Marine Chronometer model · verification for a build · 28 October 2026 · with it: the model against real chronometers

Would a chronometer made to the model run and keep time?

The aim has changed from parity with the 1948 Hamilton Model 21 to a working marine chronometer, made with a CNC and a 3D printer, steady enough to navigate by. The model already agreed with the manual and with recordings of real Model 21s; this report covers the checks added to answer the maker's question: will the gears run when the holes are bored by a machine, will it hold together and stay in adjustment from cold to hot, does the hairspring do what its design says, and how will the made chronometer be tested and rated. Every figure here is computed by a script in the repository, named beside it, from the model's own geometry and numbers.

Approach

A chronometer keeps time if three things hold: its train delivers a steady torque to the escapement (so the balance's swing, and with it the rate, stays put), its geometry stays in adjustment under load and temperature, and its balance and hairspring oscillate at a rate that changes little with swing and temperature. The model's existing checks cover the escapement in detail (its figures against the manual, its motion in time, the energy budget, the stresses and an error budget for a voyage). Nine checks were added for the build:

#QuestionMethodScript
1Do the gears mesh, and how exactly must the holes be placed?Tooth-by-tooth contact on the model's own cycloidal outlines, with friction; the train run in time through the escapement's equationtools/gearing.js
2Does it stay in adjustment from 40 to 100 °F?Differential expansion of brass against steel through the depths, the endshake and the escapement solvertools/thermal.js
3Do the plates and bridges hold the arbors where they belong?Linear finite elements (voxel bricks, multigrid) on every part that carries a pivot, assembled by superpositiontools/fea.py
4Does the hairspring behave as designed at large swings?Co-rotational beam elements, Newton's method, against the existing shooting elasticatools/hairspring-fe.js
5What must the made parts measure?The model's figures turned into bench tests with pass figures, in the build booktools/bench.js
6Does the made chronometer keep time?The Navy's 30-day test worked on the maker's log; the rate fitted, backtested and projected for a voyagetools/rating.py
10How much does the mainspring lose to its own coils, and does the fusee still even it?The coils' slip followed on the solved spring from state to state, the packed coils' pressure by the ring equations; the fusee's output runningtools/mainspring_friction.py
9Do the gimbals keep it level at sea, and does anything resonate?The case and ring weighed on the model, run as two rigid bodies on their pivots under a ship's random roll, pitch, yaw and heave; what reaches the escapement run through it in timetools/gimbals.js, tools/massprops.py
8Does a knock, or a swing too large, let a tooth past the locking stone?The detent as a frame of beam elements from the drawn parts, its modes carried in time with the balance and the wheel through elastic contacts; validated against the rigid detenttools/tripping.js

Units: millimetres and micrometres (µm), newtons, seconds a day (s/d, + gaining). The swing is the balance's amplitude, each way; the manual asks for 1⅜–1½ turns, 247.5–270°, and the model runs at 255°. A rate of 4 s of time is a mile of longitude at the equator.

1. The gears, tooth by tooth

The going train is four meshes: the fusee wheel drives the centre pinion, then centre → third → fourth → escape pinion. The model draws its teeth to BS 978 Part 2's clock proportions (cycloidal wheel teeth: an epicycloidal addendum, radial flanks; pinion leaves with round tips), which the manual doesn't give and the model estimates. gearing.js takes those outlines from the model's own code (core.js gearOutline), sampled finer (400 points an addendum, 240 a tip), and places wheel and pinion at their centre distance m(z₁+z₂)/2. For each of 120 steps through one pitch of the wheel it turns the pinion back against its load until the outlines touch (bisection to well under a micrometre), takes the contact's normal from the touching outline and the slip from the two bodies' velocities, and so the torque the pinion receives against the wheel's, with tooth friction µ (0.10–0.20: hardened steel on brass, unlubricated, the range physics.js uses). A ratio of 1 is a perfect, frictionless mesh.

Figure 1. Torque the pinion receives over one pitch of the wheel, against the nominal ratio, µ = 0.20. The ratio is highest at the pitch point and falls as the contact runs away from it, most before the line of centres (left), where friction acts against the drive; the fusee mesh's deep dips are where the wheel's tooth, capped at 1.15 module, works on the leaf with its corner. The small zigzag (about 0.003) is the outlines' sampling, not the teeth. gearing.js, stage().

None of the four meshes drops (the next tooth always engages before the last lets go), about a fifth of the action is before the line of centres, and the mean loss is 2–5 % a mesh. The escape wheel's torque is the product of the four ratios as their wheels turn (the centre once an hour, the fusee wheel once in 6.4 h); the balance's swing follows it through the escapement's own equation (shared/escapement.js: A² relaxes towards ampAt(drive)² with the balance's free run-down time TF = 67 s), and its rate through rateAt.

Figure 2. The balance's swing over half an hour of the train, the teeth's ripple included, at the two friction coefficients; the band is the manual's 247.5–270°. The slow wave is the fusee wheel's teeth (one every 4.3 minutes), the fast one the third and fourth. Over six hours the swing stays within the same limits (). gearing.js, day().

The ripple moves the rate by a few hundredths of a second a day for minutes at a time and repeats every day, so a navigator's daily rate takes it in; the time it wanders about its mean is . What matters is that the swing never leaves the band.

The depthing window and the CNC's tolerance

Each mesh's centres were then moved, the others left as drawn, until one of three things happened: it jams (the wheel's tips reach the pinion's root, or no backlash is left), its extra friction takes the day's mean swing out of the band with the set-up as drawn, or, the set-up restoring the mean (Op. 13), its ripple takes the swing out at some moment.

Figure 3. Each mesh's depthing window about its drawn centres (bar), the closer side ended by jamming, the wider by the ripple or the search's end at one module; the dark mark is the CNC's tolerance band, ± the hole tolerance doubled plus the pivots' shake. gearing.js, gearing.json.

Each arbor has two holes, each off by up to t in any direction, and its pivots shake about 10 µm in their holes (estimated: 1–2 % of a jewelled pivot's diameter), so a mesh's centres close by at most 2t + 10 µm. Aimed at the drawn centres (BS 978's depth, kept for the teeth's sake: wider would leave less engagement for wear and knocks), the tightest mesh sets t = ± µm: a good CNC holds it with the holes bored, not drilled, and checked with gauge pins. Every mesh takes several times as much error wide as close, so a hole that errs should err wide.

2. From 40 to 100 °F

The movement is brass where it holds things (plates, bridges, pillars, wheels) and steel where it turns or locks (pinions, arbors, escape wheel, roller, detent). Brass grows × 10⁻⁶ a °C more than hardened steel (published ranges: leaded clock brass 18.7–20.9, carbon steel 10.8–12.5). From 20 °C, where the movement is put together, to 40 °F and 100 °F (wider than the Navy test's 55–90 °F, for a cabin at sea):

A mesh's depth changes only by the pinion's radius times the difference, because the wheel and its hole are both brass. The endshake, over the train's 16.8 mm span (measured on a side photograph), changes by : set at µm at 20 °C, it stays inside Ops. 15, 69 and 74's 0.001–0.003 in at both ends. In the escapement, the balance's and the detent foot's places are brass-held and the wheel, roller and detent are steel; makeEsc rebuilt with those distances grown moves lock and drop a tenth of a degree and the rate a few hundredths of a second a day. The temperature error is the balance's and its spring's alone, which is measured, not computed (section 5).

3. Finite elements on the plates and bridges

The largest loads in the movement are the fusee chain's pull (the mainspring's moment on the barrel's pitch radius, physics.js) and the first mesh's tooth force. fea.py takes the six parts that hold pivots from the page itself (each part's solid, closed and consistently wound, solids.py): the pillar plate, the train, barrel, escape and lower bridges, and the bridge on the dial side carrying the third's and fourth's lower jewels. A tetrahedral mesher (gmsh) was tried first; it found self-overlapping facets in the page's surfaces, so each solid is instead cut into voxel bricks (a ray up each column finding where it is inside), 0.25–0.55 mm in plan and 6–14 layers through the thickness. Trilinear bricks, brass at E = 100 GPa (97–110), ν = 0.34, assembled in numpy and solved by conjugate gradients preconditioned with smoothed-aggregation multigrid (pyamg), the six rigid-body modes given.

Figure 4. The solver checked: a brass cantilever 20 × 2 × 2 mm, 1 N at its tip, against Timoshenko's deflection PL³/3EI + PL/κGA, as the bricks get finer through the depth. Trilinear bricks are stiff in bending until fine; the parts here have 6–14 layers. fea.py --check.

Loads, every stage lossless (the largest): the chain's pull along the external tangent of the fusee and barrel at its height on them (fully wound on the fusee's top turn, run down on its bottom), and at each mesh the tooth force, tangential from the torque and radial from the mesh's mean pressure angle (section 1, 9–11°). Each is split between its arbor's two bearings by the lever rule; a pivot's load goes on its hole's wall over the half it bears on, as cos θ. Supports: a bridge is held where it is screwed (every node within the screw head's radius, through its thickness); the pillar plate at its rim, where the case's seat ring takes it. Assembly by superposition: each fixing's reaction is put on what is beneath it (another part, or a pillar, a 6.7 mm brass rod fixed at both ends, bending as a beam); each part then follows its fixings as a rigid body fitted in plan, plus its own bending. The run is repeated with coarser bricks as a convergence check.

Figure 5. The barrel bridge fully wound, in plan as the dial side sees it (3 o'clock right, 12 up): each 0.6 mm column's largest in-plane movement. The chain pulls the fusee's bushing and the barrel's toward each other; the bridge is held at its three screws, round its rim. fea.py --json.
Figure 6. How far each arbor's two bearing holes move under load, both cases; the hole tolerance from section 1 is ±13 µm, off the scale. fea.py.

The plates and bridges are many times stiffer than they need to be: the hole positions and the depthing are what the maker must hold, not the parts' stiffness.

4. The hairspring, by a second method

The model's hairspring is designed, not drawn by eye (shared/hairspring.js): a helix of 12 turns, r 5.1 mm, with terminal curves solved to Phillips' conditions (1861) so its pull on the balance is a pure couple, and checked at large swings as an inextensible planar elastica solved by shooting (hairspring.js). hairspring-fe.js solves the same path a different way: co-rotational beam elements (Crisfield 1991, ch. 7; each element straight between its nodes, axial and bending stiffness, the tangent consistent), the strip extensible (0.221 mm thick: EA/EI = 12/t²), the stud's end clamped and the collet's end turned about the staff in 2.5° steps to ±300°, each step to equilibrium by Newton's method with a block-tridiagonal factorization.

Figure 7. The couple's departure from linear, each method against its own stiffness at 5° (parts per million): the spring stiffens a little at large swings, which makes it gain there. hairspring-fe.js, hairspring.js.
Figure 8. The force the spring leaves on the balance's pivots, as a share of couple / R (0 for a pure couple, about 1 for a spring with plain ends).

The two agree on the couple's shape to a few parts in 10⁷ and on the force to two figures; their absolute stiffnesses differ by 5 × 10⁻⁴, the elements' chords being shorter than the curve. The spring's own isochronism, the rate it gives across the manual's band, is . What neither holds: the strip's own mass, and the elastic modulus's change with temperature, which is the spring maker's to set and the bench's to measure.

5. Tests on the bench

The analysis says the design works; only tests on the made parts say a chronometer made to it does. bench.js turns the model's figures into tests with pass figures; makers.py puts each on its part's card (its maker's sheet) and the whole list in the build book's last chapter (Making it → Build book). benchcheck.py judges the maker's measurements against the same figures (a JSON file it makes to fill in) and says what to do where one fails: the set-up turns that bring the mainspring into its band or even the fusee (from the curve's slope and the fusee's 1.96:1 profile), the timing weights or the millimetres to take at the stud, the manual's operation that sets an escapement figure. Its --demo plants eight faults in a made-up set and finds all eight.

6. The rating trial

The made chronometer is rated as the Navy rated the Model 21 (Sec. IX: thirty days in six periods of five, at 90, 72½, 55, 55, 72½ and 90 °F, compared daily with a time signal). rating.py reads the maker's log (the time, the chronometer's error, the temperature) and works the test's figures exactly as the model's test card does (app.js testCalc), fits the daily rates as a constant, a linear drift and a temperature curve with the rest as day-to-day scatter, backtests a ten-day rate carried on through the log itself, and projects 2000 voyages. Below it is run on a made-up log with known properties (the error budget's chronometer like No. 3390: rate +0.35 s/d, scatter 0.042 s/d, a temperature curve), sixty days, the first thirty the Navy's periods.

Figure 9. The made-up log: the chronometer's error against the time signal, the shaded spans the Navy test's periods and their temperatures. rating.py --demo.

The fit recovers what the log was made with (the rate to 0.01 s/d, the scatter to 0.002), so on a real log its figures can be trusted as far as the log goes. Good enough to navigate by is taken as 95 % of projected voyages under 30 s (7½ miles) after 30 days.

7. The build's deviations from the Model 21

The checks above found nothing that needs changing for the chronometer to work; two changes, both with 1950s materials, make it easier to make and steadier to keep. Each is on its part's card in the page as the build's, with the Model 21's original beside it and a button back to it (the link keeps orig= for one set back); the 3D model stays the Model 21. tools/deviations.js works their numbers and writes deviations.json, which makers.py puts on the sheets.

The balance's temperature compensation, from Table IV

The build keeps the Model 21's own balance, a stainless rim on an Invar arm, and takes only the hairspring in Nivarox. Until 13 October 2026 it took a one-piece Glucydur balance too; that has nothing to trim, so the spring's thermoelastic coefficient would have had to be held within ± × 10⁻⁶ a °C to meet the Navy's 1.20 s a day between 55 and 90 °F, about a seventh of the ±10 a bought spring holds. On the Model 21's balance, Table IV's screw moves (Op. 8) trim what the spring misses.

Method. The rim grows as stainless, while the arm's ends hold it to Invar's growth, so the rim is pulled in at the arm and bulges at the quarter, as a thin ring under two opposite forces does: its radial deflection goes as Σ cos nθ / (n² − 1)², n = 2, 4, 6 … (the ring's bending by Fourier series, which gives Roark's 0.1366 and 0.1488 P R³ / E I for its two diameters). A screw pair of mass M at radius r moved from the quarter (hole 7) to hole n then changes the rate in heat by M r C (S(θn) − S(θ7)) ΔT / I of a day. The one unknown, C, is fitted to all forty of Table IV's figures (holes 8–12 under eight heads), each pair's mass from the parts list. Then the whole balance's expansion (rim, arm, hub, the standard screws and the weights) gives the coefficient the spring needs, and every placing of the four standard pairs in holes 3–12 gives what Table IV can trim.

Figures.

Estimated: Nivarox's expansion (7–10 × 10⁻⁶/°C) and modulus (165–195 GPa), as Elinvar-type alloys'; Invar 1.0–1.6 × 10⁻⁶/°C; the ring taken as inextensible and the arm as rigid; the 0.120 in head's mass extrapolated from the parts list's; I_REST (the 8 % of the moment the drawing lacks) anywhere between Invar's growth and the rim's. A bought spring's ±10 × 10⁻⁶ is an alloy's target (EP3502288A1), not a grade's published tolerance. The temperature test on the bench (the rate, the last of section 5's tests) settles the pair and says which pairs to move.

8. The detent under shock: does a tooth ever get past?

The classic way a detent escapement fails at sea is tripping: the escape wheel lets a tooth go without the balance having unlocked it, and the chronometer gains a beat. The detent is the weak point: a thin blade on a thinner spring, held on its stop by a preload of a few millinewtons and nothing else, with the locking stone at its end. dynamics.js takes it as one rigid body turning about the root of its spring; that answers whether it is home in time, but not whether the blade, still ringing from its landing or from the tooth's lock, lifts the stone off the next tooth, nor how hard a knock must be once its own bending is counted. tripping.js takes the detent as what it is.

Method. The detent becomes a plane frame (): beam elements with stretching, bending and their consistent mass for the two-strip spring, the blade, the arm out to the horn and the trip spring, and rigid bodies for the cross-piece with its bracket, the jewel block with the stone, and the horn; every size is the drawn part's (movement.js, ESC.pieces), every material dynamics.js's (beryllium copper 128 GPa, Elinvar-type 180 GPa). Its modes, clamped at the foot, carry it in time with the balance and the escape wheel as dynamics.js does, but every contact is elastic: the discharge jewel's own outline on the trip spring's tip (pushing it onto the horn on the active swing, bending it away on the return), the trip spring on the horn, the blade on its stop button, the tooth on the locking stone (the stone's outline and draw from makeEsc, so the tooth is pushed back as the stone lifts) and on the impulse jewel. A contact is a stiff spring with a dashpot set for its restitution (penalty contact). Op. 78 sets the preload: the spring is bent so a 0.770 g weight on the locking jewel parts the blade from its button by approximately 0.005 in (p. 61), which leaves it 0.56 g of preload at the jewel as drawn and none for a spring 0.122 mm thick (the spring 3.30 mm long, measured since 20 October 2026: 0.63 g and 0.137 mm at the 3.82 drawn before).

Validated. Made rigid and hinged at the spring's root, with the draw counted as dynamics.js counts it, the frame is dynamics.js's detent (stiffness, preload and first mode to 0.1 %) and runs as dynamics.js does: . Two things the frame finds that the rigid model leaves out:

Figure 10. The locking stone through one oscillation at the running swing, each run from 10 ms before its unlocking, as a share of the depth of lock (the depth of lock, µm at the stone): the highest each 0.1 ms. The shaded spans are the unlocking (the discharge jewel lifts the stone clear, past 100 %, off the chart) and the return swing (the jewel bends the trip spring). tripping.js, section 2.

Running. The largest lift while a tooth is held comes ; by the time the next tooth arrives the stone has settled (). None of the estimates moves that far (the table below: the detent's damping, every contact's restitution and stiffness, the blade's stiffness, the modes kept).

Figure 11. The swing pushed by raising the drive: the stone's largest lift while a tooth is held, the frame as drawn. A dot on the top line is a run where a tooth got past. The dashed line is makeEsc's TRIP, where the discharge jewel comes round to the trip spring a second time in the swing; the band is the manual's 247.5–270°. tripping.js, section 3.

Pushed and knocked. Knocks are found the way dynamics.js finds them, the smallest that lets a tooth go or loses a beat at the worst of eight moments in the swing, for the frame and for the rigid detent by the same runs:

A chronometer in its gimballed, padded box meets a few g at sea; the margin is wide. What the frame leaves out: the detent's motion out of the plate (its spring is 0.95 mm deep against 0.079 thick, twelve times stiffer that way, and moving that way doesn't unlock), and the true restitution of each contact, which only the made detent can show.

What the bench checks

Two things the frame shows a maker can measure. The lock (Op. 85): run at shallower locks (each in its own process, makeEsc's depth of lock changed, the frame as drawn and with every contact four times stiffer, 20 beats each), the stone's largest lift while it holds a tooth stays at whatever the depth, so a shallower lock leaves less margin: The blow that trips it hardly changes with the lock (the first table below): against knocks it is Op. 78's preload that holds the detent, which makes the weight test the one to do carefully. The detent spring, whose section no source gives: clamped by its foot and flicked, the detent rings at its first mode, , and the pitch rises nearly as its spring's thickness to the 1.5, so a phone's recording measures the spring (benchcheck.py --ring reads one). Both are in the build book's tests chapter and on the detent's card; benchcheck.py judges them.

The largest lift over 20 beats is a maximum of a chaotic chatter: two runs a hair apart differ by about ten points. The figures are margins, not sharp limits; the lock's floor is where the margin falls below two, and the spring's band is the drawn spring to the thickest that stays under that margin and that Op. 78's test can still preload (0.137 mm, tolerances.js) ().

9. The box and gimbals in a seaway

A chronometer at sea hangs in gimbals so that the balance's staff stays upright whatever the ship does. The checks above took that as given. Here it is worked out: does the gimballed case keep the movement level, can it be set swinging by the ship or by the balance's own beat, and what reaches the escapement?

Method. The case with the movement in it and the gimbal ring are weighed on the model (massprops.py: every solid integrated at its material's density, for mass, centre of gravity and inertia about the gimbal centre), then run as two rigid bodies on their pivots, the ring on the box's at 3 and 9 and the case on the ring's at 6 and 12 (gimbals.js: Newton-Euler about the gimbal centre, which moves with the box; RK4). The ship's roll, pitch, yaw and heave are narrow-band random about each motion's period, the box some metres above the ship's centre of motion. Merchant ships roll at periods of 10-30 s (MARIN; Jachowski et al.); NORDFORSK's operability limit puts the box's athwartships acceleration at 0.12 g rms at the bridge, and one sea is set to it. The other levels, a small vessel's motion and the box's place in the ship are estimated. What the escapement feels, the force in its plane and the turn about the staff, is run through dynamics.js for 100 s against still water.

Figure 12. The case's lean from the apparent vertical for each degree of a regular roll, against the roll's period (the case rocked about the box itself). The shaded spans are merchant ships' roll periods and a small craft's (estimated); the gimbals' own periods and the balance's beat are marked. gimbals.js, section 2.

Figure 13. Two minutes at NORDFORSK's limit: the case's lean from the true vertical and from the apparent vertical (gravity less the box's acceleration), which is what the balance's pivots feel. gimbals.js, section 3.

10. The mainspring's coil friction

A mainspring takes more torque to wind than it gives back running: its coils slide on each other as it winds and unwinds, and rub. The energy budget had allowed for it as a published 5-15 %, the largest number in it taken on trust. Here it is worked out on the model's own spring.

Method. mainspring.py solves the strip's shape at every quarter turn of the wind as the one of least bending energy, its coils kept apart but frictionless (js/mainspring.js). Each of those states is polished by the same solve (mainspring_friction.py). The slip follows the steel: the two pieces that touch, on a coil and on the one over it, are followed from one state to the next, and the difference of their movements along the strip is how far they slide on each other; on the barrel's wall against the barrel's turn, on the core against the arbor's. The pressure between packed coils comes from the packed spiral's ring equations: each coil a ring bent by its own moment, EI(k − k₀), the torque the same at every section, so each coil's hoop force is (τ − M)/r, and each ring's radial balance passes the pressure on to the next from the pack's free edge (the solver's own multipliers and the rod's equations by differentiation were tried first, and failed their own checks in the packs: the second derivatives of a sampled shape are too rough). The friction's work, μ times pressure times slip, over each quarter turn is a torque that adds to winding and comes off running. μ: greased steel on steel, 0.08-0.2 lubricated (the CRC Handbook, as cited in US 3,939,081); 0.10 taken, 0.08 and 0.16 tried.

Figure 14. How much more winding takes than running gives, over the spring's turns from its inner end, at three friction coefficients. The shaded span is the fusee's working range, from run down to fully wound; past 19.35 turns the coils are packed solid. mainspring_friction.py.

Figure 15. The fusee wheel's torque running, over the fusee's turns from full wind, against its mean: the solved spring without friction and with it. The band is the bench's, the escape wheel's torque that keeps the swing in the manual's 247.5-270°. mainspring_friction.py.

The build's strip, in Nivaflex

Method. The build's mainspring (section 7) is Nivaflex 45/5, 0.386 mm thick for the model's torque (its modulus 220 GPa against carbon steel's 207), the same length, width and free curve (a spring set the same way). mainspring.py --nivaflex solves its states over the wind as for the model's spring (mainspring_t0.386.js, beside the tool: the page still draws the model's), and mainspring.py --bound finds, for both strips, where the coils pack solid to a twentieth of a turn: the states past the last solved one, each solved warm from it and from an even spiral (the quarter-turn table alone had put the model's room at 0.13 turn). The fusee's own rule then sets the spring up (where its pull falls by the chain's lever over the turns the chain takes to leave the fusee; checked to reproduce the page's full wind for the model's spring), mainspring_friction.py --nivaflex works the coils' friction, and the swing over the wind is the escapement's (ampAt) at the fusee wheel's torque running against its mean.

Figure 16. The balance's swing running, over the fusee's turns from full wind, at μ 0.10: the model's strip and the build's Nivaflex. The band is the manual's, 247.5-270° (1⅜-1½ turns). reportdata.js from mainspring_friction.py.

The model's strip against the sources' proportions

Why. The room in turns depends only on the spring's proportions, not on its size: the packed coils' pitch, the core and the length against the barrel's inside radius. The sources give exactly those ratios without needing the barrel in millimetres: the videos' packed coils against the barrel's inside edge (KLUwI2UUCMQ 15:54 and 33:06, 1/41.0-1/41.8 of the inside radius; TN24H 3SFDplGq6vs 0:16, 1/43.3-1/43.7), the length over the pitch from the packs' area (2,630-2,790), the core at 0.30-0.34 of the inside radius; and the British Museum's Model 21 (Randall and Good 1990, cat. 117A: strip 0.44 mm, barrel 38 mm inside, arbor 13 mm), whose strip is 1/43.2 of its inside radius and its core 0.34. Method. Each set of proportions is put in the model's barrel (16.3 inside) with the model's free curve and fusee and solved as above. The pitch the videos measure is the coils' spacing packed, the strip and what lies between: the solver keeps packed coils its film apart (none now; 0.01 mm until 12 October 2026), so a strip matching a pitch is the pitch less the film. Until 12 October 2026 the model took the videos' pitch, 0.394, as its strip and kept the film on top, so it packed at 0.404, 1/40.3 of the inside radius: coarser than every source. It now packs at the pitch (the strip 0.394, no film: no source gives one, and the pitch is the strip at most). The texts agree on what a spring wants: the arbor a third of the barrel's inside (Britten, Watch and Clockmakers' Handbook, 1896, p. 229: "for watches and 2-day marine chronometers"), the strip 1/30 of the arbor's diameter for chronometers (Rozé, in Saunier, Traité d'horlogerie moderne, 1887, §1214), and for a fusee about half a turn spare past full wind (Britten); turns in a real barrel are "always less" than the ideal (J. and H. Grossmann, Lessons in Horology, 1905, §§105-111), which the solve includes.

Figure 17. The balance's swing running over the fusee's turns from full wind, at μ 0.10, for the model's strip and the sources' proportions in the model's barrel. The band is the manual's, 247.5-270°. mainspring.py, mainspring_friction.py --variants.

11. Every mechanism, each in turn

The sections above follow the going train and the escapement. This one takes the rest of the movement's mechanisms in turn: the key and the winding stop, the ratchets and their pawls, the sustaining spring, the motion work and the hands, wear at every sliding contact, the train in time behind the stopping and starting escape wheel, and the whole movement made to a maker's tolerances at once. Each is its own tool; the figures here are those tools' own (mechdata.js).

Their geometry. mechanisms.py reads off the built model what its constants don't give: each pawl's contact on its ratchet, found by turning the ratchet the way its load turns it until its teeth first reach into the pawl's outline (only a contact that grows counts, since a drawn outline may sit a few micrometres into a tooth's back; the load's way is the one whose first contact falls on a face or a tooth's tip corner), the pawls' masses and inertias from their outlines, and every arbor's moment of inertia from its solids at their densities (tools/mechanisms.json).

11.1 The key forced

Method. What a hand puts on the key's paddle ( mm wide): finger-and-thumb torque on small handles, men tightening flanged handles 9.5 and 19 mm wide (Imrhan and Munoz-Vasa 2013) interpolated to the paddle, and Swain's 95th-percentile men on a 12.7 mm knob (1970, in the same paper); the larger is taken as a strong hand's. Wound against the stop, the stop-bar and its stud take the key's torque less the chain's (the stop holds the fusee, so the chain keeps the mainspring's pull); turned the wrong way, the key's torque adds to the chain's through the winding pawls and the sustaining spring into the train, which the escapement holds still. Each part's stress against its material's range, as strength.js does it, and the key torque at which it reaches yield. overwind.js.

Figure 18. The key torque at which each part in the key's load path reaches the weak end of its yield, wound against the stop (blue) and turned the wrong way (orange); the dashed line is a strong hand's. overwind.js.

11.2 The ratchets and their pawls

Method. Each pawl as the model seats it under load. The line of action through the contact against the pawl's pivot gives the draw: a load drawing the pawl in seats it whenever the draw angle exceeds the friction angle (μ 0.10-0.20, oiled steel). The force is the ratchet's torque over the line's arm about its centre; the tooth is bent at its root as a triangle on 0.97 of the pitch, the face's bearing taken over the pawl's depth in the teeth (a fitted end), the pawl bent by its pivot and its pivot sheared. Each spring as a cantilever (3 E I / L³): the force it adds as the pawl rides over a tip, and, for the winding pawls, the force they need to follow the fusee's teeth down when the key lets go and the mainspring turns the fusee forward under them. ratchets.js.

Figure 19. Each pawl on its ratchet as the model seats it, in the ratchet's frame: the teeth, the pawl's outline, its pivot, the load on it as drawn (solid) and with its end fitted to the face (dashed). mechanisms.json, ratchets.js.

11.3 The sustaining spring

Method. The spring's stiffness from its section: a band bent in its plane between the pins that hold its two lobes (each pinned once, so the band is loaded by a pair of equal forces along the line through the pins), by Castigliano, F ∫ h² / (E I) ds along the band from pin to pin, h a point's distance from the pins' line and I its section's there (the band traced on KLUwI2UUCMQ 27:26 tapers, 1.3-1.9 mm wide, its height read on 28:42.5 and 28:43.0). Then, the key held, the train runs on it, the fusee wheel turning at its rate and the spring opening, the balance's swing following its equation (A² towards ampAt(drive)², the free balance's TF) and the escapement's rate at that swing. Since 14 October 2026 the spring is a split ring weaker than the drive, as the period's chronometers had it (Gould, 1923; Glasgow, 1885): the fusee's pull holds it closed on its banking in running, and when the key takes the fusee it pushes what its section gives at its measured 17.4° gap (traced by hand; 12.5° until 18 October 2026, the first reading), falling as it opens (the model's SPU, which this checks against the section). The manual's "enough power to drive the chronometer for five to ten minutes" (Sec. IV) is read as the time it keeps the chronometer running, and as the least it gives (since 18 October 2026: the measured outline and height give about ten and a half) (since 17 October 2026, when the spring was traced on the video; the time it kept the balance over a turn until then, which the designed band gave and the traced one can't). sustaining.js.

Figure 20. The balance's swing while the key is held, on the spring as drawn (its section's push) and as the model runs it (SPU); the two lie together. The band is the manual's, 247.5-270°; 180° is a turn; under 41° the escapement can't unlock and the train stops. sustaining.js.

11.4 The motion work and the hands

Method. The motion work's two meshes and the wind indicator's, tooth by tooth on the model's own outlines as gearing.js works the going train; the hands' play from their meshes' backlash; and the cannon pinion's friction on the centre arbor (Op. 58), between what a knock asks (the hands' out-of-balance, off the built model, at the 45 g that already frees the escape wheel) and what setting the hands forward allows (the slipping torque adds to the train's drive, and must not swing the balance to the trip, past a full turn, within a minute's setting). motionwork.js.

Figure 21. The cannon pinion's friction (mN·m): above the knock's line the hands hold; below the setting's lines the balance stays short of the trip; the bench's aim in the middle. motionwork.js.

11.5 Wear

Method. Every sliding contact by Archard's law (the volume worn is k F s / H: the wear coefficient, the load, the distance slid, the softer part's hardness; Archard 1953), k from Rabinowicz's table for lubricated metals: about 10⁻⁷ well oiled, 10⁻⁶ poorly, 10⁻⁵ once the oil has gone. Each depth a year against the tolerance it eats: a pivot hole worn open moves its arbor toward the load and changes its meshes' depth (each mesh's close side less the hole tolerance and shake: gearing.json); the balance staff's pivot ends wear the endshake; the escape teeth's tips and the escape arbor's pivot wear the depth of lock toward its floor (tripping.json). wear.js.

11.6 The train in time

Method. The train's inertia at the escape wheel from the solids (each arbor with what it carries), and dynamics.js run again with it. Then the train as five bodies on four meshes, each a spring (the teeth as cantilevers, estimated) with its backlash, driven at the fusee's even torque, the escape wheel moved as dynamics.js's escapement moves it through a beat: locked, let go, catching the impulse jewel and carried at its speed while its drive pushes it, let go, landing on the stone. traindyn.js.

11.7 The whole movement to a maker's tolerances

Method. Movements made with every tolerance drawn at once (the holes, the meshes' friction, the mainspring's and the escapement springs' strips, the sustaining spring's band, the pawls' springs' set), each judged on all of the above: the train turns, the balance swings in the manual's band, the maintaining work drives five minutes, the winding pawls follow the fusee down and the sustaining pawl's wire presses it in (each pawl against its own need, 11.2). Judged first as made, then after the bench (section 5's tests, by bench.json's own figures) as a maker would apply it: the depths test re-bushes a mesh closer than its window less the pivots' shake (benchcheck.py), the pawls' test bends a spring under its least set, or a wire that leaves its pawl loose, back to the aim, the swing test has the set-up turned where the swing is out of the band (Op. 13), and the maintaining work is timed (nothing on the bench re-makes the sustaining spring, so a short one stays failed, but caught). A movement still failing after the bench must have been caught by one of its tests: one that isn't means the bench's figures fall short of the analyses', and the check fails. montecarlo.js.

Figure 22. The balance's swing as made, over the movements drawn (the escapement adjusted to the manual's figures, the set-up as drawn); the band is the manual's. montecarlo.js.

12. The estimates swept

Method. Five design values rest on estimates no source settles: the balance's free run-down, the detent's springs' thicknesses, the discharge jewel's back radius, the sustaining spring's band (traced on the video since 17 October 2026: swept over its readings' uncertainty), and the hairspring's upper terminal curve. Each is swept over the range its sources allow, and where the bench has a correction for it (the run-down test choosing the strip, the set-up, the ring test, the adjusting screws and the trip spring's tip, the maintaining test) it is applied as a maker would, to show that a chronometer made to the model works whichever value the real part has. estimates.js.

Figure 30. The balance's swing at each free run-down, with the drawn strip unadjusted and with the strip the run-down test chooses. The band is the manual's, 247.5-270°. estimates.js.

13. The mainspring the fusee asks for, and how much the balance loses

Why. Two readings the sources leave open are tied together: the mainspring's strip (0.394 mm drawn, from the videos' coils; the parts list's 0.419; the British Museum's 0.44) and the balance's damping, its free run-down TF (67 s in the model; 68 until 14 October 2026). The balance's steady swing fixes only their product, drive against loss, and the model's TF was set to give 255° with the drawn strip. So each is worked out here from evidence that doesn't involve the other: the spring from the fusee's measured profile, the damping from real balances whose swing changes on video (PLAN-drive-and-damping.md; OPEN-QUESTIONS.md 1 and "The balance's free run-down").

The spring the fusee was cut for

Method. The fusee's profile is measured (the groove's floors fit r₀/√(1−a·m) to 0.085 mm, a 1.96 : 1 taper; model README), and a fusee is cut to even one spring's pull: the spring it was cut for is the one whose pull falls by that taper over the chain's turns at a set-up near the manual's "about one turn" (Op. 53), and whose pull it evens best. Each candidate strip, 0.374, 0.394, 0.419 and 0.44 mm, at the videos' length over thickness (2,700), on the cores r 4.9 (drawn), 5.4 (a third of the barrel's inside) and 6.5 (the museum's 13 mm arbor), is solved in the model's measured barrel by mainspring.py (the strip's least bending energy over the wind from its measured free curve, each a variant beside the model's), its coil-bound found by mainspring.py --bound, its coils' friction by mainspring_friction.py, and drive.js then applies the page's fusee rule (the set-up where the pull falls by the taper over the 4.51 turns the chain takes to leave the fusee) and runs the escapement's averaged swing (ampAt) at the fusee wheel's torque against the model's spring's at 12 h, both running at μ 0.10. The earlier proportions' variants (section 10) are included. The taper's uncertainty: a bootstrap of the profile's fit puts the levers' ratio to ±1.4 %; the floors' turns placed a quarter turn either way along the cone move it ±4 % (the formula extrapolated over the last turn and a quarter, where it climbs steeply): the set-up is given at ±3 %.

Figure 31. The balance's swing over the fusee's turns from full wind at the model's TF, for each strip the fusee can set up, at μ 0.10. The band is the manual's, 247.5-270°. drive.js from mainspring.py and mainspring_friction.py.

The damping, from real balances whose swing changes

Method. How fast a swing settles after it is disturbed says how little the balance loses, almost whatever the drive. Two real movements show it on video, each read frame by frame: We1dLNXiBj0's balance let go by hand from rest (swing.py, 66° to 256° over 68 s, the absolute swing known to a screw hole), and xM_mxtGfeqQ's recovering after the movement was handled on its side (170° to 229° over 107 s, its 186° reading at 9:12 picked by the sound and the clicks' interval; References/VIDEOS.md). Each is run in the model's escapement in time (dynamics.js, through startup.js) from its first reading, at run-downs of 52 to 320 s, the drive free in 5 % steps; the best drive's miss at each TF is plotted. And the published figures for balances' Q, gathered from the literature (below).

Figure 32. How well each real swing is followed by the model's escapement in time at each run-down TF, at the drive that fits best (rms of the miss). The model runs at 67 s. startup.js.

Table 32. Published figures for balances' damping (Q = π TF / T; the loss a swing 2π/Q of the balance's energy). Read 12 October 2026; second-hand where marked.

Together

14. Every moving part in 3D

Why. Section 11 takes each mechanism in turn and strength.js each loaded part fully wound, by formula. Here every part that moves is one system in time, and each part's own stress is worked out on its solid under the worst load that system gives it.

Method. hamilton_transmission.js runs the whole transmission as the Wempe's wempe_transmission.js does (the barrel on its solved mainspring with its coils' friction, the chain in tension only, the fusee with its winding ratchet and pawls, the sustaining ratchet, pawl and spring with its bank, the going train tooth by tooth on the page's outlines, the escapement dynamics.js's own, the key and the stop) through four runs: a beat at 12 hours, the key let go outright and eased back, and wound into the stop. motion3d.py takes the page's solids and moves each part as the page's own update() moves it; motion3d_fe.py solves each part as finite elements (voxel bricks on its solid, a modulus by its material, finer windows at its contacts, the root's bending kept apart from the contact's pressure) at its worst load in the four runs, held where it is held; the springs (the mainspring by its solved curvature, the hairspring, the sustaining spring, the trip spring) from their large-deflection solutions.

Open the 3D viewer: every moving part in 3D through the four runs, coloured by its stress, its stress against its material's limit or its speed, the deviations magnified, with each part's load, margin and motion, the kinematics against the tooth counts, and the assumptions.

The checks the model already had

These ran before this work and still run in ci.py; their current output is below.

tools/escapement.js passes

The escapement against the manual's figures (Ops. 84-97), its running swing and its isochronism; the locking stone swept against the teeth.

its output today
ok  centre distance                                  9.40 mm   
  ok  roller shake (Op. 84)                           0.056 mm   about 0.002 in (0.051 mm)
  ok  teeth dip into the roller's crescent (Op. 76)    0.34 mm   > 0
  ok  lock (Op. 85)                                       6.1°   about 6°
  ok  let-off (Op. 86)                                    7.8°   at least 6°
  ok  overall (Op. 87)                                   28.3°   26-30°
  ok  drop (Op. 97)                                       2.1°   about 2°
  ok  horn clearance to the unlocking jewel (Op. 88)   0.25 mm   about 0.010 in (0.25 mm)
  ok  angle between the jewels                           88.6°   about 90° (Fig. 90)

  balance angle: jewel meets trip spring -27.4°, wheel released -21.2°, detent falls back -13.5°;
  impulse ends 20.7°; on the return swing the trip spring falls off at -41.8°. Lift at release 0.21 mm.
  running amplitude 255.0° (it needs 41.8°); the tooth lands on the impulse jewel at -2.8° and drives it through 11.17° of the wheel;
  unlocking takes 15.1% of the impulse's work. Escapement error +0.00 s a day against the model's settings (the escapement's own error -2.91 s a day, which the timing absorbs at the model's settings: impulse -2.40, draw -0.29, detent spring -0.91, trip spring 0.68).
  isochronism (the escapement's; the hairspring taken as isochronous): 247.5° 95% of the torque, -0.11 s a day; 255.0° 100% of the torque, +0.00 s a day; 270.0° 111% of the torque, +0.19 s a day.
  ok  the locking stone never inside a tooth through the period (10 swings, 41.8° to 340°, 20,000 phases each): deepest 0.36 µm at 62.5°, drive 15 %, the balance at -25.24° (phase 0.18385), tooth -2 (want under 1 µm)
  the detent leaves the discharge jewel at -13.98° and is home on its stop at -7.05° (at 255.0°); the escaping tooth's lock runs on from 20.72°.

tools/physics.js passes

The energy budget from the mainspring to the balance, the escape wheel's chase onto the jewel, the balance's damping, temperature from the materials.

its output today
A3: the energy budget
  the mainspring, solved: 0.998 N·m fully wound, 0.526 as the chain's last turn leaves the fusee (1.90:1 against the fusee's 1.90:1 over 4.62 turns); a linear strip of its section and length (84.4 mN·m a turn, E b t³ / 12 L, 2π) would give 0.824
  ok  its bending stress fully wound 2.85 GPa (6M / b t², from its set curve): under a set strip's limit, 1.5 times hardened carbon steel's elastic limit, about 2.55-2.85 GPa (a rectangle's plastic moment; the strip formed past its elastic limit, ICMFF9)
  the fusee's even torque: 0.388 N·m at its wheel (the pull 51.7 N on the chain)
  the train fusee wheel to escape wheel 2893:1; the escape wheel's torque 67-104 µN·m (each stage 0.85-0.95)
  A4: the escape wheel with mass
  ok  the escapement's wheel inertia (5.65 g·mm²) is the escape wheel's 4.46 and the train's behind it, reflected through the pinions (1.2: their solids weighed, mechanisms.json, dynamics.js)
  unlocked at -21.2°, the wheel starts from rest at 10942 rad/s² (61.8 µN·m, the model's): it catches the impulse jewel, already running away at 23.9 rad/s of the wheel, 18.5° of the balance later (at -2.8°), running 63.1 rad/s: it strikes the jewel
  ok  the impulse ends inside the swing (the tooth leaves the jewel at 20.7°)
  the escapement's own efficiency: 56 % of the tooth's work reaches the balance (the strike 33.3 % of the impulse, the unlocking 15.1 %, of it the detent's 7.8 %, and the wheel's speed lost at each lock)
  the balance at 255°: 0.905 mJ; it is given 30-55 µW, so it holds 255° if its Q is 207-378 (its free swing decaying in 33-60 s)
  ok  the budget closes: that Q is inside a balance's (about 100-400, mechanical watch and chronometer balances in air)
  ~~  the model's TF 67 s (Q 421, shared/escapement.js) is inside the budget's: it swings the balance 316° at each stage's 0.9, 269-363° over 0.85-0.95 (the manual's 1⅜-1½ turns, 247.5-270°)
  A6: friction from the measured parts
    the centre arbor (pivots r 0.5, 0.5): pivots 2.2-3.2 %, mesh 1.3-5.2 %: the stage 0.917-0.966
    the third arbor (pivots r 0.3, 0.31): pivots 1.9-2.8 %, mesh 1.5-5.9 %: the stage 0.914-0.967
    the fourth arbor (pivots r 0.25, 0.35): pivots 2.7-4.1 %, mesh 1.8-7.1 %: the stage 0.891-0.956
    the escape arbor (pivots r 0.089, 0.089): pivots 0.9-1.3 %, mesh 1.8-7.1 %: the stage 0.917-0.974
  the four stages together 0.685-0.869 (as 0.910-0.965 each), against the 0.522-0.815 the budget assumed (0.85-0.95 each)
  ok  the train's efficiency from its pivots and teeth lies inside or at the top of the budget's assumed range
  the balance's constant friction, measured (Jd2c3x8VKsE 14:04, 16:10): its free swing falls 0.16-0.238° a second, 32-47 nN·m (k c T / 4): Q 6759-10054 at 255°, 4-6 % of the model's loss there; a flat end bearing across its face at mu 0.10-0.15 would give Q 378-793 (the measured friction is that of mu 0.008-0.012 on it), a domed end 22141
  ok  the measured friction lies between a flat end's and a domed end's (the pivot's end bears over less than its whole face, or its oil film carries part of the load)
  the balance's damping: air on its screws, nuts and arm Q 761-909; the endstone and any constant friction 6759-10054 (measured); side friction from the spring 123627; the spring itself 1000-5000
  together the balance's Q is 405-710 (free decay 64-113 s); with the train's efficiency from its parts the budget needs 191-259 to hold 255°
  the swing they would give a spring without its own losses: 344-456° at the train's middle efficiency (the manual's 247.5-270°)
  with the mainspring's own losses (2.1-11 %: its coils' friction, mainspring_friction.py, to what mainsprings measure, running 10-20 % under winding), the budget needs 196-304: against the parts' 405-710, the model's Q 421 outside it; the swing 321-426° at the train's middle efficiency and 10 %
  ~~  the two overlap: the balance as built loses about what the train, as built, gives it, the mainspring's own losses counted (the model's TF 67 s, Q 421: inside the damping's range, inside the budget's)
  ok  the model's TF 67 s is inside the decay the balance's parts give (64-113 s): derived from the budget before, now from the friction and the air too
  B2: the detent spring, from Op. 78
  its preload 5.49 mN at the jewel (Op. 78's 0.770 g less the spring's 20.8 µN·m over its 0.127 mm bend at the stop button); to the release (0.21 mm at the passing spring's tip) that is 0.85 µJ, but the jewel carries the detent on to 1.37 times that lift: 1.24 µJ an oscillation, 7.8 % of the impulse
  ok  the detent's share leaves the escapement running at the manual's swing
  A5: temperature from the materials
  a plain steel spring on a steel balance: -10.3 s a day a °C (-5.7 a °F): why a chronometer with one needs a compensating balance
  the Model 21: for no linear error the spring's thermoelastic coefficient must be -22 to 10 × 10⁻⁶ a °C (the balance's expansion between the Invar arm's and the stainless rim's)
  ok  inside what an Elinvar-type spring is heat-treated to (about ±50 × 10⁻⁶ a °C round zero): the linear term is the spring maker's to set, as Table IV's screw moves then trim it
  a split bimetallic rim (brass 19 on steel 11.5, 1.6 thick, r 14.5, the model's variant): the free ends come in 1.48 µm a °C, gaining 7.0 s a day a °C
  ok  against the 10.3 s a day a °C a steel spring loses: 68 % with the screws at 90° on average; moving them toward the free ends (or the rim's thickness) brings it to 100 % (Table IV's job)
  ok  the essay quotes swingM as 316: found 316
  ok  the essay quotes stage as 0.91-0.97: found 0.91-0.97
  ok  the essay quotes Qparts as 405-710: found 405-710
  ok  the essay quotes TFparts as 64-113: found 64-113
  ok  the essay quotes Mup as 1.00: found 1.00
  ok  the essay quotes sUp as 2.85: found 2.85
  ok  the essay quotes esc as 67-104: found 67-104
  ok  the essay quotes chase as 18: found 18
  ok  the essay quotes TFm as 67: found 67
  ok  the essay quotes Q as 207-378: found 207-378
  ok  the essay quotes steelRate as 10.3: found 10.3
  ok  the essay quotes comp as 7.0: found 7.0
  ok  the essay quotes eNeed as -22 to 10: found -22 to 10

tools/strength.js passes

Every loaded part fully wound against its material's range (Lewis' formula for teeth, beams for pivots and pins), and the knocks it takes.

its output today
The going train, fully wound (each stage lossless: the largest forces)
  ok  fusee wheel → centre pinion (90:14, m 0.392)   22.0 N, 388.1 mN·m   126 MPa   brass's yield (41k cycles in 30 y) 300-450: margin 2.4-3.6
  ok  centre wheel → third pinion (90:12, m 0.3245)  4.1 N, 60.4 mN·m      68 MPa   brass's yield (263k cycles in 30 y) 300-450: margin 4.4-6.6
  ok  third wheel → fourth pinion (80:10, m 0.251)   0.802 N, 8.05 mN·m    25 MPa   brass's endurance (2.0M cycles in 30 y) 120-170: margin 4.9-6.9
  ok  fourth wheel → escape pinion (75:10, m 0.2457) 0.109 N, 1.01 mN·m     2 MPa   brass's endurance (15.8M cycles in 30 y) 120-170: margin 48.6-68.8
  the pinions are hardened steel, 2.7-6.0 times as strong as the brass teeth that drive them: the wheels' teeth are the weaker side of every mesh
  ok  escape wheel tooth, struck (the tooth alone springing) 15.2 N peak, 2.0 µJ    77 MPa   steel's yield (118M strikes a tooth in 30 y; below the endurance limit, about half the yield, it lasts) 1200-1800: margin 15.6-23.4
  Pivots and bearings
  ok  fusee arbor's pivot (r 1.85, 3.4 long in its bushing) 73.7 N (chain + wheel)    25 MPa   steel's yield 1200-1800: margin 47.6-71.5
  ok    its bushing, the pressure (oiled brass, turning once in 6.4 h) 73.7 N                 6 MPa   oiled brass at a crawl, typical 10-35: margin 1.7-6.0
  ok  centre arbor's thinner pivot (r 0.5)           26.1 N               133 MPa   steel's yield 1200-1800: margin 9.0-13.5
  ok  third arbor's thinner pivot (r 0.3)            4.9 N                 70 MPa   steel's yield 1200-1800: margin 17.2-25.8
  ok  fourth arbor's thinner pivot (r 0.25)          0.911 N               22 MPa   steel's yield 1200-1800: margin 53.9-80.8
  ok  escape arbor's thinner pivot (r 0.089)         0.109 N               59 MPa   steel's yield 1200-1800: margin 20.3-30.4
  The fusee chain, fully wound
  pull 51.7 N (the spring's 1.00 N·m on the barrel's chain radius 19.32 mm, physics.js); plates 1.95 × 0.18 mm, rivets r 0.17, pitch 2.19
  ok  an inner link: its one plate across the rivet hole 51.7 N               178 MPa   steel's strength 1400-2000: margin 7.9-11.2
  ok    the same plate at its waist                  51.7 N               449 MPa   steel's strength 1400-2000: margin 3.1-4.5
  ok  a rivet, in double shear                       51.7 N               285 MPa   steel's shear strength 840-1200: margin 3.0-4.2
  ok  the rivet bearing on the inner plate           51.7 N               845 MPa   steel in bearing (1.5 × strength) 2100-3000: margin 2.5-3.6
  ok  the fusee end's crook across its throat (1.21 × 0.18), in tension 51.7 N               237 MPa   steel's strength 1400-2000: margin 5.9-8.4
  ok    its nose at the root (1.22 × 0.18), in shear 51.7 N               235 MPa   steel's shear strength 840-1200: margin 3.6-5.1
  ~~    the nose bearing in its hole in the fusee (0.18 × 0.84 deep) 51.7 N               344 MPa   brass in bearing (1.5 × yield) 450-675: margin 1.3-2.0
  ok  the hook's nose (r 0.28, 0.3 into the wall), in shear 51.7 N               210 MPa   steel's shear strength 840-1200: margin 4.0-5.7
  ~~    bearing on the barrel's brass wall           51.7 N               308 MPa   brass in bearing (1.5 × yield) 450-675: margin 1.5-2.2
  The springs
  ~~  mainspring, fully wound (physics.js)           1.00 N·m            2848 MPa   a set carbon-steel strip's limit, 1.5 times its elastic limit (physics.js: a rectangle's plastic moment, the strip formed past its elastic limit) 2550-2850: margin 0.9-1.0
    as a mainspring is meant to, it works near its limit: wound once a day (10.9k cycles in 30 years) it fails by taking a set (the model's MSET, Op. 13), not by fatigue
  ok  hairspring at 255° (0.221 × 0.23 mm, 409 mm: hairspring.js) 407 µN·m             217 MPa   Elinvar-type's endurance (1.9G cycles) 350-500: margin 1.6-2.3
  ok    carried past a full turn to trip (333°)      530 µN·m             283 MPa   Elinvar-type's yield, once 1100-1400: margin 3.9-5.0
  ok  detent spring, its swing (preloaded 56, lifted 89 MPa) 87 µN·m               16 MPa   beryllium copper's endurance at that mean (Goodman, 1.9G cycles) 225-294: margin 13.9-18.2
  ok  trip spring, bent 0.51 mm at its tip           3.0 mN               227 MPa   Elinvar-type's endurance, its clamped root taken at 1.5 × (estimated) 350-500: margin 1.5-2.2
  Knocks (the chronometer in its gimbals, the staff upright)
  the balance staff's pivots (r 0.18, the load 0.35 mm out): a sideways knock bends them at 670-1403 g (the balance 3.4-4.8 g, the steel's yield)
  end-on, the flat pivot's end on its endstone takes 4342-6513 g before the ruby's 2-3 GPa (the balance's heavier mass)
  the escape arbor's pivots (r 0.089, its mass 178 mg from its inertia at a 5 mm radius of gyration, estimated): 2532-3798 g
  the detent, against its preload, its spring and the tooth's draw, frees the wheel under a steady 35 g across the plate (a short blow needs more: dynamics.js): the first thing a knock upsets, and harmlessly (a tooth escapes, the hands jump half a second)
  so a knock that would bend a balance pivot (some 670 g) has long since let teeth go; the gimbals and the box's padding are what keep it from either

tools/dynamics.js --quick passes

The escapement integrated in time through every contact, against the averaged model; the detent's return.

its output today
The bodies
  the balance 578.5 g·mm² on k 91.4 µN·m a radian, damped as its free run-down TF 67 s (Q 421)
  the train at the escape wheel 5.66 g·mm² (its arbors' solids, mechanisms.json: the fourth wheel's reflected 0.81); driven at 61.8 µN·m, the torque the model is calibrated at
  ok  the model's bodies and springs (JW 5.65, the detent's inertia ID 2.51, its spring's 1.512 mN·m a radian and 55.5 µN·m, the trip spring's) are these, from the drawn parts
  the detent: 45.4 mg, 2.510 g·mm² about its flexure, its centre of mass 6.76 mm from it (the drawn pieces, the trip spring with it)
  its spring: two strips 0.50 + 0.45 mm tall, 0.079 thick, 3.30 long: 1.512 mN·m a radian (beryllium copper, 128 GPa), preloaded 55.5 µN·m by Op. 78 (as if bent 2.10° past its stop)
  the wheel is free at 0.881° of the detent (lRel 0.205 mm at the passing spring's tip); the trip spring 6.4 × 0.046 × 0.37 mm, 5.9 N/m at its tip (Elinvar-type, 180 GPa)
  ok  the contacts are makeEsc's, worked out apart: the detent's lift over the whole active swing (693 points, to 0.00 µm) and the trip spring's bending (1203 points, to 0.00 µm)
  ok  the integrator: kicked at the dead point the balance keeps −0.063 s a day, the damping's own (−ζ²/2: −0.061; at half the steps −0.062)
  (1 s)
  1. The model, beat by beat (its springs, the detent all but massless and dead on its stop, as makeEsc has it)
  makeEsc: the swing 255.0°, the tooth lands on the impulse jewel 18.5° after the release, the detent falls back at -13.5°, the escapement's own rate −2.91 s a day
  in time:  the swing 255.3°, the tooth lands 18.4° after the release, the detent falls back at -13.5°, the escapement's own rate −2.85 s a day (at half the step −2.85)
  ok  it runs, beat by beat: 140 beats, no tooth lost or gained
  ok  the swing within 3° of the model's: +0.3° (the wheel's chase from rest, its strike on the jewel and the springs as the model counts them)
  ok  the tooth lands where the model has it, within 1.5° of the balance
  ok  the detent falls back where the model has it, within 1°
  ok  the escapement's own rate within 0.3 s a day of Airy's sum: +0.06 s a day apart
  the drive 95 % → 111 %: the model 248° → 270°, the rate +0.30 s a day; in time 248° → 270°, +0.29 s a day (+0.130 s a day for each 10° of swing)
  ok  isochronism (the rate's change over the manual's range of swing) within 0.1 s a day of the model's (−0.00)
  ok  and small (under 0.5 s a day for each 10° of swing): what the fusee keeps even, the escapement doesn't undo
  (6 s)
  2. The detent as a body: its inertia, bouncing on its stop
  the swing 253.6°, the escapement's own rate −3.05 s a day; the tooth lands 18.2° after the release
  ok  the detent's own inertia and bounce change little: the swing −1.4°, the rate −0.14 s a day against the model
  the detent leaves the discharge jewel for the last time at -13.97° (first at its acting corner, falling back onto its end) and first lands on its stop at -10.70° (makeEsc's fall: -13.96° and -10.70°; home for good at -7.09°)
  ok  the detent's fall back to its stop in the model is the integration's, within 0.5° of the balance (it went home in one step at -13.5° until 11 October 2026)
  ok  no tooth gets past the stone: 140 beats; the detent is back on its stop 10.30 ms before the next tooth reaches it, still 0.0 % lifted as it arrives
    spring half as stiff                   254°, trips 0, back 10.26 ms before the tooth, 0 % lifted as it arrives
    spring twice as stiff                  255°, trips 0, back 10.26 ms before the tooth, 0 % lifted as it arrives
    spring 0.06 mm thick (0.44× as stiff)  254°, trips 0, back 10.26 ms before the tooth, 0 % lifted as it arrives
    preload half                           256°, trips 0, back 9.95 ms before the tooth, 0 % lifted as it arrives
    preload 1.5×                           252°, trips 0, back 10.45 ms before the tooth, 0 % lifted as it arrives
    detent twice its inertia               254°, trips 0, back 9.85 ms before the tooth, 0 % lifted as it arrives
    bouncy stop (restitution 0.7)          254°, trips 0, back 10.26 ms before the tooth, 0 % lifted as it arrives
    dead stop (restitution 0)              254°, trips 0, back 10.26 ms before the tooth, 0 % lifted as it arrives
    the train 1.5× its inertia             246°, trips 0, back 10.51 ms before the tooth, 0 % lifted as it arrives
  ok  over all of these the detent is home before the next tooth: what brings it back is the preload Op. 78 sets, not the spring's thickness
  with no preload at all the spring's own stiffness still brings the detent home in time: Op. 78's preload is for the locking's safety against knocks (below), not for the return
  (11 s)

tools/errorbudget.js passes

A Monte Carlo of the escapement made to a maker's tolerances, and a 60-day voyage's error from the test card's figures, the barometer, the oil and the sea.

its output today
1. As made: the escapement with a maker's tolerances (Monte Carlo, the reference torque and balance)
  1000 escapements: 75.8 % run; their swing 242-261° (5-95 %), 62 % inside the manual's 247.5-270°; 22 % meet every adjustment figure as made
  out of the manual's figures, as made: overall (Op. 87) 49 %, would not run 24 %, drop (Op. 97) 15 %, let-off (Op. 86) 8 %, lock (Op. 85) 0 %: the adjuster's work (the lock-adjusting and detent-adjusting screws, Ops. 84-97, set lock and the detent's place; the trip spring's tip is bent to its figure)
  the escapement's isochronism over them (averaged model): +0.07 to +0.18 s a day for each 10° of swing (5-95 %); the model's own +0.13
  (20 s)
  2. At sea: 60 days from a port in winter (55 °F) into the tropics (85 °F), rated over 10 days in port, the error the navigator carries
  the escapement's isochronism in time, near 255°: +0.142 s a day for each 10° of swing (dynamics.js; 247° → 269°)
  ok  small, as the fusee and the detent escapement are meant to make it
  the barometer: the air takes 50 % of the balance's damping, so 1 % denser air swings it 0.64° less: −0.0091 s a day
  heavy weather (yaw ±10°, ±0.3 g sideways, 8 s): +0.010 s a day (dynamics.js)
  like No. 3390: scatter 0.042 s a day, drift ±2.85 ms a day each day (calibrated on its rating 0.03 and recovery 0.06); the 30-day test then reads (median) rating 0.03, recovery 0.06, temperature 0.08 / 0.06 / 0.02 (the card 0.08 / 0.06 / 0.02); it passes 100 % of the time
  ok  the stochastic model reproduces the card it is calibrated on
  a borderline chronometer (the card's scatter, drift and linear temperature term 5.1 times as large): scatter 0.216 s a day, drift ±14.63 ms a day each day, 0.41 s a day over 17½ °F; it passes the Navy's test 49 % of the time
  the error the navigator carries (median, and 95 % of 2000 voyages; in miles of longitude on the equator, 4 s of time a mile):
  like No. 3390                      10 days: 0.4 s (0.1 mi), 95 % 0.9 s (0.2 mi); 30 days: 1.4 s (0.3 mi), 95 % 3.9 s (1.0 mi); 60 days: 4.3 s (1.1 mi), 95 % 12.3 s (3.1 mi)
  borderline (passes half the time)  10 days: 2.0 s (0.5 mi), 95 % 4.9 s (1.2 mi); 30 days: 15.1 s (3.8 mi), 95 % 29.7 s (7.4 mi); 60 days: 37.9 s (9.5 mi), 95 % 87.4 s (21.9 mi)
  what each source adds at 30 days, like No. 3390 (median |error|, alone):
    the temperature curve (55 → 85 °F)     0.96 s
    day-to-day scatter                     0.31 s
    drift                                  1.15 s
    the oil thickening                     0.16 s
    the barometer                          0.14 s
    heavy weather one day in five          0.06 s
  ok  a chronometer like No. 3390 carries 1.4 s (0.3 miles of longitude) after a month at sea, well under 30 s (7½ miles): it keeps time
  (23 s)

What only the bench can settle

Reproduce

From marine-chronometer-source/chronometer-working-model/tools/ (Node.js 22, Python 3.13 with numpy, scipy, pyamg and Playwright's Chromium; times on the author's machine):

node gearing.js            # about 2 min (--quick 15 s; --write rewrites gearing.json)
node thermal.js            # under a second
python fea.py              # about 3 min (--check 5 s; --reuse skips the browser)
node hairspring-fe.js      # about 20 s
node bench.js              # a second (--write, then python makers.py)
python rating.py --demo    # under a second; python rating.py LOG.csv for a real log
node tripping.js           # about 6 min (--quick 10 s: the frame, the validation, the bench's pitches)
python massprops.py        # about 30 s (the browser): the case and ring weighed; --write writes gimbals.json
node gimbals.js            # about 2½ min: the gimbals in a seaway
python mainspring_friction.py  # about a minute (14 processes): the coils' friction (mainspring_friction.json, read by physics.js and bench.js)
python mainspring.py --bound   # about 6 min (14 processes): where the model's coils pack solid, to 0.05 turn (mainspring_bound.json)
python mainspring.py --nivaflex   # about 20 min (14 processes): the build's strip solved (mainspring_t0.386.js); then --nivaflex --bound
python mainspring_friction.py --nivaflex   # about a minute: its friction (mainspring_friction_t0.386.json); then node ../verification/build-verification/reportdata.js
MC_MS_TAG=pitch MC_MS_T=0.391 MC_MS_GAP=0.003 python mainspring.py   # about 20 min, then --bound: a strip in the sources' proportions (also klu, klu52, tn24h, bm: their parameters in mainspring_bound.json)
python mainspring_friction.py --variants   # about 6 min: every strip mainspring_bound.json lists (mainspring_friction_NAME.json); --check in ci.py --full
node tripping.js --write   # about 2 min: tripping.json, the bench's lock floor and the detent's ring (then node bench.js --write, python makers.py)

and this report, from verification/build-verification/:

node reportdata.js                                                 # about 35 s: assets/data.json
python ../../tools/fea.py --reuse --json assets/fea.json           # about 3 min
python ../../tools/rating.py --demo --json assets/rating.json      # under a second
node ../../tools/tripping.js --json assets/trip.json               # about 6 min
node ../../tools/gimbals.js --json assets/gimbals.json             # about 2½ min
node mechdata.js                                                  # about 45 s: assets/mech.json (section 11)
python build_report.py                                             # index.html (runs the earlier checks for their output, about 30 s)

python ci.py at the repository's root runs the quick ones on every change, --full adds gearing.js in full and fea.py.

Scripts

ScriptWhat it doesFails when
tools/gearing.jsThe four meshes tooth by tooth; the train in time; the depthing windows and the hole tolerance (gearing.json)a mesh jams or the swing leaves 247.5–270° at the drawn centres; gearing.json stale
tools/thermal.js40–100 °F: the depths, endshake, the escapement's figures, the chaina depth uses a tenth of its window, endshake leaves the manual's, a figure leaves the manual's
tools/fea.pyVoxel finite elements on the six pivot-holding parts, superposed; the cantilever checka depth moves a tenth of its close side; stress past two thirds of brass's yield; the check off by 3 %
tools/hairspring-fe.jsThe hairspring by co-rotational beam elements against hairspring.js's elasticathe couple's shape, the force on the pivots or the rate part from the elastica's
tools/bench.jsThe bench tests and their figures (bench.json, through makers.py into the cards and the build book; book.py checks them printed)bench.json stale
tools/benchcheck.pyA maker's bench measurements judged against bench.json's figures, with what to do(--demo) a planted fault missed; (a real file) a measurement fails
tools/rating.pyA made chronometer's log: the Navy test's figures, the fit, the backtest, the projection(--demo) the fit misses what the log was made with
tools/tripping.jsThe detent as a frame: its modes; the frame made rigid against dynamics.js; running as drawn, the estimates varied, the swing pushed, the knocks that trip itthe rigid frame parts from dynamics.js; a tooth gets past at the model's settings or under any estimate varied; a held tooth's stone uses half the depth of lock; a blow under 20 g trips it; tripping.json is stale
tools/gimbals.js, tools/massprops.pyThe case and ring weighed on the model (gimbals.json); their swing at sea, the resonance curve, the stops, what reaches the escapementthe gimbals' periods among a ship's or the beat; a working sea tilts the case 2° from the apparent vertical, reaches its stops or moves the rate 0.05 s a day; what reaches the escapement comes within half of what trips it; gimbals.json stale
tools/mainspring_friction.pyThe mainspring's coil friction over the wind, the fusee's output running, the set-up against it (mainspring_friction.json); --nivaflex the same for the build's strip, set up by the fusee's rule (mainspring_friction_t0.386.json)(--check) the file is stale; invariants.py holds its fusee to the page; the rule must reproduce the page's full wind
tools/mainspring.pyThe mainspring's shape over the wind (js/mainspring.js, the page's); --bound where the coils pack solid, to 0.05 turn; --nivaflex the build's strip, beside it (mainspring_t0.386.js); MC_MS_T, _L, _RA, _GAP, _TAG any strip, length, core and film (mainspring_NAME.js)(writes) states that can't keep the coils apart and the length inside the range
tools/acoustic.pyThe swing from a recording of the beat (now any recording's path, a made movement's)—
tools/mechanisms.pyThe mechanisms off the built model: each pawl's contact on its ratchet, the pawls' and arbors' inertias (mechanisms.json)about 40 s
tools/overwind.jsThe key forced: against the winding stop and the wrong way into the train (11.1)under a second
tools/ratchets.jsThe ratchets and pawls: draw, strength, springs, the let-go (11.2)under a second
tools/sustaining.jsThe sustaining spring: its stiffness by Castigliano, the train run on it while the key is held (11.3)about 15 s
tools/motionwork.jsThe motion work's and wind indicator's meshes, the hands' play, the cannon pinion's friction (11.4)about 7 s
tools/wear.jsWear at every sliding contact by Archard's law, against each tolerance (11.5)a few seconds
tools/traindyn.jsThe train in time: its inertia at the escape wheel, its modes, a beat on four sprung meshes with backlash (11.6)a few seconds
tools/montecarlo.jsThe whole movement to a maker's tolerances at once (11.7)about 30 s
reportdata.js, build_report.pyThis report's data and page—

Sources

  1. Bureau of Ships, Overhaul, Repair and Handling of Hamilton Model 21 Chronometer, NAVSHIPS 250-624 (1948): Secs. II, IX (the Navy test and its limits), XI (the parts list), Ops. 13, 15, 69, 74, 78, 84–97. In the repository, References/; served beside the page as navships-250-624-1948.pdf.
  2. British Standards Institution, BS 978 Part 2, Specification for fine pitch gears: cycloidal type gears (the tooth proportions the model draws; estimated, not from the manual).
  3. NORDFORSK, Assessment of Ship Performance in a Seaway, Marintek (1987): the operability criteria, lateral acceleration 0.12 g rms at the bridge (as applied in later seakeeping studies, e.g. Wulandari et al., 2026).
  4. MARIN, On empirical methods to predict the rolling period of ships (IMDC 2024), and Jachowski et al., KONES (2012): merchant ships' roll periods, 14-32 s in the latter's cases.
  5. Mainspring torque, winding against unwinding: "the torque during unwinding is 10 to 20 per cent less than that during winding" (ICMFF9, gruppofrattura.it).
  6. Friction of lubricated steel on steel, 0.08-0.2: the CRC Handbook of Chemistry and Physics, 43rd ed., p. 2181, as cited in US 3,939,081; Misumi's table (grease 0.16 static, oil 0.08 kinetic).
  7. J. S. Przemieniecki, Theory of Matrix Structural Analysis, McGraw-Hill (1968), chs. 5 and 11: the plane frame element and its consistent mass.
  8. P. Wriggers, Computational Contact Mechanics, 2nd ed., Springer (2006), ch. 2: penalty contact.
  9. L. L. Howell, Compliant Mechanisms, Wiley (2001), ch. 5: the small-length flexural pivot, which turns about its middle.
  10. M. A. Crisfield, Non-linear Finite Element Analysis of Solids and Structures, vol. 1, Wiley (1991), ch. 7: the co-rotational beam element.
  11. S. P. Timoshenko and J. M. Gere, Mechanics of Materials: the cantilever's deflection with shear (the solver's check).
  12. E. Phillips, "Mémoire sur le spiral réglant des chronomètres et des montres", Annales des Mines (1861): the terminal curves' conditions.
  13. L. N. Olson and J. B. Schroder, PyAMG: Algebraic Multigrid Solvers in Python (2018), github.com/pyamg/pyamg: the multigrid preconditioner.
  14. Material ranges (published handbook values, as the tools state them): leaded brass CuZn39Pb3, E 97–110 GPa, expansion 18.7–20.9 × 10⁻⁶/°C, yield 300–450 MPa; hardened carbon steel, expansion 10.8–12.5 × 10⁻⁶/°C; tooth friction 0.10–0.20 (physics.js).
  15. Jacques Allemann, Nivaflex 45/5 data sheet JAS-014 (2021): composition, E 220 GPa, density 8.50 g/cm³, expansion 11.2 × 10⁻⁶/K, tensile strength over 3000 N/mm², non-magnetic, resistant to seawater: jacques-allemann.com.
  16. Vacuumschmelze, Nivaflex 45/18 brochure: tensile strength up to 2900 MPa, mainsprings among its uses: vacuumschmelze.com.
  17. Nivaflex (patented 1948, the name registered 1957): Wikipedia; Grail-Watch wiki, Reinhard Straumann.
  18. Nivarox (perfected 1933, patented 1935; Fe-Ni-Cr-Be, its thermoelastic coefficient set by alloy and heat treatment, graded for monometallic balances such as Glucydur, beryllium bronze): WatchesbySJX, Hairspring Materials and Evolution Part II; VintageWatchstraps, Nivarox balance springs; Wikipedia, Glucydur; Grail-Watch wiki, Glucydur.
  19. Finger torque on small handles: S. N. Imrhan and M. J. Munoz-Vasa, "Finger Pinch Loosening and Tightening Torques on Small Cylindrical Handles Among Industrial Workers", Industrial and Systems Engineering Review 1(2), 2013 (and Swain et al. 1970, cited there).
  20. Breguet's "tipsy" key (clé de sûreté), about 1790: a ratchet in the key that slips when it is turned the wrong way.
  21. J. F. Archard, "Contact and Rubbing of Flat Surfaces", Journal of Applied Physics 24, 981 (1953); E. Rabinowicz, Friction and Wear of Materials, 2nd ed. (Wiley, 1995): wear coefficients for lubricated metals.
  22. A spring alloy's thermoelastic coefficient within ±10 ppm/°C gives an oscillator within ±0.6 s a day a °C: EP3502288A1 (Nivarox-FAR), Method for manufacturing a hairspring.
  23. The thin ring's bending under radial loads by Fourier series: S. Timoshenko, Strength of Materials, Part II; a ring under two opposite forces, its diameters' changes 0.1366 and 0.1488 P R³ / E I: Roark's Formulas for Stress and Strain. The balance and Table IV (p. 74), Op. 8: the manual.
  24. Xometry, Beryllium bronze C17200 data sheet: density 8.25 g/cm³, expansion 17 × 10⁻⁶/°C: xometry.asia.
  25. The working model's own README (model of the escapement, the balance's dynamics, "Checked in time", "Estimated, not from the manual") and the report Does the model behave like a real Hamilton Model 21? (the model against recordings of real Model 21s).

The scripts' outputs that feed this page are embedded in it, so it opens from the file. Built by build_report.py; do not edit index.html by hand.