Methodology

What this is

The Bitcoin Weigh-In dataset records daily closing prices in US dollars for a curated set of fungible commodities from 2013-01-02 to the most recent completed UTC day. From those closes it derives per-BTC equivalents (how many troy ounces of gold, pounds of copper, or barrels of crude one bitcoin could have purchased on each day) and pairs them with a deterministically computed BTC circulating supply. The artifact is a single small file — around 800 KB as CSV, 700 KB as Parquet — that any analyst, journalist, or hobbyist can download once and analyse offline without an API key.

This document describes how the data is collected, what the published flags mean, how cross-validation works, how versions are cut, and how to report corrections. The companion dataset page ships the artifacts; this page describes the rules behind them.

Data sources

Three providers between them cover every live series. Each commodity is pinned to a single primary endpoint so the dataset has one parser, one rate-limit regime, and one place to look when something disagrees with the rest of the financial press.

CoinGecko

The primary source for BTC-USD (coin id bitcoin) and for gold, priced via Pax Gold (pax-gold) — a token redeemable for one fine troy ounce of LBMA gold that tracks spot within a small premium. Both come from CoinGecko's keyless public API (market_chart); the daily job records the last price of each UTC day. No API key is required, so the shared pool is IP-throttled and the job backs off on HTTP 429.

GoldAPI.io

The primary source for silver spot (XAG/USD, USD per troy ounce). The daily job sends the key in the x-access-token header, and a redacted form of every fetched URL is recorded in /health.json so an authentication failure surfaces clearly rather than presenting as silent forward-fill.

FRED (St. Louis Fed)

The primary source for Brent crude (DCOILBRENTEU). FRED redistributes the EIA spot price daily, typically with a one business-day lag. The daily job retries transient HTTP errors on a backoff and forward-fills if the value never arrives.

Stooq (retired)

Stooq was the original source for BTC, gold, silver, and several deferred commodities (platinum, copper, CBOT wheat, ICE coffee). It was dropped on 2026-06-13 after it began blocking automated access. BTC, gold, and silver moved to the providers above; the deferred commodities are not rendered in the interface and their historical values remain frozen in the dataset.

Derived (no API)

BTC circulating supply is computed in scripts/sources.ts as a pure function of days-since-genesis. Genesis is 2009-01-03; the protocol targets 144 blocks per day, the initial block reward is 50 BTC, and the reward halves every 210,000 blocks. The implementation walks halving eras and accumulates supply era-by-era. Because every input is a constant of the protocol, the column has no API dependency and is unit-tested against known halving block dates.

Cash (derived, no API)

The Cash commodity has no price to fetch: one $1 Federal Reserve Note is worth exactly one dollar, so its "price" is the constant 1, and the note count is simply the live BTC-USD value. The only external data are fixed physical constants from the U.S. Bureau of Engraving and Printing — note length 155.956 mm (6.14 in), width 66.294 mm (2.61 in), thickness 0.10922 mm (0.0043 in), mass 1 g — identical across every denomination and unchanged since the note's current size was adopted in 1929. These live in src/lib/billStack.ts and are cross-checked in that file's tests against the Bureau's own public trivia that a banded strap of 100 notes runs about 0.43 inches thick.

The cash stage

The Cash tab draws every note at that true size, in the units cash is actually handled in (cashParts in src/lib/billStack.ts): a loose stack under 100 notes; from 100, straps of 100 under the blue $100 band (the American Bankers Association colour for a strap of $1s), in piles of six; from 1,000, bundles of ten straps, stacked roughly as a cube; from ten million, pallets of 1,000 bundles (10 × 10 × 10, 0.66 × 1.56 × 1.09 m and one tonne of paper, on a 14 cm wooden pallet under shrink-wrap). Past 60 pallets they stand as one block faced with its pallets. The count is exact: a remainder is a shorter last bundle or a part-loaded last pallet, never rounded up to a whole one, and the few notes lying loose on the floor (up to seven, curled, folded or face down) come out of that remainder — never an extra, and never by breaking open a whole strap or bundle. The note art is printed at runtime and is deliberately not a reproduction of genuine Federal Reserve Note artwork: Sat takes the portrait, the serials are invented, and the note is "weighed, not signed". The bundle and pallet arrangement is set dressing; the sizes, counts and mass are not.

The Moon ride

Once the notes would stack taller than a doorway, the Cash stage offers to restack them as one column of notes laid flat, 0.10922 mm each (1 BTC at $84,550 is 9.2 m), and to ride up it (src/lib/moonRide.ts). The climb is logarithmic in altitude, every factor of ten taking the same time, because the tallest stack spans nine orders of magnitude. It passes real heights: the Statue of Liberty (93 m), the Eiffel Tower (330 m) and the Burj Khalifa (828 m), drawn as true-scale silhouettes around the column, then Mount Everest (8,849 m), airliners' cruising height (about 11 km), the edge of space (the Kármán line, 100 km), the International Space Station (about 408 km), the GPS satellites (20,200 km) and geostationary orbit (35,786 km). The Moon is at its average distance, 384,400 km, the same figure the readout's "of the way to the Moon" has always used, drawn with its near side there. The Earth is painted from the site's own world map, with the column standing at New York. In the closing side view the stack is drawn as a line once its true width falls under a couple of pixels, and the Earth and the Moon, drawn at true size, are ringed, because at that scale they are a few pixels across. The closing card gives the whole 21-million supply's height at today's price, and the price at which it reaches the Moon: about $167,595 per BTC. That figure doesn't move with the market, since one note is one dollar (384,400 km ÷ 0.10922 mm ÷ 21,000,000).

Manhattan (land, not a commodity)

The Manhattan tab shows how much of Manhattan's ground a sum of bitcoin buys (src/lib/manhattan.ts). The price is the ground itself, from the most thorough estimate of the island's land value: Jason Barr, Fred Smith and Sayali Kulkarni, "What's Manhattan worth? A land values index from 1950 to 2014" (Regional Science and Urban Economics, 2018), which valued all of Manhattan's developable land at about $1.74 trillion in 2014 (range $1.54–1.95 trillion), from some 3,600 sales of vacant lots. That total is spread evenly over the developable land as drawn, so the share of the island a sum buys is simply its value over $1.74 trillion, whatever the exact area. Land, not apartment prices: an apartment's price per square foot is for floor space stacked many storeys up, while an area on a map is ground. It is a 2014 figure, the latest rigorous total, so the tab is marked illustrative, like cocaine. All 21 million bitcoin buy the whole island's developable land once one bitcoin is worth about $82,900.

The map is New York City's own open data, built once by scripts/build-manhattan.ts into a 3.6 MB file the page loads: the Department of Finance's digital tax map (every tax lot), the Department of City Planning's PLUTO land use and shoreline-clipped borough boundary, and building footprints with measured roof heights (Office of Technology and Innovation). Tax-lot outlines run out over the rivers in places, so every lot is clipped to the shoreline first: 42.3 km² of lot outlines become 32.4 km² of dry land, once parks and open space (PLUTO land use 9) are left out as the land-value study did — about 55% of the borough's 59 km², close to the study's 60%. Your land fills lot by lot in order up the island from the Battery, whole lots at a time with their buildings, and the last lot part-owned as a slice from its southern edge; then the outer islands. A sum smaller than a lot is a true-size square on the most open ground of the first lot (the build finds the spot farthest from any wall). The cross street named in the readout is the northernmost whose land to the south the sum covers.

Forward-fill logic

Markets close on weekends and public holidays. Source endpoints occasionally drop a single day's row even on a normal trading session. In both cases the daily job carries the previous known value forward so that every calendar date from coverage start to last update has a row in the dataset. The decision to forward-fill rather than emit nulls is an honest one — analyses that join across commodities need a value for every date, and the alternative (per-commodity NaN) silently propagates into derived calculations.

v1.0 of the dataset ships a forward_filled column populated as empty string for every row, because per-row fill provenance is not reconstructable from the historical NDJSON that was bootstrapped before this column existed. Prospective per-row tracking begins in v1.1, at which point the column will hold a pipe-delimited list of the column names that were forward-filled on that date — for example brent_usd on a day where most feeds returned values but the FRED Brent series had not yet published.

The daily cron writes a separate fill record per source into /health.json on every run, so the present day's fill state is always visible even before per-row tracking lands. The cron also exits non-zero if every source returns zero rows on a UTC weekday — a signal that authentication, rate limits, or upstream infrastructure has changed, rather than silent fill propagating an undetected outage.

Forward-fill is why the year-by-year snapshots can quote a value for every year-open and year-close: those boundary dates frequently fall on a weekend or holiday, and the figure shown is the last known close carried forward, not a trade that happened on 1 January.

BTC supply derivation

The btc_supply column is deterministic. For a date D:

1. Compute days since the genesis block at 2009-01-03.
2. Multiply by 144 (the protocol's target blocks per day) to get an approximate cumulative block count.
3. Walk halving eras of 210,000 blocks: era 1 pays 50 BTC per block, era 2 pays 25, era 3 pays 12.5, era 4 pays 6.25, era 5 pays 3.125, and so on. For each era, add min(era_end, total_blocks) − blocks_so_far times the era's reward.
4. Round to an integer count of BTC.

The approximation drifts a few thousand BTC from reality (real interblock times vary around the 10-minute target, and mining hashrate growth nudges blocks slightly faster than schedule), but the error is small enough — under 0.1% across the full coverage range — that the column is fit for the visualisation's purpose: showing where on the supply curve any given date sits. Analyses that need block-exact supply should pull from a node or a block explorer; this dataset's column is a clean closed-form schedule.

Illustrative pricing

Two of the five commodities rendered in the visualisation — Plutonium-238 and cocaine — do not have public spot markets. Their prices on the site are illustrative composites constructed from named sources, with the as-of date carried alongside. They appear on the main visualisation but they are not in the live dataset published under /data, which holds only live market closes. A third commodity, the LEU uranium fuel pellet, follows the same pattern but is currently deferred from the visualisation; its illustrative price record persists in the repository for later re-enable.

Plutonium-238

Composite material-cost estimate of ~$5,000/g (midpoint of a $4,000–$8,000 range) derived from the DOE Office of Nuclear Energy, NASA Planetary Science Division publications on the Pu-238 production program (~$150M/year for ~1.5 kg/year), the Cassini OIG report from 1997 ($1,968/g escalated to 2024 dollars), and Atomic Insights' analysis of RTG heat sources. A separately cited fully-loaded program cost (~$100,000/g) reflects the facility maintenance and regulatory infrastructure required for production but is less directly comparable to other commodities' market prices, so the material-cost figure drives the BTC equivalence on the visualisation. Uncertainty bounds: roughly ±60% around the midpoint at the material-cost layer. As-of date: 2024-12-31.

Density and the cube. The visualiser sizes the plutonium cube from the oxide fuel — plutonium-238 dioxide (PuO₂), the ceramic form that radioisotope thermoelectric generators actually burn and that glows — not the pure metal. The cube edge is computed at PuO₂'s theoretical density of 11.46 g/cm³. Real sintered fuel pellets are deliberately pressed to roughly 80–90% of theoretical density (a controlled porosity that accommodates helium from alpha decay without cracking), so an actual pellet of the same mass occupies 10–25% more volume — a slightly larger cube — than the one drawn here. We render the theoretical-density cube because it is the single unambiguous figure; the caveat is that real fuel is a little less dense, and therefore a little bulkier, than the idealised block.

LEU uranium fuel pellet

Composite cost of ~$20 per 7 g pellet from the World Nuclear Association "Economics of Nuclear Power" methodology, cross-checked against the IAEA/OECD-NEA Red Book 2024. Decomposes as: U₃O₈ feed at ~$100/lb, conversion to UF₆ at ~$20/kgU, enrichment at ~$150/SWU, fabrication at ~$300/kgU, yielding ~$3,000/kgU of finished fuel; divided by 7 g/pellet ≈ $20/pellet. Uncertainty bounds: ±30% by contract terms, enrichment level, and market conditions. As-of date: 2025-01-01.

Cocaine (three-tier)

There is no spot market for cocaine. The composite presents three tiers reflecting the market's actual structure: producer (~$2,500/kg, range $1,500–$3,500, raw refined base, UNODC World Drug Report 2024); wholesale (~$30,000/kg, range $25,000–$35,000, ≥80% pure US wholesale standard, UNODC 2024 / DEA NDTA 2024); and retail purity-adjusted (~$120,000/kg, range $80,000–$250,000, normalised to 100% for cross-tier comparison, DEA / EMCDDA). Wholesale is the primary tier for BTC equivalence because it is the most directly comparable to how other commodities are priced (standardised purity, kilogram-scale transactions). As-of date: 2024-12-31.

The cocaine stage

The Cocaine tab draws the bought mass at true scale beside Sat, in the units it would actually come in (src/lib/cocaine-scene.ts): under a gram, lines of 30 mg chopped out on a mirror, each ~6 cm × 4.5 mm and, at a loose-powder bulk density of 0.5 g/cm³, under half a millimetre tall, with a 43 × 22 mm razor blade alongside; from 1 g, 1 g zip-lock baggies (5 × 6 cm) heaped where they fall; from 1 kg, pressed, taped 1 kg bricks (21 × 14 × 3.8 cm, about a hardback book, ~0.9 g/cm³ as wrapped), in a row and then stacked five by seven, a pallet's layer; from 1 t, shrink-wrapped pallets (1.2 × 1.0 m) of 1,000 bricks, 29 courses high. Past 120 pallets they stand as one warehouse block at the true pallet count. A trailing part-unit is drawn (a short line, a half-filled bag, a cut brick, a part-load) once it reaches 15% of a unit; the count above the stage carries the exact figure. The tape colours and the ink stamp are set dressing; sizes, counts and masses are not.

Impossibility lines

Past a certain quantity, comparing to a real-world object stops being honest — there is no gold bar, silver hoard, or production run of that size. Four figures mark where the readout switches from "here's what this looks like" to "this has never existed," and each overshoot is stated as a multiple computed from these figures, never a hardcoded number:

• All gold ever mined: ~213,000 t (World Gold Council 2024) — already the largest gold quantity anchor on the site.
• All silver ever mined: ~1,740,000 t (USGS / The Silver Institute) — already the largest silver quantity anchor on the site.
• Global annual cocaine production: ~2,250 t/yr (UNODC 2024 estimate, the figure already used by the denomination and brick-stack copy above). The same report's newest headline number, for 2022 output, is a record ~2,757 t — sources on annual cocaine manufacture range roughly 2,250–2,800 t/yr depending on report vintage; the site keeps its previously established, more conservative figure rather than introducing a second, slightly different total for the same concept.
• Global annual Pu-238 production: the US restart at Oak Ridge has ramped from ~50 g/yr (2015) through ~400 g/yr (2023) toward a ~1.5 kg/yr steady-state target by 2026 (DOE Office of Nuclear Energy / NASA Planetary Science Division). The site uses the 1.5 kg/yr target — the higher, more conservative figure — for its "years of global production" framing.

Plutonium-238 decay heat

Pure Pu-238 metal generates ~0.567 W/g of decay heat. The visualiser depicts the oxide fuel (PuO₂, see "Density and the cube" above), whose effective specific power is lower — approximately 0.40–0.42 W/g of oxide — because RTG-grade material is not isotopically pure Pu-238 and because PuO₂ carries the oxygen's extra molecular weight. The readout uses the midpoint, 0.41 W/g of oxide, applied to the same oxide mass the cube is sized from. Source: DOE Office of Nuclear Energy / NASA Planetary Science Division published RTG fuel specifications.

The live visualiser: camera and staging

The home page renders each commodity as a real-time 3D cube at true physical scale, beside a rigged Shiba Inu that acts as the constant scale anchor (40 cm at the shoulder). Because the cube spans roughly six orders of magnitude — from a sub-millimetre fleck of gold to a silver monolith tens of metres on a side — a single fixed shot cannot stay convincing across the whole range. The scene solves scale the way a photographer does: with the camera. The honesty rule is that every apparent size on screen derives from one declared camera geometry per frame, never from an artistic fudge.

Scale reference: Sat, Shiba Inu, 40 cm at the shoulder, 9 kg. Constant.

Camera model

A single perspective camera (35° field of view) frames the scene, and exactly one geometry is in force at any instant. As the cube grows the camera dollies along a banded path — macro framing for the speck (the cube held at a fixed fraction of the frame so it never vanishes, with a 5 cm floor), the familiar two-shot when cube and dog are comparable, and a wide shot when the cube towers. The transitions are continuous at the band crossovers by construction, and the camera's height is capped at one metre so the largest cubes are genuinely looked up at rather than viewed from above. The damped easing between framings is itself the scale cue: the longer the camera travels, the bigger the change in size you are being shown.

Staging honesty

Once the cube grows past about 1.2 m on a side, the Shiba walks to a fixed mark in the near foreground while the cube recedes into the distance — the standard photographer's trick for conveying the size of something enormous. This means the dog and the cube are no longer the same distance from the camera, so their on-screen sizes are governed by real perspective (an object twice as far away appears half as large) rather than by a shared scale factor. That is a true depiction, not a trick of the eye, but because it differs from the simple side-by-side comparison the readout says so explicitly — "Sat is standing nearer the camera" — whenever the dog is staged in the foreground. Nothing on screen is ever resized by feel; the apparent sizes always follow from the one declared camera geometry.

The drop

While you hold the slider, or drag the cube itself, the camera stays where it is and the cube grows or shrinks in place — it is allowed to outgrow the shot rather than being quietly re-framed. When you let go, the camera steps back to its proper framing and the cube is dropped from its own height (its underside one edge-length above the floor) under standard gravity, 9.80665 m/s², in real time. Nothing about the fall is slowed down or sped up, which makes the fall time a scale cue in its own right: a speck lands in milliseconds, the whole supply of gold takes more than a second. The readout states the fall time and the impact energy, m·g·h; where that exceeds a gram of TNT it also gives the TNT equivalent, using the standard convention of 4.184 GJ per tonne. The camera shake, the dust and Sat's reaction are all driven by that one impact-energy figure; none of them is tuned per amount.

The same line gives the cube's resting bearing pressure — its weight over its own footprint, which for a cube reduces to density × g × edge — and compares it with the presumptive load-bearing values in the International Building Code (Table 1806.2): 1,500 psf for clay, 2,000 for sand, 3,000 for gravel, 4,000 for sedimentary rock and 12,000 for crystalline bedrock. Past the bedrock value the floor is drawn cracked. These are allowable pressures, which include the code's safety margin, so exceeding one means a structural engineer would not sign it off, not that the ground would necessarily give way. A gold cube reaches the bedrock value at about 3.03 m on a side, whatever the gold price.

The sound

With the sound on, you hear the landing, and a block landing flat on a concrete floor is heard mostly through the floor. The blow is modelled as a mass on a spring: the cube's mass against the stiffness of its flat face pressed into the concrete (a flat punch on an elastic half-space, in series with the cube's own stiffness). So it lasts τ = π√(m/k) and delivers the cube's momentum, m·v·(1 + e). The concrete follows ACI 318: E = 4,700√f′c MPa at f′c = 30 MPa, about 25.7 GPa, with Poisson's ratio 0.2 (Eurocode 2). Mass grows with the cube of the size and stiffness only in proportion to it, so the blow lengthens in step with the edge: about 0.09 ms for 1 BTC of gold, 0.7 ms for 500 BTC, 26 ms for the whole supply. That is why small cubes tick and big ones thud. The peak force is startling: about a tonne for 1 BTC dropped 3 cm, and about 190 tonnes for 500 BTC.

What you hear is that force pulse radiated by the floor slab, taken as 15 cm of concrete. A point force on a plate, below the plate's coincidence frequency, radiates a pressure that follows the force itself (Cremer, Heckl & Petersson, Structure-Borne Sound). On top of that comes a sharper crack from the cube itself stopping dead, which dominates for small cubes. Past the building code's bedrock allowance, where the floor is drawn cracked, a burst of fracture clicks is added; their timing and texture are illustrative, and only whether the floor cracks is computed. The whole thing plays in a stated room: a large, hard-floored studio with a 1.4-second reverberation time.

The cube's own ring is left out on purpose. We computed its free-vibration modes with the Rayleigh–Ritz method used in resonant ultrasound spectroscopy (Visscher et al., J. Acoust. Soc. Am. 90, 1991), checked against beam theory and against Lamé's exact solution for a cube. Its loudest note is the shear wave speed divided by (√2 × edge): 27 kHz for 1 BTC of gold. But even undamped, the ring carries under a three-hundredth of the thud's acoustic energy for a 500 BTC cube, and a block pressed flat on concrete loses it into the floor within a few cycles. The wave speeds used are 3,240 and 1,200 m/s for gold and 3,650 and 1,610 m/s for silver (CRC Handbook of Chemistry and Physics); for plutonium dioxide they come from the measured shear modulus of 89 GPa and Poisson's ratio of 0.32 (Kato & Matsumoto, IAEA INIS). That abstract's own Young's modulus, 219 GPa, is 7% lower than those two imply, which would change the blow by less than half a percent.

Loudness is compressed from the impact energy. A real whole-supply landing would be deafening, and most of its energy is too low for a phone speaker to play; you would feel it more than hear it. Sound is off unless you switch it on; the clip recorder always includes it.

Camera optics

The stage camera is modelled as a real one: a full-frame sensor (24 mm tall) with the focal length its 35° field of view implies (about 38 mm), at f/2.8, focused on the cube. Every pixel is blurred by that lens's thin-lens circle of confusion for its distance, c = A·f·|d − S| / (d·(S − f)). That is why small amounts come out looking like macro photographs, with Sat's paws soft behind a sharp speck of metal, and large amounts come out sharp from front to back: a real camera at 20 cm and at 25 m behaves exactly that way, and it is the same cue that makes tilt-shift photos of cities look like toys. The blur is capped at 5% of the frame height for legibility, so it is only ever understated; at the closest framings a real lens would blur the background more. No atmospheric haze is added: over the distances in the scene (under about 100 m) real air adds none.

The magnifier

Below about eight pixels on screen the cube cannot be resolved at all, so a circular magnifier appears. It re-renders the scene from the same camera position through a narrower field of view — a true optical magnification, like a loupe or a telephoto lens — and states its power, chosen from a fixed ladder (×2, ×5, ×10 … ×10,000) so that the magnified cube reads at a legible size. One satoshi of gold, about 68 µm on a side, appears at ×200 on a typical screen.

Hashweight: network physical mass

The Hashweight panel estimates the total physical mass of the hardware that secures the Bitcoin network. It is an order-of-magnitude estimate — treat all figures as having roughly ±30% uncertainty — derived from three independently sourced inputs: live network hashrate, published ASIC specifications, and publicly disclosed node counts.

To see what one of those machines does inside, the Inside a Bitcoin miner page takes a mining chip apart and simulates a single SHA-256 hash core on real block headers.

Live hashrate

Current network hashrate is fetched at page load from the mempool.space mining API (/api/v1/mining/hashrate/1w), which returns a 7-day rolling average in H/s. The fleet model and its timeline use /api/v1/mining/hashrate/all, which provides weekly averages back to Bitcoin's origin (the panel starts in 2014). If the history is unreachable, the panel shows today only, from the fixed fleet average described at the end of this section; if the live figure is unreachable too, it falls back to 800 EH/s.

ASIC fleet model

Hashrate alone doesn't give a weight: one exahash per second took about 870,000 machines in 2015 and about 3,700 in 2026. So the panel models the fleet in cohorts (src/lib/hashweight/fleet.ts):

1. Capacity. Installed capacity is the running maximum of the 30-day mean hashrate. Machines switched off during a dip still exist and still weigh: when the 2021 China ban halved the hashrate for months, the machines were in transit, not destroyed.
2. Arrivals. Every rise in capacity is built from the most efficient machine on sale at the time (the frontier, below).
3. Retirement. Each cohort runs for 5 years, then is replaced, like for like in hashrate, by that day's frontier machine.

The frontier is Bitmain's Antminer line, the most-deployed family. Weights include the separate power supply the S5–S9 generation needed; later models have it built in.

Before the S5 shipped (December 2014) the model still uses it, though earlier ASICs were heavier per TH/s, so figures for 2014 are a lower bound. Other makers (MicroBT, Canaan), fleets that lag the frontier and slower retirements all move the answer, so treat every figure as ±30%. In September 2026 the model gives roughly 5–6 million machines and 80,000–90,000 tonnes.

The pile

The scene stacks every machine case edge to edge (plus the S5–S9 era's separate supplies) into one cube and draws it in true-scale side elevation beside a 1.75 m person, a 40-ft ISO shipping container (12.19 × 2.59 m) and RMS Titanic, keel to funnel tops (269 m long, 53 m tall, out of the water). Close up, the cube's face shows the front of one machine per cell. "Per bitcoin in existence" divides the fleet's mass by the supply at that date, interpolated between halvings.

Without the history

If the history can't be fetched, today's figure falls back to two blended constants: 150 TH/s and 13.5 kg per machine (an S19/S21 mix). That is within the ±30% band of the cohort model.

Node mass

Full nodes contribute negligibly to the total: approximately 20,000 reachable nodes (source: bitnodes.io) at a blended average of 0.5 kg each (Raspberry Pi at 45 g through NUC/small server at ~1.2 kg) ≈ 10 metric tonnes — under 0.02% of total network mass. The true node count including behind-NAT nodes is likely 50,000–100,000+, but even at that scale the contribution remains under 50 tonnes.

Titanic comparison

The comparison reference is the RMS Titanic's loaded displacement: 52,310 long tons = 53,150 metric tonnes. This is the actual physical mass of the ship, passengers, cargo, and fuel when she sailed. Note: the commonly cited figure of 46,328 is the ship's gross register tonnage — a volumetric measure (100 cubic feet = 1 gross ton), not a mass. Comparing a mass to a volume figure would be dimensionally incorrect, so the displacement figure is used here.

Solo miner estimate

Solo miners — predominantly Bitaxe open-source boards, home Antminers, and Nerdminers — are estimated at ~40 PH/s total hashrate and ~60,000 devices. CKPool Solo routinely reports 10–20 PH/s; allowing for other solo pools and direct-connected miners, 30–50 PH/s is a plausible range. At ~667 GH/s average per device (Bitaxe Ultra/Gamma range: 400–1,200 GH/s), 40 PH/s implies ~60,000 units. Average device weight of 0.18 kg blends bare Bitaxe boards (~0.12 kg) with heavier home ASICs. Total solo mass ≈ 11 metric tonnes, representing roughly 0.012% of total network mass.

Cross-validation

After the primary CoinGecko, GoldAPI.io, and FRED fetches complete, the daily job queries a secondary source — Massive — for the same day's close on BTC-USD, XAU-USD, XAG-USD, and (where available) XPT-USD. For each ticker where both providers return a value, the job computes the absolute percent difference. When the difference exceeds 0.5%, an entry is appended to a cross_validation_flags array in /health.json recording the date, ticker, both values, and the percent diff.

The cross-validation step is a quality signal, not a build gate. It does not fail the daily cron — a missing API key, an HTTP error, a parse failure, or a Massive ticker that doesn't exist all produce a "skipped" status without emitting a flag. This is deliberate: a secondary-source disagreement is information for an analyst, not an infrastructure outage that should block publication of the primary feed. Tickers Massive doesn't cover (continuous futures, FRED-only series like Brent) are skipped silently.

Versioning and updates

The dataset uses semantic versioning for schema changes: a major bump for removed or renamed columns, a minor bump for added columns or sources, and a patch for fixes that preserve the schema. The current version is pinned in dataset-config.json at the repository root; the artifact builder uses that value to decide which static/data/v{X.Y}/ directory to write to. Bumping the version is a manual one-line edit committed by the maintainer.

Daily updates happen at 02:00 UTC. A GitHub Actions cron fetches the previous UTC day's close from every source, appends a row to data/prices.ndjson, rebuilds static/prices.json, regenerates every artifact under static/data/v{X.Y}/, and commits the result to main. Cloudflare Pages redeploys automatically from the commit. The latest aliases at /data/prices.csv, /data/prices.json, and similar always point to the current version's artifacts; the versioned directory at /data/v{X.Y}/ persists indefinitely so prior versions remain downloadable.

Archival to Zenodo is triggered manually by cutting a GitHub release tag, at which point Zenodo's GitHub integration mints a DOI and archives the source tarball. The DOI is copied back into dataset-config.json and the next build surfaces it on the dataset page. Release cadence is keyed to schema-meaningful changes rather than the daily content updates, which keeps DOIs sparse and citable.

Corrections

To report a suspected error, email info@sortathing.com with the affected date(s) and column(s), the value the dataset shows, and where the corrected value should come from with a link. Corrections that affect a single row land in the next daily commit; corrections that affect the schema or a historical methodology trigger a minor version bump and a CHANGELOG entry. Either way, the original row stays in git history — the dataset is the current best truth, but the prior shape remains inspectable in the commit log.

Credits and licences

The dataset itself is published under Creative Commons CC-BY-4.0; see the dataset page for citation details.

Shiba Inu 3D model

The Shiba Inu used as the live visualiser's scale reference is a third-party model licensed under CC-BY-4.0, which requires visible attribution. Per the licence:

This work is based on "Animated Dog, Shiba Inu" (https://sketchfab.com/3d-models/animated-dog-shiba-inu-9abfce885a834399b2c3ccaed51cd474) by quander (https://sketchfab.com/quander) licensed under CC-BY-4.0 (http://creativecommons.org/licenses/by/4.0/)