness-dependent and is the wrong physics for this). And the LOOK: light print on a smoked ground, which §196's ink solve arrives at by itself as the ground darkens. What is not taken is the reason the coating exists. The stack is not public and is not guessed at here; and this movement has no luminous material anywhere, so it takes the tint and not the point of it — docs/BUILT.md §220 says so under "Not modelled", and a lume entry would be its own §. The transmittance dial.plate.smoke is this project's own knob, derived against its own legibility gates; Lange publishes no figure for it. Only the qualitative "most of the visible light" is borrowed, and only to put the Lumen-dark coat below T 0.5 — which is where the shipped 0.45 sits (§222).

• A. Lange & Söhne, The "Lumen" models
• A. Lange & Söhne, Datograph Up/Down "Lumen"

The hidden alarm hand — Parmigiani Fleurier Tonda PF Chronographe Mystérieux. 2026, cal. PF053. The idea is a hand that vanishes by SUPERIMPOSITION rather than by being taken away: at rest its chronograph hands lie exactly under the time hands and the dial reads as a plain time display; the pusher splits them out and they appear. This movement's alarm hand is cut shorter than the hour hand so it hides exactly beneath it, and while the alarm is disarmed it tracks the hour hand through a heart cam and a sprung follower — a rattrapante's own mechanism, which is what §25 C calls it — swinging out to the set time when the alarm is armed, and visibly travelling both ways. The differences are large and worth keeping in view: Parmigiani's is a chronograph across five coaxial hands on a triple clutch, where this is one hand on one heart cam; and theirs hides a complication that is merely idle, where this hides one that is switched OFF, so here the hidden state is the resting state.

• SJX, Hands On: Parmigiani Fleurier Tonda PF Chronographe Mystérieux
• Parmigiani Fleurier, Novelties 2026

A note on what these citations do and do not license. They describe mechanisms; they are not evidence that this movement reproduces them. What holds the movement to its own claims is the instrument battery and TODO.md, not this list — and where a source describes something the build does not do, the honest place to find that out is the debt item, which is why two entries above point at one.

This project was made possible by Claude, Anthropic's AI model. The movement, the layout solver, the inspector and every document in this repository — this file included — were designed and built in collaboration with it, and the reasoning behind each shipped decision is recorded in docs/BUILT.md rather than left implicit.

That is a statement about authorship, not a warranty. Nothing here asks to be believed because of who wrote it: the constants are derived from stated constraints, the derivations are written next to them, and the checks above are what hold them — which is exactly why so much of this repository is instruments and so much of TODO.md is the gap between what is modelled and what is genuinely simulated.

Copyright 2026 timesim authors. Licensed under the Apache License, Version 2.0; you may not use this project except in compliance with it.

The Apache grant covers this project's own code — src/, index.html, dev_server.py, test-geometry.html, and the documentation. It does