Goldray

A metal detector hears one note and forgets it. Goldray is a design for keeping every note, with the place the coil was when it sounded, until a swing becomes a picture of the ground.

Contents
  1. tl;dr
  2. One afternoon at Kingower
  3. Seven tones into the ground
  4. Fewer wires, fewer grams
  5. Ground the tones cannot handle
  6. One a stranger could buy
  7. Where it stands
  8. Method and sources

Can a phone, a Raspberry Pi and an audio chip turn the swing of a metal detector into a picture of what lies underneath?

Goldray is a metal-detecting instrument designed in five blocks, from a headphone lead into a phone to a module on the shaft that sends seven tones into the ground and images what answers. In simulation it finds a 1 g nugget 21 cm down in mild soil. None of it has been built.

BLOCK I
AUD 10–30
A lead from any detector’s headphone jack to a phone. Rough estimate
Depth, 1 g nugget
21 cm
Predicted, one pass, mild soil; 5 cm in moderate soil
Position to name it
2 mm
Over two passes 90° apart; a phone gives about 10 mm
Units built
0
Nothing wound, soldered or swept

tl;dr

Goldray is designed to keep what a metal detector throws away: every measurement, with the place the coil was when it was taken, so that a swing adds up to an image. It comes in five blocks, none of them built, and every depth below is a simulation’s.

The first block is a headphone lead into a phone, for AUD 10 to 30, and it turns an afternoon of beeps into a map. The second, the Core, is the real instrument: a small module on the shaft that drives the coil with seven tones at once, about 94 times a second, and stamps each reply with the coil’s position. A swing of those replies behaves like one large antenna, and it can give a buried object a depth and a signature that tells a nugget from a nail.

In simulation the Core finds a 1 g gold nugget 21 cm down in mild soil, and only 5 cm down in moderate soil, because 5 mm of error in the coil’s height lets the ground drown it. Saying what an object is needs the coil’s position to about 2 mm, over two passes at right angles. A phone on its own manages about 10.

one swingthe imageground
Figure 1Five positions along one swing sight the same buried object. Not to scale.

Five blocks, each a whole instrument

IPhone and leadAUD 10–30IIThe CoreAUD 450–700IIICarrierAUD 300–450IVFront endAUD 400–800VIntegratedAUD 150–300rough cost per unit, unverified · dotted: software exists · dashed: outline only, waiting on the block below
BLOCK I · Phone and leadSoftware written. Never run on a phone.
Hardware
A cable or USB-C audio dongle from any detector’s headphone jack to a phone on the shaft. Optionally the same audio into the Core’s software on a laptop or a Pi, for up to about AUD 150 in all.
What it does
The Goldray app takes the tone’s loudness and pitch, tags each reading with the phone’s position, draws the map and logs digs.
What it proves
Whether a phone knows its position well enough on real ground, and the map and dig log. Nothing about the physics.
Starts when
None. It needs a detector and a phone, and no soldering.
Costs are rough estimates. No block has been built.

BLOCK I · A PHONE AND A LEAD

One afternoon at Kingower

On an imagined afternoon, a detectorist parks outside Kingower with a borrowed detector, a phone and a headphone lead. Under the old diggings, say, a 1 g nugget lies 15 cm down among a century and a half of nails, tins and lead shot. By evening there would be thirty holes and no record of which beep led to which.

Kingower sits in Victoria’s Golden Triangle, the old goldfields between Bendigo and Ballarat. On 26 September 1980 a man with a metal detector dug the Hand of Faith out of the ground near the town: 27.2 kg of gold, about 30 cm down by the usual account, and still the largest nugget found with a detector. In 2023 a 4.6 kg rock holding about 2.6 kg of gold came out of the same country under a budget detector.

No detector needs help with 27 kg a foot down. The hard targets are the small ones lying among the junk, and each block below is measured against that one gram at 15 cm.

0102030cmreach, moderate soilreach, mild soiltopsoilironstone gravelclayquartz1 g, 15 cmillustrative
Figure 2What an afternoon’s ground might hold. Illustrative, not surveyed.

A camera with one pixel

A metal detector is a camera with a single pixel that forgets each frame as soon as it has shown it. What it measures is one complex number: how much whatever lies under the coil changes the coupling between a transmit winding and a receive winding, split into a part in step with the drive and a part a quarter-cycle behind it.

The detector squeezes that number into a tone, throws it away within a few tens of milliseconds, and never knew where the coil was when it took it. Depth, size and shape are rebuilt, if at all, in the operator’s head, from how the tone changes as the coil is wiggled. Goldray sets out to keep what the detector discards: the whole number, its history and its place.

heardeach frame forgottenkept, with where the coil was
Figure 3Each frame heard and forgotten, or kept with the place the coil was.

A lead from the headphone jack

The first block keeps only the place. A lead runs from the detector’s headphone socket to a phone clamped on the shaft, and the detector works as it always has. The Goldray app listens to the tone, takes its loudness and pitch as the signal, and tags each moment with the phone’s own estimate of where it is, worked out from its camera and motion sensors. By the end of the afternoon at Kingower there would be a map of every signal and a note against each hole: nail, shot, nothing, gold.

A detector’s audio holds no physics worth extracting, so this block can say nothing about depth or identity. What it can test, on real ground and for the price of a cable, is the question every later block rests on: whether a phone on a swinging shaft knows its own position well enough.

AUD 10–30 of cableany detectorpose+ tone
Figure 4The detector unchanged; the phone remembers where each beep happened.

One optional step stays inside the block. The same audio can feed the Core’s software running on a laptop, or on a Pi with a power bank, and nothing new reaches the screen. It is a rehearsal on a real clock. A simulator keeps perfect time and its hardware answers instantly, and software that has only ever met a simulator can hide timing faults that real hardware exposes.

A map of beeps says where something is, but not how deep it lies or what it is. That takes BLOCK II, the Core.

BLOCK II · THE CORE

Seven tones into the ground

Back on the patch, the map shows two beeps a pace apart and equally loud. One is the nugget and one a rusted nail, and the map cannot say which. The Core takes the detector out of the loop: a Raspberry Pi Zero, an audio chip and a hand-wound coil ride the shaft and send seven tones of their own into the ground.

Its predicted reach for a 1 g nugget is 21 cm in mild soil and 5 cm in moderate soil. The gram at 15 cm sits between the two, and which side of it the afternoon falls on depends on the ground.

The method comes from clearing unexploded ordnance. There, carts carrying arrays of coils and survey-grade satellite positioning work across the ground and sort buried shells from scrap by how each one answers a magnetic field. Goldray tries the same physics with one coil, a phone for positioning and an arm in place of the cart.

A swing as an array

Radar showed long ago that many small looks from known positions add up to one large antenna, and Goldray applies the trick to a search coil. Each record, which the project calls a ray, holds the response at all seven tones, the time it was taken, and the coil’s position and tilt. A survey is a cloud of rays, and nothing in it is thrown away. For a small buried object, each ray is set by three things.

Y(ω,p)=hR(r)TM(ω)hT(r)

Here r is where the object sits relative to the coil, h_T and h_R are the field patterns of the two windings, known from the coil’s shape, and M(ω) is a 3 × 3 array of numbers called the magnetic polarizability tensor. M belongs to the object alone. Its three principal values describe size and shape, and the way they change with frequency says how well the object conducts and whether it is magnetic.

Where the coil is listening

windings, top viewTXRX5 cm below · peak 2.8410 cm below · peak 0.6820 cm below · peak 0.11
Figure 5Sensitivity on slices 5, 10 and 20 cm under an idealised double-D, each scaled to its own peak.

One reading cannot tell a small, shallow object from a large, deep one, because both can give the same number. A cloud of readings can: the bump an object makes as the coil passes over it widens with depth in a way the geometry fixes.

So the Core lays columns 5 cm across under the swept ground, cuts each into eight layers from 2 to 40 cm down, and asks of every cell how well a target there would explain the rays so far. The image sharpens while the operator swings, and each new pass adds to the last.

5 cm tiles8 layers · 2 to 40 cmpeak
Figure 6Tiles 5 cm across, with eight depth layers under each. Not to scale.

A swing becomes an image

-40-2002040-40-2002040coil track, plan view, cm-40-2002040back-projected SNR at 11.1 cm, 3 of 3 passespeak 1.1 cm from truthpeak SNR 1,892-40-2002040one line (dashed) · |Y| at 10.22 kHz00.51235711172640depth layer, cmSNR down the peak column, of max
Figure 7A 2 cm ball of gold-like metal 10 cm down, in air. The buttons stop the image after one, two and three passes.

On a reference scene, a 2 cm ball of gold-like metal 10 cm down, three passes put the brightest cell 1.1 cm from the ball. Depth is the weak axis. Forty centimetres down the same column the image still holds 88 % of its peak, because a bigger, deeper object explains the same bumps almost as well. Pinning the depth, and naming the object, takes a second step: a fit that moves a trial object about until one tensor per tone explains every ray.

Seven primes

A single-frequency detector asks a buried object one question. The Core asks seven at once, from 1.59 to 72.09 kHz, about 94 times a second. The tones have to sound together and stay locked to one clock, and a 24-bit audio chip running at 192 kHz does that by construction, because its output and inputs share a clock. One input watches the current going into the transmit winding and the other the voltage coming off the receive winding, so every measurement is a ratio and the chip’s own drift cancels out.

Seven tones, five gaps

1101001k2k5k10k20k50k100kfrequency, Hz (log)TX current per tone, mA peak (log)19.95–20.05 kHz gaptime-signal spotscodec −3 dB?173159109211397769spec test vector (17 … 811)
Figure 8Stem height is drive current. Hatched bands are frequencies the rules exclude; teal ticks mark the tones first proposed.

Each tone has to land exactly on one of the frequencies the Core can separate cleanly, 93.75 Hz apart, and stay off the harmonics of 50 Hz and 60 Hz mains. None may sit where two or three of the others would create a false tone in imperfect electronics, and none in a band the Australian class licence closes. A search for seven that meet every rule returns slots 17, 31, 59, 109, 211, 397 and 769, and every one of them is prime.

The licence closes the band from 19.95 to 20.05 kHz outright, and the total drive is held to 0.1 A rms, about half the public exposure guideline for magnetic fields, which gives away roughly 6 dB.

A nugget and a nail

Each tone comes back in two parts, one in step with the drive and one a quarter-cycle behind it. Eddy currents in a good conductor take time to build and decay, so its lagging part peaks at one frequency, roughly where a cycle matches the time the currents take to die away. Iron does the opposite: it draws the field in, which shows as an in-step part of the other sign, largest at low frequency.

A sphere is the one shape whose response has an exact formula, which makes ball bearings Goldray’s calibration kit. Real nuggets are lumpy alloys and will sit somewhere off the sphere’s curve.

drivein phasequadraturecoilobjectdelay
Figure 9Each tone returns two numbers: the part in step with the drive and the part behind it.

Signatures in frequency

-10121001k10k100k1MHzin-phase, β′ / 2πa³the seven tones00.20.40.61001k10k100k1MHzquadrature, −β″ / 2πa³gold 0.1 ggold 1 gcopper ball 10 mmsteel ball 10 mm
Figure 10Closed-form sphere responses at textbook conductivities. Real nuggets are irregular alloys and will sit off these curves.

Size sets where the peak falls; it moves down with the square of the radius. A 1 g gold sphere, 2.3 mm in radius, peaks near 6.8 kHz. A 0.1 g sphere, 1.1 mm in radius, peaks near 32 kHz, and at 72 kHz it has still reached only about 60 % of its full in-step response. The smallest gold speaks only at the top of the band, so losing the top tone would cost far more than a seventh of the data.

Shape does the rest: the tensor has one principal value per axis, and common finds differ in how many of them matter.

Nugget
Compact: three similar values.
Nail
One dominant, magnetic axis, in step with the drive at low frequency.
Ring-pull, foil
A shorted loop: one value, along its normal, with a single peak.
Hot rock
Like ferrite: in step, positive and nearly flat.

A coil that cannot hear itself

The transmit winding is many orders of magnitude louder than any nugget, and the receive winding sits right beside it. So the coil is two D-shaped windings laid with their flat sides overlapping. At one overlap the transmit field passing through the receive winding cancels exactly; for two D’s of 130 mm radius, sized for a 280 mm printed former, that point falls at 17.90 mm.

R 130 mm260 mm17.90 mmoverlap±3 mm trimTX · 40 turnsRX · 120 turns
Figure 11Former in plan. Two D windings of 130 mm radius, flat sides overlapping by 17.90 mm. Modelled, not printed.
pod: preamp on the coilcoverRX · 120 turns · slides ±3 mmTX · 40 turns · fixed
Figure 12Former, taken apart. Tray, sliding half, cover and the pod that carries the preamplifier.

The null is sharp enough that holding the leakage 40 dB down means setting the overlap to about a tenth of a millimetre before the coil is potted: 40 turns go on one D and 120 on the other, then one half slides against the other while the Core reports how much of its own transmitter it can still hear. What leaks through after that is cancelled electronically. With the coil held a metre clear of ground and metal, the Core’s second output plays the seven tones back against the leak, adjusting their weights until the residue is 40 dB down.

Then the weights are frozen. A canceller left adapting during a survey removes targets as readily as leakage.

Ground that answers first

Goldfields ground is iron-stained, and the coil hears it before any target, differently each time the coil dips. A detector handles this with a ground balance that tracks one or two numbers and subtracts them. Goldray fits the ground to a physical model instead.

With seven tones and the coil’s height known, the ground splits into two parts. Fine magnetic grains give an in-step part that falls slowly with frequency and a steady lagging part; conductive soil gives a lagging part that rises with frequency. Soil changes from place to place rather than moment to moment, so the tracker updates as the coil moves and stops while a target is in view. A by-product is a map of mineralisation, which a prospector may value as much as the targets.

hgrains · 1/h³eddies · 1/h
Figure 13Magnetic grains and conductive ground answer differently, and both depend on the coil’s height.

How loudly the ground answers

1e-41e-31e-21e-125102040coil height above ground, cmmagnetic soil, S_χ (m)1e-51e-41e-325102040coil height above ground, cmconductive soil, S_σ (m³)DD, at its nullsingle loop
Figure 14Ground coupling against coil height for the double-D and a single loop, one turn each.

A single loop hears more and more ground as it nears the soil. The double-D, unexpectedly, does not: its ground signal peaks with the coil about 5 cm up and falls again below that, so a coil bobbing close to the surface disturbs it less. Ordinary soil conducts at 0.001 to 0.1 S/m and the model holds to about 0.14 near the top of the band; salt ground, at 1 to 10 S/m, is far past it.

Twenty-one centimetres, then five

Moving from mild soil to moderate costs a 1 g nugget most of its predicted depth, and the loss traces to the coil’s height.

How far down it can see, by soil

no soilmildmoderatesevere5101520250depth, cm2121lost5 cm16 cm lostnoise, × thermala 5 mm height error leaks ground1.6×462.95.510203772×8tone, kHz · log scale to ×1000
Predicted, one pass at 0.6 m/s, coil 5 cm up, its position reported with 5 mm of error. Only these four soils were computed; nothing between them is interpolated.

The tracker subtracts the ground at the height it is told, and the coil is somewhere else. With 5 mm of error in the reported height, moderate soil lifts the noise to 46 times its floor at the lowest tone and 8 times at the highest. In severe soil the factors are 862 and 143, and of the test objects only the steel and aluminium balls are found.

In moderate soil, 5 mm of error in the coil’s height costs a 1 g nugget 16 of its 21 centimetres.

Against a commercial detector, the exposure cap costs about 6 dB, and with no pulse mode it loses to pulse detectors in the worst ground. The cure for the cliff is a ground balance that does not trust the reported height. It has not been written, so the moderate and severe figures stand as the forecast.

Two millimetres, or no name

Finding a target tolerates a centimetre of error in the coil’s position; naming it needs far less.

Signature error against position and clock error

coil track as reportedtrue sphere dashed · 2.1 mm off6.9 %eigenvalue error, 14 mm sphere at 10 cm · inside the limit10 % limit1101001000
Predicted. Position error and clock error were simulated one at a time, so moving one slider returns the other to zero. Each stop is one simulated case.

The simulation runs the whole chain on a 14 mm gold-alloy sphere 10 cm down, swept three times with a realistic wobble in the position estimate. With perfect position the recovered values are 1.5 % out; at 2 mm of error they are 6.9 % out, and at 3 mm 12.0 %, against a limit of 10 %. At 5 mm the fit loses the object and places it 127 mm away.

Timing is kinder. At a swing of 0.6 m/s a millisecond is 0.6 mm, and even 10 ms of mismatch between the phone’s clock and the Core’s leaves the sphere 2.0 % out.

A coil swept one way pushes on a buried object mostly along one axis, so it can place the object and still misjudge its shape. At 2 mm of error one pass leaves the values 12.9 % out, two passes 9.1 % and three 6.9 %. A class is therefore shown only once a target has been crossed twice, at right angles.

90°pass 1pass 2a class only after both
Figure 15Two passes at right angles over one target.

A phone manages about 10 mm, and closing the gap to 2 is the largest open problem in the design. One candidate is a motion sensor on the Core to bridge between the camera’s fixes. Another is autofocus: a coin is swept from two directions and the assumed geometry between phone and coil is tuned until the image is sharpest. In simulation that recovers a 3.6 cm error in the geometry to 1 mm.

When the phone goes blind

A few steps into scrub, the phone’s camera loses sight of the ground. Position stops, or arrives late, and the coil is still swinging. So the Core answers at four speeds, and the fastest needs no position at all. It subtracts the tracked ground, sums what is left across the tones and sends loudness and pitch to the phone inside 80 ms, like an ordinary detector.

The slower answers build the map, and then a named target, as position allows. With no phone at all, a buzzer on the Core sounds the fast tone and the survey downloads afterwards.

Four speeds of answer

10 ms100 ms1 s10 s1 min10 minlatency (log)Tier 0 · instantTier 1 · mapTier 2 · targetTier 3 · offline
Figure 16Time allowed for each answer. None has yet been timed on hardware.

Seventeen objects and three blanks

The first real test would be a garden: a plot 2.4 by 3.0 m with twenty marked spots 60 cm apart, seventeen holding a buried object and three left empty. The score is what the Core finds against what it invents.

It could stall before then on the hardware. The audio board’s maker quotes a response to about 70 kHz, and if the top tone rolls off it may have to come down to about 63 kHz, which is the tone the smallest gold needs most. And on a fast desktop machine the ground tracker and the fast tier already take 2.4 ms a frame, against 2 ms allowed on a much slower Pi.

0.60 m2.4 m17 objects, 3 blanks
Figure 17Twenty sites, three of them blank. Which three is illustrative.

If the Core gets through the garden, the next complaint will be about the wiring and the weight. That is BLOCK III.

BLOCK III · A CARRIER BOARD

Fewer wires, fewer grams

Suppose the Core passes and goes out for a second season. A hand-soldered wire lifts every few outings, and the enclosure, phone arm and battery drag on the wrist: 436 g carried against a 350 g target. This block adds no physics. It moves everything hand-wired onto one board and takes grams off the printed parts, and none of it is designed yet, because no Core exists to show what needs changing.

Nanovolts beside an amplifier

DAC L+−OPA548T3 kΩ1 kΩgain 4 · ±15 V4.7 Ω1 Ω22 Ω10 nFZobelcurrent sense → ADC RM12TX winding40 turns · 0.635 mH303 Ω at 76 kHz
Figure 18Transmit stage. A power amplifier at a gain of 4 drives the coil, through a small resistor that turns its current into a voltage the Core can read.
RX winding120 turns · 7.88 mH6.8 kΩdampingDAC R · nullin series with the return1 kΩ over 10 Ω+−OPA1612191 Ω : 10 Ω×20gain 402 · 52 dBM12 → ADC Ldifferential120 kHz band limit1.77 nV/√Hz at 1.6 kHz8.42 nV/√Hz at 72 kHz
Figure 19Receive pod. The cancelling signal enters in series with the coil, ahead of a gain of 402.

The board has to carry two circuits that dislike each other. One pushes 0.1 A through the transmit winding without oscillating. A hand’s width away, the other has to hear signals measured in nanovolts. Both hold in simulation.

Noise referred to the receive winding runs from 1.77 nV/√Hz at the bottom tone to 8.42 at the top, so a 2 nV target holds only to about 8 kHz. Above that the winding itself sets the floor, and a different amplifier does not help.

Where the noise target is lost

0.10.20.5125101k2k5k10k20k50k100ktone, Hz (log)receive noise referred to the coil EMF, nV/√Hz (log)target 2total, 8.42 at 72 kHzpreamp current noise, 6.07 at 72 kHzdamping resistor, 5.57 at 72 kHzpreamp voltage noise, 1.12 at 72 kHzwinding, thermal, 0.85 at 72 kHzfirst estimate, 5.66 at 72 kHz
Figure 20Receive noise at the seven tones, term by term, as modelled. The 2 nV target holds to about 8 kHz.

Eighty-six grams over

battery sledRay Corephone armlower rod, boughtreceive poddouble-D coilrod shortened · 436 g carried against a 350 g target
Figure 21Goldray as modelled. The whole instrument, with the rod shortened to fit the drawing. Modelled, not printed.
lid, O-ring groovePi Zero 2 W + codec HATamplifier, shunt, supplydovetail to the shaft clampM12 to the coil
Figure 22Core enclosure. Lid lifted, boards seated; 200 × 109 × 44 mm.

Core under the grip, phone ahead of it, battery behind the elbow, and every swing for the rest of the day moves all of it. Seven separate boards wired together are the obvious place to look for 86 g.

Stiffness matters more than mass. Phone, Core and coil must not move against one another during a swing, because the Core treats the phone’s position as the coil’s. Under a 300 g phone the arm is estimated to bend 0.1 mm against a 2 mm budget, though flex in the clamps, the rod and the coil’s mounting ears is not modelled.

fork for the phone holderdovetail footsplit clamp25 mm rod
Figure 23Every joint a clamped dovetail; no ball joints.

A tidier Core still listens on one channel and still has no answer for the worst ground. Both are left to BLOCK IV.

BLOCK IV · ITS OWN FRONT END

Ground the tones cannot handle

An ironstone flat in Western Australia’s goldfields: 40 degrees, hot rocks underfoot and soil thick with fine magnetic iron oxide. On ground like that the seven-tone Core is predicted to find only the steel and aluminium test balls, and a pulse detector would go deeper. This block stops borrowing an audio chip. Its own converters would listen on several channels at once and add a pulse mode, and it gets built only if the board before it shows the need.

More channels
A three-axis receive cube or a small array, so a full signature could come from a single pass.
Pulse induction
A pulser and a fast converter for ground the tones cannot handle, after its own high-voltage safety review.
Rough cost
AUD 400 to 800 a unit, unverified.

Learning from a simulator

More channels and more computing invite a neural network on top of the physics. A first one, called RayNet, has been trained on 700 simulated windows of rays and scored on 300 more. It rates 0.78 on detection where the plain physics rates 0.81, on a scale where 1 is perfect and 0.5 is a coin toss.

A network trained on the same simulator that built the physics has nothing to learn that the physics lacks, so the loss is no surprise. Any real advantage would have to come from what the simulator leaves out (lumpy nuggets, odd soils, interference, the person swinging), and only real dig results can show it. Either way the network is never allowed to report a target the physics cannot explain.

A learned detector, no better than the physics

00.250.50.7510.050.10.20.5false alarms per square metre sweptprobability of detectionPhysics, AUC 0.808RayNet, AUC 0.780Physics × RayNet, AUC 0.811
Figure 24Detection rate at four false-alarm rates, on 300 simulated windows.

By then the instrument would be several boards in several boxes, each put together by the person who swings it. BLOCK V folds them into one.

BLOCK V · ONE SEALED UNIT

One a stranger could buy

A stranger at a prospecting club asks where to buy one. The last block would be a single sealed board in a moulded housing, with a small radio chip in place of the Pi and the map and target-naming moved onto the phone, at a rough AUD 150 to 300 a unit in volume. It is two blocks of evidence away.

The idea of blocks comes from Apollo, whose spacecraft were built as Block I and Block II. Goldray’s version freezes the interfaces and changes only what a measurement has condemned. The ray record, the link between phone and Core and the plug on the coil stay the same from block to block, so a recording made with one block replays through the next, and a coil wound for the Core plugs into the sealed unit.

Where it stands

Nothing in Goldray has been built, wound, printed or swept over real ground. Software exists for the first two blocks and its physics has been checked against exact solutions, and every depth and noise figure in this piece is a prediction.

The next step needs a person with a soldering iron, a printer, a coil former and a patch of ground free of metal. The cheapest one needs less. BLOCK I asks for no coil, no Core and no soldering, and if a phone on a shaft cannot hold its position to a few millimetres over a target, the project learns it for about AUD 30, before any other block is built.

Method and sources

Facts and figures come from the goldray repository at commit 70f75a5, read on 20 November 2026: its README, its concept, architecture, signal-processing, app-integration and hardware-blocks documents (docs/00, 01, 04, 07 and 13), its decision and progress logs, the simulation reports S1 to S6 and the RayNet evaluation, with the draft branch m7-cad at 31c99f0 for the CAD, the circuit values and the 436 g mass. The depths, noise multipliers and soil cliff are parsed from S5, the pose, sync and pass figures from S6, the coil and null from S1 and S2, the ground model from S3, the receive noise from S4, and the 0.78 and 0.81 detection scores from ml/reports/0.0.1-sim.md. The aperture and pose charts were made by running goldray’s own code (scripts/goldray/export_figures.py in this site’s repository, writing components/labs/goldrayData.ts); the CAD plates are hidden-line projections of its CadQuery parts (scripts/goldray/cad_plates.py), and the other plates are drawn by hand and schematic. Sphere conductivities are textbook values (gold 4.1 × 10⁷ S/m, copper 5.8 × 10⁷ S/m, steel 5 × 10⁶ S/m with μr 100), and nugget masses are converted to radii at 19.3 g/cm³. Costs are the hardware-blocks document’s rough estimates. The project’s own documents number the blocks 0 to IV; this piece renumbers them BLOCK I to BLOCK V. The Kingower afternoon and the other scenes are hypothetical, the ground section is illustrative and not surveyed, and the two finds are cited only to place the country. The goldray specifications, code and CAD were produced by Claude Code sessions under the author’s direction. No hardware was built, no field was transmitted and nothing was measured.

RefSourceWhereAs ofAccessUse
COConcept, architecture and risksgoldray README.md, docs/00-concept.md, docs/01-architecture.md2026-11-11internalcontext
SPSignal processing: tone rules, lock-in, null, ground model, forward model, answer tiersgoldray docs/04-signal-processing.md2026-11-11internalcontext
TNTone planner output, product clearance and amplitudesgoldray src/goldray/dsp/tones.py; tests/test_tones.py2026-11-11internalderived
S1S1 field and kernel library: validation and idealised double-D nullgoldray sim/reports/S1-kernels.md2026-11-11internalderived
S2S2 reference coil: turns, inductance, resonance, null overlapgoldray sim/reports/S2-coil.md2026-11-20internalderived
S3S3 ground models: coupling against height, conductive term, validity limitsgoldray sim/reports/S3-ground.md2026-11-11internalderived
S4S4 analog chain: transmit-loop stability, receive noisegoldray sim/reports/S4-analog.md2026-11-20internalderived
S5S5 end-to-end link budget: detection depth by soilgoldray sim/reports/S5-link-budget.md2026-11-20internalderived
S6S6 inversion conditioning: pose noise, sync error, number of passesgoldray sim/reports/S6-conditioning.md2026-11-20internalderived
BPReference-scene back-projection and pose-noise sweep, run for this piece with goldray’s codescripts/goldray/export_figures.py (APERTURE, POSE)2026-11-11internalderived
TGAnalytic target polarizabilities: sphere, shorted loop, ferritegoldray src/goldray/sim/targets.py2026-11-11internalderived
RNRayNet 0.0.1-sim evaluation on simulated windowsgoldray ml/reports/0.0.1-sim.md2026-11-20internalderived
RGRegulatory check: LIPD Class Licence 2025, ICNIRP 2010 exposuregoldray docs/regulatory-check.md, docs/11-safety-regulatory.md2026-11-11publicderived
HBHardware blocks, costs and the standalone appgoldray docs/13-hardware-blocks.md; docs/07-goldfall-integration.md2026-11-20internalcontext
DPDecisions and progress loggoldray docs/DECISIONS.md, docs/PROGRESS.md2026-11-20internalcontext
CDMechanical CAD and schematics: parts, assembly, massgoldray branch m7-cad: cad/goldray_cad/, cad/PRINTING.md, hw/schematics/2026-11-20internalderived
HFHand of Faith nugget: 875 troy ounces (27.21 kg), found with a metal detector near Kingower, Victoria, on 26 September 1980, about 30 cm downWikipedia, “Hand of Faith”. Other accounts give a shallower depth and slightly different weights2026-11-20publiccontext
GTA 4.6 kg rock holding about 2.6 kg of gold, found in Victoria’s Golden Triangle with a budget metal detectorKitco News, 12 April 2023, updated 14 April 20232026-11-20publiccontext
  • Electromagnetic Induction
  • Synthetic Aperture
  • Signal Processing
  • Inverse Problems
  • Prospecting
  • Raspberry Pi
  • Methods