# Stuff, money, value: build plan

**Status:** Steps 0–8 and the coda built 2026-10-29 (M0–M3, plus M4's coda). M4's §5 presets are not built yet. Owner decisions are in §12, the conceptual sweep in §13, and the build log in §15.

**What we're building:** a small stepped toy model that separates three things people run together: the *stuff* an economy produces, the *money* spent on it, and the *value* people get from it. It walks from Wilkes's identity (nominal = real × prices) through "what counts as one unit of stuff?" to how ONS, building on work by Diane Coyle and colleagues, changed the way telecoms output is measured, and what that did to measured growth and inflation. Along the way it shows where ideas about utility (value to the buyer) get into the price indices, and where cost to the maker is used instead.

**Decisions recorded (owner, 2026-10-29):**

1. **"Value" means how utility gets into the measures.** Deflators aren't only a mechanical stuff-and-money adjustment; there is a utility conception behind them. New Step 5 shows this with a hedonic laptop example.
2. **Telecoms revenue:** illustrative first, real data at M2.
3. **A standalone page** in `docs/` is the product.
4. **Double deflation (Step 8) is in the first full release.**
5. **Electricity is included in the cost of light by default.** Review once it's built.
6. **No link to the defence-vs-welfare model for now** (parked; see §5).
7. **A copy of D3 per model.**
8. **The toy is open, not closed:** a "made here or abroad?" world dial in Step 4, from M1 (decided after the sweep, §13).

**Key sources:**

- [claude_research/wilkes-bank-nominal-vs-real.md](../../../claude_research/wilkes-bank-nominal-vs-real.md), the starting point.
- The earlier chained-volume toy: [indexnumbers.js](../../../claude_research/quarto-blog-demo/posts/nominal-real-toy/indexnumbers.js) in the Quarto demo, and the plan in Dan's local `~/Code/claude/econ_measures_chainedvolume/initial_summary.md`.
- The modelling_demand research notes, especially [measuring-gva-and-productivity.md](https://github.com/DanOlner/modelling_demand/blob/master/research/notes/measuring-gva-and-productivity.md) (double deflation) and [data-streetlight-effect.md](https://github.com/DanOlner/modelling_demand/blob/master/research/notes/data-streetlight-effect.md) (the "what would an official observer see" design idea).
- The two Abdirahman, Coyle, Heys & Stewart telecoms deflator papers (2020), and ONS's Blue Book 2021 articles. All are in §14.

---

## 1. The point in one paragraph

Wilkes's argument is that the Bank of England steers *nominal* spending and the supply side decides how it splits into real growth and inflation. This model adds a second thing that decides the *measured* split: what statisticians count as one unit of stuff. Hold the world fixed (the same pounds spent, the same bulbs bought, the same phone contracts) and change only the unit. Measured real growth and measured inflation then move by equal and opposite amounts, and nominal never moves. That is roughly what happened to UK telecoms in Blue Book 2021. The research behind ONS's new method has telecoms prices falling 58% over 2010–17, where the old deflator had them rising 3%. On roughly flat revenue, that turns measured real output from a small fall into more than a doubling, with no change in the pounds spent (Step 7).

Behind every choice of unit sits an assumption about what counts as "the same thing". For consumer prices the textbook test is value to the buyer, which is a utility idea, even when it looks like a physical unit such as grams, lumens or bytes. For producer prices the textbook test is the cost of making it (Step 5). Hedonic methods try to read the valuation from market prices instead of assuming it.

That conservation law is the spine of the whole thing. Every step shows it explicitly: nominal is unchanged, and real growth plus price growth adds up to nominal growth (in log points).

## 2. Three words, kept apart

The title's three words carry five meanings, and most confusion comes from sliding between them. The toy keeps each as a separately visible layer.

| Layer | Meaning in the toy | How it is known | Adds up across goods? |
|---|---|---|---|
| **Money** | Pounds spent: price × quantity | Observed (business surveys, turnover) | Yes |
| **Stuff** | Physical things bought, in their natural unit: loaves, bulbs, laptops, contracts | Counted | No |
| **Service** | What the stuff is bought *for*: grams of chocolate, lumen-hours of light, computing speed, bytes moved | Needs a choice of unit, and that choice is the argument | No |
| **Volume ("real")** | Quantities valued at some year's prices. Real GDP | Built by statisticians from money and prices | Only in the reference year and the year after it, once chain-linked |
| **Welfare** | What it's worth to people, including the surplus above the price they paid | Never measured directly. With the right unit, real growth tracks *changes* in it quite closely (Step 6); the total is not measured at all | Not meaningfully |

"Value" is the slippery word. It can mean market value (money), real value (volume), or worth to people (welfare or utility). GDP measures the first two, but **utility is never far away**. The "economic approach" to index numbers defines the ideal consumer price index in utility terms: how much more money would you need to be exactly as well off as before? Quality adjustment (lumens, hedonic regressions) pulls the volume measure towards the service people actually value, and values it at market prices.

That is less of a limitation than it sounds. When the unit is right and the index is chain-linked, real growth tracks the *change* in welfare from cheaper stuff quite closely (Step 6). What GDP never measures is the *total* value of things (Adam Smith's water: cheap, but priceless), or the gains from brand-new and free goods. The toy should never let a step quietly move from one meaning to another.

## 3. Design principle: the world and the lens

The engine has two halves, and the interface shows them as two panels.

- **The world.** What is bought each year, at what price, with what physical characteristics (grams, lumens, lifetime, watts, speed, bytes), and where it is made. In the toy this is ground truth, set by a scenario.
- **The statistician's lens.** The choices that turn the world into official-style numbers: the unit of quantity, how much an improvement in a characteristic is assumed to be worth, the index formula, fixed base year or chain-linking, how a new product is linked in, and single or double deflation.

Each step's main switch changes the lens while the world stays put, so any movement in "real growth" is visibly a measurement effect. World dials (how much more light people use when it's cheaper, how much extra speed costs, where bulbs are made) are labelled as world dials and kept visibly apart from lens switches. The same split also lets a later version change the world while holding the lens fixed, which is Wilkes's supply-side story. This is the data-streetlight note's idea ("then generate what an official observer would see") applied to prices.

**An honesty note that must appear in the model itself:** in the lighting steps the world is generated from a demand for *light* (lumen-hours), with bulbs and electricity bought to supply it. That builds in the assumption that light is what people want, so the lumen lens is "right" *by construction* in the toy. In reality that assumption is the very thing being argued about. It is easy to accept for light and much harder for bits, which is why the ladder goes in that order (§4). The national accounts don't even have a "light" product (Step 3). Step 5 then names the assumption for what it is.

## 4. The ladder

The steps climb a **gradient of contestedness**. The first unit question is one almost nobody argues about (grams of chocolate). The second is mostly accepted (lumens of light). Then Step 5 shows that each of these is really a claim about what counts as the same thing, usually settled by what it is worth to people. The last is genuinely contested (bits of data), and ONS took a middle road on it. Each step adds one switch.

| Step | Question | One new switch | Core? |
|---|---|---|---|
| 0 | Money = stuff × price | Spending growth vs capacity growth (from the Wilkes toy) | Core |
| 1 | You can't add loaves to bulbs | Base year; fixed-base vs chain-linked | Core |
| 2 | What is one unit? (shrinkflation) | Per bar vs per gram | Core |
| 3 | The bulb that lasts longer | What is a unit of lighting: bulbs and kWh, linked-in bulbs, or lumen-hours | Core |
| 4 | What it does to GDP and inflation | Lighting lens, plus two world dials: how much more light people use when it's cheaper, and where bulbs are made | Core |
| 5 | Where does utility get in? (the laptop that got better) | What is a laptop worth: the box, its speed, or what the market charges for speed (hedonic) | Core |
| 6 | When does volume track welfare? | How the gain is measured: fixed-base, chain-linked, or the area under the demand curve | Should |
| 7 | Bits: telecoms 2010–2017 | Six published telecoms deflators | Core |
| 8 | Bits are mostly an input | Share of telecoms bought by firms; single vs double deflation | Core (owner decision), since without it Step 7 overstates the GDP effect |
| Coda | Computing, cloud and free goods | None (narrative) | Could |

### Step 0: Money = stuff × price

One good (bread). This recaps the Wilkes identity with the slider pair from the Quarto demo: spending growth and capacity growth, with inflation as the residual (assuming, as that toy does, that output tracks capacity). Point: with one good and an obvious unit, splitting nominal into real and prices is trivial. Everything that follows is about what happens when it isn't.

### Step 1: You can't add loaves to bulbs

Two goods with different price trends. Quantities can only be added once they are weighted by prices, and the weights you choose change the answer. This reuses `indexnumbers.js` and is the toy that `econ_measures_chainedvolume` set out to build:

- Switch the base year and watch real growth change.
- Chain-link (previous year's prices, the ONS method) and the base-year dependence goes away.
- The cost: components no longer add up to the total, except in the reference year and the year after it. (Chain-linking can also drift when prices bounce up and down; that's small for annual national accounts, and it's part of the RPI clothing story in §5.)
- The implied deflator is a residual (nominal ÷ real), never measured directly.
- ONS chain-links a Laspeyres index; the US Bureau of Economic Analysis uses a chained Fisher. Step 6 shows why the difference is small.

On desktop, a workbench table shows the previous-year's-prices arithmetic year by year. On phone it becomes a "show the sums" drawer for one year at a time.

### Step 2: What is one unit? (shrinkflation)

A 200 g chocolate bar becomes 180 g at the same price. Counted per bar, there is no inflation and real consumption is flat. Counted per gram, the price has risen 11.1% and real consumption has fallen 10%. Money is identical under both. Almost nobody argues that per gram is wrong, and UK consumer price statistics do adjust for pack size (the measuring-gva note says CPIH does; check the current ONS practice wording at build time). Even per gram is a claim, though: that the last 20 g are worth as much as the first 20 g (β = 1 in Step 5's terms). This is the warm-up: choosing a unit is ordinary and uncontroversial here, and the same move is about to become startling.

### Step 3: The bulb that lasts longer

**World (decade scale, roughly 2008–2020):** households buy lighting. Incandescent bulbs give way to halogen and CFL, then LED, a shift that in the UK was pushed by regulation (the EU incandescent phase-out from 2009, and the UK halogen ban in 2021; check the dates at build time). Each technology has a bulb price, light output in lumens, lifetime in hours and wattage, and electricity is bought to run them. The LED's price when it first goes on sale is a world dial: early LEDs were far dearer than today's (check UK prices around 2010).

**Lens switch: what is one unit of lighting?**

- **(a) Count the stuff:** bulbs per bulb, electricity per kWh. Bulbs bought fall by more than half, because LEDs last about 15 times longer. What happens to electricity depends on a world dial. If people take the whole saving as more light (elasticity 1, the default), electricity use barely changes; if not, it falls. So measured lighting volume ranges from about flat (−5% at the default) to a steep fall (about −60% at elasticity 0.5). Measured lighting *prices* rise slightly, because an LED costs more per bulb.
- **(b) Link the LED in at overlap:** as (a), but the LED enters the bulb index while old and new bulbs are both on sale, and the price gap at that moment is treated as quality. After that only the LED's own price changes count, so its later price fall is recorded; whatever it offered beyond its launch premium is not. This changes the bulb figures a lot and the lighting total very little, because bulbs are a small slice of the cost of light.
- **(c) Count the light:** lumen-hours delivered, with bulb and electricity cost together per million lumen-hours. The price of light falls about 6.5-fold. Volume rises about 6.5-fold at the default elasticity, and about 2.6-fold at elasticity 0.5.

**The lesson the three lenses share:** most of the LED's gain is electricity saved. Counted per kWh, electricity can't show that, so (a) and (b) miss most of the gain whatever they do with the bulbs. Lens (c) also dissolves the new-goods problem: under it the LED isn't a new good at all, just a cheaper way to buy an old service.

Nominal spending on lighting is identical under all three. Each lens shows lighting's real growth and lighting inflation, with the conservation check beside them and a breakdown into bulbs and electricity.

**Rough calibration, to be checked at build:** an incandescent bulb (60 W, about 800 lumens, 1,000 hours, £1) at 27p/kWh costs about £21 per million lumen-hours, nearly all of it electricity. An LED (9 W, about 800 lumens, 15,000 hours, £3) costs about £3.30. That is roughly a 6.5-fold fall at a constant electricity price. The calculation should be shown, not hidden.

**Electricity is in the cost of light by default** (owner decision; review once it's built). The price of light is mostly electricity, so leaving it out would mislead. A **bulbs-only view** switches electricity off and shows the bulb market alone, where the three lenses diverge most sharply: about −56% for (a), somewhere in between for (b) depending on the LED's launch price, and about 6.5-fold for (c). Counted per kWh, efficiency shows up as less electricity, or with full rebound as no change at all, even though households get more light. That is the energy-efficiency paradox, and it leads into the less extreme examples in §5.

**What the national accounts would actually do.** There is no "light" product in the national accounts. There are bulbs and there is electricity, each with its own price index. Lens (c) is the *household-production* view used by Nordhaus and by Fouquet & Pearson: households combine bulbs and electricity to make light.

- **Corrected while building M1 (see §15).** This plan used to say that a national-accounts version of lens (c) would match the bundle's total. That version credits each LED with the light *and* the electricity savings it delivers over its life, and keeps electricity in kWh. It only matches when valued at incandescent-era prices. Chain-linked at market prices it gives about +46% lighting volume over 2008–20, against +519% for the light itself, and the figure depends on how fast LEDs took over. Once LEDs are cheap, market prices value the electricity they save at pennies. So it is not robust enough to show.
- What a statistical office *can* do robustly is re-price the bulb per million lumen-hours of its own lifetime light, keeping electricity in kWh. As built, Step 4 calls this "Count each bulb's light". It catches only a sliver of the gain, because bulbs are well under a tenth of the cost of light and most of the LED's gain is electricity saved.
- That saving turns into more light inside homes. No industry sells light, so it sits outside GDP's production boundary. Only the household view (lens (c), the bundle) sees it.
- Statistical offices don't normally adjust bulb prices for running costs (check ONS practice for light bulbs), so official measures in practice sit closer to (a) or (b).

**Honesty notes:**

- Lumens are not everything. Colour, flicker, warm-up time, dimming and looks all matter, and that is why the toy states "people want lumens" as an assumption (§3).
- A hedonic regression on lumens, lifetime and watts would, in this toy, reduce to the national-accounts version of lens (c), because those are the only characteristics the toy has. Say so, rather than presenting hedonics as a fourth independent answer. Step 5 is where hedonics gets its own example.
- Real UK electricity prices rose a lot over this period, which partly offsets the efficiency gain. Make the electricity price a world dial.
- **Bulbs are durables.** A bulb is bought once and lights a room for years. The national accounts record it when it is bought; they impute a flow of services only for owner-occupied housing. So lens (c) re-prices each year's purchases, valuing each bulb for what it will deliver over its life, rather than counting the light used that year. That keeps nominal unchanged. The two agree in a steady state. During the switchover they don't: a household buys about fifteen years of light in one go, so measured volume runs ahead of the light used, then settles. The world needs a stock of bulbs with staggered replacement, so purchases are smooth rather than lumpy. Flag this as a structural choice (§9), and compare steady states, not single years, in Step 6.

### Step 4: What it does to GDP and inflation

The whole toy economy: bread plus light, with an optional nominal anchor of Wilkes's 4% a year. It starts *closed*, with everything made at home, and the "made here or abroad?" dial then opens it (owner decision).

**As built (M1):** total spending is held flat rather than growing 4%, so every change comes from light. The lens switch offers the three product-level ways a statistical office could count bulbs: count the stuff, link the LED in, and count each bulb's light. Electricity always stays in kWh. "Real consumption if light itself were a product" is shown alongside as a reference. Under that household view there is no GDP to measure. At the defaults over 2008–20:

- re-pricing bulbs by their light adds about 0.16 points to real consumption compared with counting stuff, and takes the same off consumer prices;
- the same 0.16 reaches real GDP when bulbs are made here, and none when they are made abroad;
- counting light itself gives about +1.5 points.

The bullets below are the original plan.

- **Big at product level, small at GDP level.** Lighting's weight in GDP is its share of spending, which is small. Tsao et al. (2010) estimate that about 0.72% of GDP has gone on light for three centuries, in rich and poor economies alike (from press summaries; confirm in the paper). If demand for light is inelastic, that share shrinks as light gets cheaper, and annual chain-linking updates the weight every year. A world dial, "how much more light do people use when it's cheaper?" (the price elasticity, or rebound), sets this. The default is 1, meaning spending on light stays constant, which is what Tsao et al.'s constant share implies.
- **Equal and opposite.** Switching to lens (c) raises real growth by x and lowers deflator growth by the same x.
- **Made here or abroad?** In the UK, bulbs are mostly imported, and laptops almost entirely so (check). If imports are re-measured the same way, real imports rise as much as real consumption, so real GDP and the GDP deflator barely move. Only the domestic parts count: shop margins, and the electricity. The gain shows up in real consumption, real incomes and consumer prices. If import prices *aren't* adjusted the same way, real GDP can move spuriously, which is a known problem with import prices. Telecoms services are produced in the UK, which is one reason that change did reach GDP. **The dial** (owner decision) sets the share of bulbs made abroad, from none (the closed toy) to all, with electricity staying home-made. Slide it and watch the lens's gain move out of real GDP and the GDP deflator and into real consumption and consumer prices, while the conservation check holds for each.
- **The Wilkes tie.** In the closed toy, the Bank's 4% is untouched but measured supply-side performance changes: part of "how much is real and how much inflation" is a statistical choice. With imports the tie shifts. Re-measuring an imported good moves consumer prices (the Bank's target) but hardly the GDP deflator (Wilkes's split). Re-measuring a UK-made product moves the GDP deflator.
- **Not papered over: which index each number is.** For consumer goods, national accounts deflators are largely built from CPI components. So a change in how CPI treats bulbs would flow into the consumption deflator *and* into the inflation the Bank targets. That is the one route by which a measurement choice changes the world as well as the lens: a mismeasured CPI misdirects everything indexed to it, from benefits to an inflation target. (The Boskin Commission estimated in 1996 that the US CPI overstated inflation by about 1.1 points a year, partly through missed quality change and new goods; check.) One sentence on the page, not a model.

### Step 5: Where does utility get in? (the laptop that got better)

Until now, choosing a unit looked like a physical question: grams, lumens. This step shows it is really a question about what counts as the same thing, and that the answer usually rests on what people value.

**The theory, briefly.** The "economic approach" to index numbers (Konüs 1924) defines the ideal consumer price index as the extra money someone would need to be exactly as well off as before. That is a cost-of-living index, and it is defined by utility. The Laspeyres index statisticians mostly use answers an easier question, the cost of last period's basket. On its own it ignores substitution, and it needs separate rules for quality change. Statistical offices differ on how far they aim at the utility ideal. The US Bureau of Labor Statistics treats a cost-of-living index as its conceptual goal. The European-style CPI the UK follows is framed as a fixed-basket price index. Check both offices' current wording before publishing.

**Two tests for "the same thing".** Triplett (1983) proposed a resolution to a long argument. For consumer prices and other *inputs*, judge quality by value to the user. For producer (*output*) prices, judge it by resource cost, meaning what it takes to make. So the utility idea sits behind consumer prices, while output deflators, like the one ONS changed for telecoms, officially rest on cost. Robert Gordon's reply was that with the right unit (one calculation rather than one computer) the two tests give the same answer. That is this toy's thesis from another angle. (Gordon's view is as summarised in a 2013 Voorburg Group paper; check the original.)

Abdirahman et al. (2020, WP58, p. 4) put the consumer side with a laptop that costs £1,000 in both 2012 and 2017, but where the 2017 model is far better: "to reflect real growth through innovation, the price ought to record a decline; there has been an increase in consumer surplus." The same page traces hedonic adjustment back to Adelman & Griliches (1961), who wanted to link prices "to a yardstick more nearly relevant to its intrinsic utility". Utility is in the method's founding statement.

**World:** each year several laptop models are on sale, at different speeds and prices. The typical laptop bought costs £1,000 every year, while speed doubles every two years. Every model's price sits on the market's price–speed curve, which shifts down over time: price rises with speed to the power β. β is a world dial, "how much more does a laptop with double the speed cost?" With β = 1 it costs twice as much; with β = 0, nothing extra. A little noise (fixed seed) makes the cross-section look like real data.

**Lens switch: what is a laptop worth?**

- **(a) Per box (unit value).** Price is flat, so there is no inflation and no real gain beyond the number of laptops sold. Speed is treated as worth nothing (β = 0). No statistical office does this for laptops; it is the "per bulb" and "per contract" lens in laptop form.
- **(b) Per unit of speed, the engineer's view.** The price halves every two years and real output doubles. Double the speed counts as double the stuff (β = 1). This is the lumen lens and the per-byte lens again.
- **(c) Hedonic.** Estimate β from the models on sale each year: regress log price on log speed with a dummy for each year (the time-dummy method). Then measure the price change holding speed constant. In the toy, the quality-adjusted price falls by a factor of 2 to the power β every two years. So when β is between 0 and 1, the hedonic index lands between (a) and (b), at the point set by what the market charges for speed.

**The punchline:** the three lenses are one formula with β = 0, β = 1, and β read from market prices. The unit arguments in Steps 2–3 are arguments about β: how value rises with the characteristic. Step 7 adds a second kind of argument: whether two services (a minute of talk, a megabyte of video) are the same thing at all. Hedonic methods let market prices choose β instead of the statistician, though the statistician still chooses which characteristics count and the shape of the curve.

On desktop, show the cross-section as a scatter: log price against log speed, one colour per year, with parallel fitted lines. The vertical gap between the lines is the quality-adjusted price change. On phone, the scatter works at full width, with the lens and the β dial underneath.

**Every lens is an assumption about what counts as the same thing.** Spell this out on the page:

- Per box assumes speed is worthless.
- Per unit of speed assumes value is proportional to speed.
- Per byte (Step 7) assumes a byte of a phone call is worth a byte of video.
- Revenue weights (ONS's telecoms choice) assume relative prices show what buyers value at the margin.

**Honesty notes:**

- **Hedonic prices are market prices.** Rosen (1974) showed that in competitive markets the slope of the price–characteristic curve equals, at the margin, both what buyers would pay and what sellers would need to be paid. With market power it reflects mark-ups too. Pakes (2003) argues the coefficients need not reflect preferences, though he still defends hedonic indices as a bound on the welfare change. In the toy, one world dial stands in for all of this, and the page says so. The dial is the market's price gradient, not any one buyer's willingness to pay for a doubling.
- **Which characteristics, which curve.** The toy has one characteristic, so hedonics is right by construction. Real hedonic work has to choose the characteristics (speed, memory, screen, battery, weight) and a functional form, and leaving one out biases the index.
- **Matched models.** What statistical offices mostly do is track the same model's price over time. In this toy's world, where every model sits on the price–speed curve, matched models give the same answer as hedonics. They come apart when quality arrives mainly through new models launched at the old price points while old models leave without being discounted. That is the case hedonics was built for.
- **Expectations move.** WP58 notes that a consumer might be "equally satisfied in 2012 and 2017, given what is available on the market and her (socially-influenced) expectations". If software gets heavier as hardware gets faster, the value of speed is partly relative, and a fixed-preference hedonic index overstates the gain.
- **Hedonic adjustment fixes the price, not the welfare question.** Whether real growth then tracks the gain to people is Step 6's question.
- **The same idea explains Step 1.** Superlative indices such as Fisher approximate the cost-of-living index (WP58, p. 4). That is why chain-linked and superlative methods are preferred to a fixed Laspeyres basket. A one-line cross-reference is enough.
- **Laptops are imported.** For the UK, re-measuring laptops moves consumer prices and real consumption, not real GDP (Step 4).

### Step 6: When does volume track welfare?

Same lighting world, one question: how much better off are people because light got cheaper? Compare steady states before and after the LED switchover, to keep the durables timing (Step 3) out of it. The toy knows the answer, because its world has a stated demand curve for light: the gain is the area under that curve between the old and new price of light (the change in consumer surplus). Then compare what each measure says.

Rough numbers at the defaults (light 0.7% of spending, the price of light down 6.5-fold, elasticity 1), to be reproduced by the engine:

| Measure | Gain in real consumption |
|---|---|
| Today's light valued at the old price (fixed-base Laspeyres) | about +3.9% |
| Old light valued at today's price (fixed-base Paasche) | about +0.6% |
| **The toy's true gain (area under the demand curve)** | **about +1.3%** |
| Chain-linked Laspeyres (the ONS method) | about +1.5% |
| Chain-linked Fisher (the US method) | about +1.3% |
| Lens (a): bulbs and kWh counted as stuff | about 0% |

**The finding, which runs against a common claim:** with the right unit and annual chain-linking, real growth tracks the welfare gain closely. Chain-linking values each year's extra light at prices from that year or the one before, and those are close to the right marginal value for that year's small change. The two fixed-base answers bracket the truth, which is the textbook Laspeyres–Paasche bound. It holds under the assumptions listed at the end of this step, and the page should state them.

**Where volume and welfare part company:**

- **The wrong unit.** Lens (a) records nothing while the toy's people are 1.3% better off. Over two centuries this is Nordhaus's (1996) finding: conventional price indices overstated the rise in the price of light by a factor of roughly 900 to 1,600 since the early 1800s, so real wages and output growth were understated. Show the Fouquet & Pearson UK series (via Our World in Data: a real fall of over 99.9% since 1700) on a log scale, from candles to LEDs.
- **Brand-new goods.** When something genuinely new appears, the gain between what people would have paid for it and its launch price is never recorded once it is linked in. Hausman's reservation-price approach tries to estimate it.
- **Free goods.** A zero price means zero weight (see the coda).
- **Total value.** GDP measures changes at the margin, not the total worth of light: Adam Smith's water, cheap but priceless. Under the default demand curve that total is infinite, so the page must never show it. A finite version needs a backstop, such as "what if we had to go back to incandescent bulbs?"
- **The assumptions behind the bound.** Preferences that scale with income (homothetic); people buying what suits them at the prices they face; a toy where all output is consumed. If light is a necessity, or price changes are very large, the link loosens.

### Step 7: Bits (telecoms, 2010–2017)

Real published deflators, with the same machinery as the lighting steps.

**The facts (Abdirahman et al. 2020, Hutchins WP58):** UK data usage grew by nearly 2,300% (about 24 times) between 2010 and 2017. Yet official real gross value added for the telecoms industry fell 8% between 2010 and 2016, and it recorded some of the slowest productivity growth of any industry.

**Lens switch: six published deflators** for telecoms services, 2010 = 100, from ESCoE DP 2020-11, Figure 5. These are price indices for telecoms *services* (products), used to deflate output; the GVA figures come after double deflation (Step 8).

| Deflator | 2010 | 2011 | 2012 | 2013 | 2014 | 2015 | 2016 | 2017 | Change 2010–17 |
|---|---|---|---|---|---|---|---|---|---|
| Current (pre-2021 official) | 100 | 100 | 100 | 100 | 100 | 98 | 102 | 103 | +3% |
| Improved SPPI (WP58 option A) | 100 | 95 | 77 | 76 | 68 | 67 | 65 | 63 | −37% |
| **Option A.1 (the method ONS adopted in Blue Book 2021)** | 100 | 91 | 71 | 67 | 55 | 51 | 46 | 42 | −58% |
| Option A.2 | 100 | 86 | 64 | 59 | 47 | 43 | 38 | 36 | −64% |
| Option A.3 | 100 | 84 | 58 | 48 | 35 | 25 | 20 | 16 | −84% |
| Data usage, price per byte (WP58 option B) | 100 | 56 | 38 | 27 | 14 | 8 | 6 | 4 | −96% |

ONS adopted the A.1 *method*. Its own published series, built with later data, may differ from the paper's A.1 line, so label the row accordingly.

**The dial:** from "economist" (separate services, revenue-weighted) to "engineer" (one product, bytes moved). The paper converts a voice minute to bytes at 480 kB per minute. That figure should be a visible, adjustable assumption, because it is what makes voice and data comparable at all.

- In Step 5's terms it is a claim about value: that a minute of talk is worth what 480 kB of anything else is worth.
- *Corrected while building:* 480 kB a minute is the paper's assumption of about 32 kbit/s each way, for a two-way call, not a phone-line standard. As built, the dial shows the rate barely matters. Voice is about 5% of 2010's bytes, so even 12 kbit/s each way only moves the 2017 index from 3.92 to 3.81. The −96% comes from treating *all* revenue, voice included, as paying for bytes, nearly all of which are broadband.

**What the step must make clear:**

- **ONS did not adopt price per bit.** Blue Book 2021 adopted option A.1. It adds broadband and mobile data to the services producer price index as unit values (revenue ÷ quantity) for each fairly uniform product (voice minutes, texts, terabytes of data), weights them by revenue, and spreads fixed-line access charges across services by revenue. So "per bit" applies *within* data, not across services. The −96% per-bit answer is the engineering bound, not the official one.
- **Why per bit gives −96%.** A price-per-byte index makes real output equal bytes moved, by definition: real = revenue ÷ (revenue ÷ bytes) = bytes. It treats a switch from expensive voice bytes to cheap video bytes as a price fall, which assumes a byte of a phone call is worth the same as a byte of video. The world-and-lens split makes this visible.
- **Nominal stays fixed.** Under each lens, telecoms real *output* is nominal ÷ deflator. With a flat illustrative revenue path (real data at M2, see §9), real output in 2017 ranges from about 3% *lower* to about 24 times higher, with nominal identical throughout.

**One more reading (arguable, and labelled as such on the page).** Triplett's rule says output deflators should judge quality by resource cost. Carrying a byte costs a network roughly the same whatever it carries, which arguably favours the engineer's per-byte view for an output deflator. ONS's revenue weights follow what buyers paid instead, and voice still sold at far more per byte than data. The authors expected prices per byte to converge as customers switch to the cheaper substitute, which "would make a data usage based unit value index a more meaningful proxy deflator" (WP58, p. 3).

**The ONS impact (to verify before use).** ONS's article says: "Average annual GVA growth for this industry was 27% over the period 1998 to 2018 in indicative BB21 estimates, compared with 6.8% in BB20."

- 27% a year for twenty years is a more than hundredfold rise, so check it against the Blue Book 2021 dataset before it goes on the page. If it holds, much of it is likely double deflation's amplification (Step 8), not the deflator alone.
- For the economy as a whole, Blue Book 2021 revised annual GDP volume growth over 2010–2019 from 1.8% to 2.0%, naming the telecoms deflator and the Financial Services Survey among the drivers. The telecoms share isn't separated out. (Check the period: the article covers 1997–2018.)

Huge in the industry, modest in GDP, for two reasons: telecoms is a small share of GDP (Step 4's lesson), and much of it is bought by other industries (Step 8).

### Step 8: Bits are mostly an input

A three-industry toy: telecoms sells to households *and* to other industries (finance, for example).

- **Switches:** the share of telecoms bought by firms vs households; single vs double deflation.
- **What you see:** under double deflation (ONS practice since Blue Book 2021), a faster telecoms price fall raises telecoms real GVA but *lowers* the real GVA of the industries that buy telecoms, because their real inputs rise. ONS's Blue Book 2021 article names finance and insurance, and gambling and betting, as the most affected. In the toy, at the previous year's prices, the economy-wide effect is exactly what reaches final users (households, exports, investment); the intermediate part cancels out. In ONS practice it is approximately so, through supply–use balancing (check how ONS deflates household telecoms spending after Blue Book 2021).
- **Double deflation amplifies.** Real GVA is a difference: deflated output minus deflated inputs. When output prices fall faster than input prices, real GVA can grow far faster than real output. With GVA at 40% of output, 13% real output growth and 3% input growth give roughly 28% GVA growth. That is a likely part of the 27% in Step 7, and a known weakness: where the margin is thin, double-deflated GVA is volatile.
- **Why it matters:** this explains "huge in the industry, modest in GDP" honestly, and shows that much of the reclassification's effect is to move measured productivity *between* industries. How much depends on the household share of telecoms (to source).

### Coda: computing, cloud and free goods (narrative, not modelled)

- The cost of computation has kept falling. Coyle & Hampton (2024) find the decline slowed after 2010 but not before, while the productivity slowdown they ask about began in the mid-2000s. They put any link to it in complementary inputs (skills, organisation), not in computing itself.
- Cloud computing price indices: Byrne, Corrado & Sichel; Coyle & Nguyen for the UK.
- Where the price to the user is zero (search, maps, social media), there is no expenditure weight at all. GDP records the advertising that pays for these services as a cost to advertisers, not their value to users. That is the end point of the ladder: the unit question can no longer be answered with a price. References: Brynjolfsson et al. (GDP-B); Coyle's *GDP: A Brief but Affectionate History* and *The Measure of Progress*.

## 5. Less extreme examples

The telecoms and two-centuries-of-light cases are deliberately extreme. These smaller ones teach the same mechanism at sizes people will find plausible, and several have UK policy consequences. Rows marked ★ are recommended as presets that reuse the Step 3–5 engine. Every number marked "verify" needs checking before it goes into the model.

| Example | The measurement question | Rough size | Why it's useful | Fits |
|---|---|---|---|---|
| ★ **Shrinkflation** (chocolate, crisps) | Per pack or per gram? | A few % on a single item | Almost undisputed; the warm-up | Step 2 |
| ★ **Energy efficiency** (fridges, boilers, insulation) | Energy bought, or warmth and cold delivered? | Efficiency gains of maybe 2–3× over decades (verify) | Lumens logic at a believable size. Counted per unit of energy, efficiency looks like less output, or none if people take it as more comfort | Step 3's bulbs-and-kWh lens, or its own preset |
| ★ **Laptops and TVs** (hedonic adjustment) | Price per box, or per characteristic (speed, screen size, resolution)? And what is the next unit of a characteristic worth? | Quality-adjusted prices fall steeply each year (verify the UK figure; check which IT items ONS adjusts hedonically) | Shows where utility gets into a deflator; the mainstream version of what Coyle and colleagues pushed further for telecoms. Mostly imported, so in the UK it moves consumer prices and real consumption more than GDP | Step 5 (laptops); TVs as a second preset |
| ★ **Public services output** (Atkinson Review 2005; Covid 2020) | Is health and education output measured as inputs, or as activity adjusted for quality? | The UK's 2020 GDP fall looked worse internationally partly because ONS measured school and health activity directly while others used inputs (verify the ONS comparison article) | Directly relevant to MONIAC2's government-spending theme | Own mini-step |
| **Medical treatment** (cataract surgery, heart attacks) | Price per procedure, or price per outcome? | Price per procedure rose while price per outcome fell (Shapiro, Shapiro & Wilcox; Cutler et al.; verify) | The "service unit" idea in health, with outcomes as the characteristic | Step 3-style preset |
| **Clothing and the RPI formula (2010)** | The same price quotes, combined with a different formula (Carli in RPI, Jevons in CPI) | Widened the RPI–CPI "formula effect" (verify, roughly 0.9 points a year) | A formula choice rather than a unit choice, with real money attached: student loans, index-linked gilts, rail fares | Side panel in Step 1 |
| **Financial services (FISIM)** | What does a bank sell? Interest margins count as output | Measured finance output *rose* in late 2008 as risk margins widened (Haldane 2010; verify) | No physical unit at all | Mention only |
| **Music and photos** | CDs and film to streaming and phone photos; price per song or photo heads to zero | Large | Stepping stone between bits and free goods | Coda |

**Parked (owner, 2026-10-29):** defence as an example. Its output is measured by its inputs, so measured productivity can't grow (verify current ONS treatment). That would bridge to the defence-vs-welfare model, but the link is left out for now.

## 6. Engine sketch

Following CLAUDE.md §7: pure functions, no DOM, all testable in the browser.

```
docs/models/stuff_money_value/
  build-plan.md          this file
  index.html             single scrolling page, one section per step (anchors per step)
  style.css
  lib/d3.v7.min.js       copied from moniac/lib (owner: a copy per model)
  engine/
    indexnumbers.js      from the Quarto demo, extended: Laspeyres, Paasche, Fisher,
                         fixed-base, chain-linked (previous year's prices), chained
                         Fisher, implied deflator, log-growth split
    hedonic.js           time-dummy hedonic regression (log price on log characteristic
                         plus year dummies); quality-adjusted index for a given or
                         estimated β
    welfare.js           area under a stated demand curve between two prices (changes
                         only, never levels)
    world.js             scenarios: goods and technologies with characteristics, and the
                         share of each made at home; lighting world generated from demand
                         for lumen-hours (elasticity dial, LED launch price, bulb stock
                         with staggered replacement); laptop cross-sections per year
                         (β dial, seeded noise)
    lens.js              unit choice; value of a characteristic (β = 0, 1, or estimated);
                         new-item linking (overlap vs characteristic); lens (c) as a
                         quality adjustment on bulbs, with electricity in kWh;
                         unit-value vs revenue-weighted; single vs double deflation
    measure.js           measure(world, lens) → nominal, real consumption, real GDP
                         (net of imports), consumption and GDP deflators, by product,
                         checks
    data/telecoms.js     the published deflator series (§4 Step 7), with source notes
    engine.test.js
  view/                  one small D3 module per step; shared chart helpers
  test-runner.html       as in moniac/
```

**Invariant tests:**

- Nominal is identical under every lens.
- log(real) + log(deflator) = log(nominal) in every year.
- Chain-linked equals fixed-base when relative prices don't move.
- Chain-linked components add up to the total in the reference year and the year after it.
- A unit-value index equals the Paasche index when the product mix doesn't change, and every index agrees when all products cost the same per unit.
- The price-per-byte lens gives real output proportional to bytes.
- With geometric means throughout, the time-dummy hedonic index with β fixed at 0 equals the change in the geometric-mean price per box, and with β = 1 the change in the geometric-mean price per unit of speed.
- The hedonic estimate recovers the world's β to within the seeded noise.
- Counting the light makes real lighting volume exactly the lumen-hours, and so does pricing bulbs per lifetime lumen-hour (for the bulbs alone). (This replaces an earlier test idea, "a bulb credited with its electricity savings matches the bundle", which turned out false; see §15.)
- With a product wholly imported, and imports re-measured the same way, changing its lens leaves real GDP unchanged.
- Step 6: for unit-elastic demand, the chained Fisher matches the area-under-the-curve gain to within rounding, chained Laspeyres sits slightly above, and fixed-base Laspeyres and Paasche bracket it.
- Under double deflation in a closed toy, at the previous year's prices, the sum of industry real GVA equals real final demand, link by link.

## 7. Desktop and phone

**My view: both can be first-class. I'm not pushing back.** The stepped ladder is already phone-shaped: each step has one chart, at most two controls and a one-sentence readout. Phones genuinely constrain only two things:

- **Comparing three to six lenses side by side.** On phone this becomes a lens picker plus one chart with overlaid lines. On desktop, small multiples.
- **Reading the arithmetic tables** (prices × quantities, chain links). On phone, a "show the sums" drawer for one year at a time. On desktop, a full workbench table.

Neither restricts the goal, so there's no need to ask people to use desktop. Same rules as the MONIAC build: no horizontal page scroll, touch targets of at least 44 px, `dvh` sizing. Tables scroll inside their own container if they must.

## 8. Tech and hosting

- **A standalone page in `docs/models/stuff_money_value/` is the product** (owner decision). Vanilla ES modules plus D3 v7, no build step, served by GitHub Pages from `docs/`, self-contained.
- **One page or one per step?** I recommend a single scrolling page. The steps share one engine and one narrative, it reads naturally on a phone, and anchors still allow linking to a step. The MONIAC build uses one page per stage because each stage is a different machine. Here each step is the same world seen through a different lens.
- **Side benefit: reusable in blog posts.** The Quarto research ([quarto-live-models-in-blog-posts.md](../../../claude_research/quarto-live-models-in-blog-posts.md)) showed that OJS cells can import plain ES modules. So the same `engine/` could later drive small widgets in coveredinbees posts, for example Step 3 alone. Not a goal for this build.
- **Testing.** Verify in Safari via `python3 -m http.server` from the repo root.

## 9. Calibration and data

- **Lighting:** illustrative, anchored to typical bulb specifications, a UK electricity price and an early-LED launch price (verify at build). Lighting's share of spending defaults to about 0.7% (Tsao et al. 2010). Show the per-million-lumen-hour calculation in the inspector.
- **Lighting history:** the Fouquet & Pearson series as republished by Our World in Data (CC BY). Vendor a small CSV, with attribution.
- **Laptops:** illustrative throughout: a £1,000 typical price, speed doubling every two years, β as a world dial. No real data is needed, because the step is about method, not the size of an effect. If a real anchor is wanted later, look for a published ONS or BLS hedonic index for computers.
- **Telecoms deflators:** the six series in §4 Step 7, transcribed from ESCoE DP 2020-11 Figure 5.
- **Telecoms nominal output or revenue, 2010–2017:** a flat illustrative path first, labelled as such. Real data (ONS industry output, or Ofcom's Communications Market Report) comes in at M2 (owner decision). Calibration here changes a displayed value, not the model's structure.
- **ONS Blue Book 2021 impacts:** telecoms GVA growth of 27% vs 6.8% a year (1998–2018), and GDP growth revised from 1.8% to 2.0%. Both are quoted from the ONS article, and both need checking against the dataset before use (Step 7).
- **Structural flags (per CLAUDE.md §8):**
  - Generating the lighting world from demand for lumens makes lens (c) right by construction (§3).
  - Lens (c) is the household-production view, outside GDP. The most a statistical office can robustly do is re-price bulbs by their lifetime light, which catches only a sliver of the gain (Steps 3–4, §15).
  - Bulbs are recorded when bought, as the national accounts do, not as a flow of light over their life (Step 3).
  - The default elasticity of 1 (full rebound) decides whether counting stuff shows lighting flat or collapsing (Steps 3–4).
  - The toy starts closed and the "made here or abroad?" dial opens it; in the UK, bulbs and laptops are mostly imported (Step 4).
  - In the laptop world, one β stands in for both what buyers value and what extra speed costs to make (Step 5).
  - The 480 kB-per-minute voice conversion is what makes the per-byte lens possible at all (Step 7).
  - All of these need to be visible in the model, not buried in code.

## 10. Milestones

Small and iterable, as in CLAUDE.md §3.4. **M1 is the natural stopping point** to decide whether the rest is worth building.

1. **M0:** engine plus invariant tests; Steps 0–1 ported from the Quarto demo.
2. **M1:** Steps 2–4, which make the lumens argument end to end, from chocolate to GDP, including the "made here or abroad?" dial. Owner review in Safari at desktop and phone widths, including how the default electricity setting looks.
3. **M2:** Steps 5 and 7. These are the utility step and telecoms, which share the message that a unit is a claim about value. Real telecoms revenue data comes in here.
4. **M3:** Steps 6 and 8, welfare and double deflation. Step 8 has to be in place before a first full release.
5. **M4:** the coda, plus the starred presets from §5.

## 11. Risks and guards

| Risk | Guard |
|---|---|
| Sliding from volume to welfare without saying so | Separate layers (§2); Step 6 shows when they meet and where they part |
| Teaching that GDP misses all gains to consumers | Step 6: with the right unit and chain-linking, real growth tracks the welfare gain; the gaps are wrong units, new goods, free goods and total value |
| Claiming every deflator rests on utility | Step 5: value to the user for consumer prices, resource cost for output prices (Triplett) |
| Presenting hedonic as *the* measure of utility | Step 5 honesty notes: costs and mark-ups as well as preferences; the statistician's choice of characteristics |
| GDP claims for goods the UK imports | Step 4's "made here or abroad?" dial |
| Presenting per bit as what ONS did | Step 7 says plainly that ONS adopted A.1 and that per bit is the engineering bound |
| Leading with the unverified 27% | Kept out of §1; flagged in Step 7, with double deflation's amplification explained in Step 8 |
| Overstating the GDP effect of telecoms | Step 8; label Step 7's GDP figure "if all telecoms were final consumption" until Step 8 exists |
| Conflating the GDP deflator with CPI | Every inflation number is labelled with its index (Step 4) |
| The toy's world quietly assuming the answer | The §3 honesty note shown in the page, not just in this plan |
| Scope creep (cloud, free goods, the full supply–use framework) | Coda stays narrative; Step 8 stays three industries |
| Unverified numbers in a published page | Nothing marked "verify" or "check" goes into the page until it has been checked |

## 12. Decisions (owner, 2026-10-29)

| # | Question | Decision |
|---|---|---|
| 1 | What does "value" mean in the title? | How utility gets into the measures: deflators carry a utility conception in the background. Added Step 5 (hedonic laptops) and reframed §2 |
| 2 | Telecoms nominal data | Illustrative first, real data at M2 |
| 3 | Standalone page or blog posts? | Standalone page in `docs/` |
| 4 | Double deflation in the first full release? | Yes |
| 5 | Electricity in the cost of light? | On by default; review once it's built |
| 6 | Link to the defence-vs-welfare model? | Not for now (parked in §5) |
| 7 | Shared D3 or a copy per model? | A copy per model |
| 8 | Make the toy open? (raised by the sweep) | Yes: a "made here or abroad?" world dial in Step 4, from M1 |

No open questions remain. New ones go here as they come up.

## 13. Conceptual sweep (2026-10-29)

A max-effort pass over the economics. What changed, most important first:

1. **Step 6 had its main claim backwards.** It said volume isn't welfare because GDP values extra light at its tiny new price. For *growth* that's wrong. With the right unit and chain-linking, real growth tracks the welfare gain closely: at the toy's defaults the true gain is about 1.3%, chained Laspeyres gives about 1.5%, chained Fisher about 1.3%, and the fixed-base answers bracket it at 0.6% and 3.9%. The real gaps are wrong units, new goods, free goods and total value. Step 6 is rewritten and §2 reworded.
2. **The toy is closed; the UK isn't.** Re-measuring an imported good (bulbs, laptops) moves real consumption and consumer prices, not real GDP or the GDP deflator. Telecoms is UK-made, which is one reason it reached GDP. Added to Steps 4 and 5 and §5, and resolved as Q8 with a "made here or abroad?" dial in Step 4.
3. **The lighting outcomes were wrong once electricity is included.** Counting bulbs and kWh doesn't make lighting "collapse". At full rebound (the default, supported by Tsao et al.'s constant 0.72% share), electricity use is flat and lighting volume falls about 5%; it collapses only if people don't take the saving as more light. Linking the LED in only changes the bulb slice. Most of the LED's gain is electricity saved, which no bulb-count lens can see.
4. **The national accounts have no "light" product.** Lens (c) is the household-production view. The national-accounts-consistent version credits the bulb with the electricity it saves, which decides which product gets the growth.
5. **Utility isn't behind every deflator.** Value to the user is the textbook test for consumer prices; resource cost is the test for output prices (Triplett 1983), and the telecoms deflator is an output price. "Economic theory says" became "the economic approach", with a note that statistical offices differ. Step 5 and §1 are reworded, and an arguable Triplett reading is added to Step 7.
6. **The 27% figure** came out of §1 and is flagged for checking. A hundredfold rise in twenty years is probably inflated by double deflation, which Step 8 now explains.
7. **Smaller fixes:**
   - Chain-linked components also add up in the year after the reference year.
   - Matched models vs hedonics, and the choice of characteristics and functional form.
   - Pakes's position is now stated in full, and the β dial is relabelled as a price gradient.
   - ONS adopted the A.1 *method*; its own published series may differ from the paper's.
   - 480 kB a minute is the old 64 kbit/s landline rate.
   - Step 7 is real output, with GVA in Step 8.
   - Free services are recorded through the advertising that pays for them.
   - The coda's Coyle & Hampton wording.
   - The CPI route by which measurement could change what the Bank does.
   - Test specifications: geometric means; unit value vs Paasche; previous-year prices for double deflation.

## 14. Sources

**Telecoms and ICT deflators (checked 2026-10-29)**

- Abdirahman, Coyle, Heys & Stewart (March 2020), ["A Comparison of Deflators for Telecommunications Services Output"](https://www.brookings.edu/wp-content/uploads/2020/03/WP58_Abdirahman-et-al.pdf), Hutchins Center Working Paper 58. Source of the +2,300% data usage, real GVA −8% (2010–16), and +3% / −37% / −96% figures. Page 3 has the convergence argument used in Step 7. Page 4 is the source of the laptop example, the cost-of-living framing and the Adelman & Griliches "intrinsic utility" quote used in Step 5. A local copy is in `~/Code/claude/econ_measures_chainedvolume/sources/`.
- Abdirahman, Coyle, Heys & Stewart (July 2020), ["Telecoms Deflators: A Story of Volume and Revenue Weights"](https://escoe-website.s3.amazonaws.com/wp-content/uploads/2020/07/16111642/ESCoE-DP-2020-11.pdf), ESCoE DP 2020-11. Source of options A.1–A.3 (−58% to −84%), the Figure 5 series and the 480 kB/minute voice conversion.
- ONS, ["Double deflation methods and deflator improvements to UK national accounts: Blue Book 2021"](https://www.ons.gov.uk/economy/nationalaccounts/uksectoraccounts/methodologies/doubledeflationmethodsanddeflatorimprovementstouknationalaccountsbluebook2021). Says ONS adopted option A.1: unit values for voice minutes, texts and terabytes, with revenue weights.
- ONS, ["Impact of double deflation on industry chain volume measure annual estimates, 1997 to 2018: Blue Book 2021"](https://www.ons.gov.uk/economy/nationalaccounts/uksectoraccounts/articles/impactofdoubledeflationonindustrychainvolumemeasureannualestimates1997to2018/bluebook2021). Source of 27% vs 6.8%, the effect on user industries, and the GDP growth revision.
- Coyle & Hampton (2024; online 2023), "21st century progress in computing", *Telecommunications Policy* 48(1). Abstract checked via a secondary listing.
- Byrne, Corrado & Sichel, ["The Rise of Cloud Computing: Minding Your P's, Q's and K's"](https://ideas.repec.org/p/nbr/nberwo/25188.html), NBER w25188.
- Coyle & Nguyen (2019), ["Cloud computing, cross-border data flows and new challenges for measurement in economics"](https://econpapers.repec.org/RePEc:cup:nierev:v:249:y:2019:i::p:r30-r38_12), *National Institute Economic Review* 249.

**Light**

- Nordhaus (1996), ["Do Real-Output and Real-Wage Measures Capture Reality? The History of Lighting Suggests Not"](https://ideas.repec.org/h/nbr/nberch/6064.html), in Bresnahan & Gordon (eds), *The Economics of New Goods*. Source of the 900–1,600 factor (checked via secondary summaries; confirm in the original).
- Fouquet & Pearson (2006), "Seven Centuries of Energy Services: The Price and Use of Light in the United Kingdom (1300–2000)", *The Energy Journal* 27(1). Data via [Our World in Data](https://ourworldindata.org/grapher/the-price-for-lighting-per-million-lumen-hours-in-the-uk-in-british-pound).
- Tsao, Saunders, Creighton, Coltrin & Simmons (2010), "Solid-state lighting: an energy-economics perspective", *Journal of Physics D: Applied Physics* 43(35), doi:10.1088/0022-3727/43/35/354001. Source of the constant 0.72% share of GDP spent on light (checked via press summaries; confirm in the paper).

**Quality, hedonics and the cost-of-living index**

- Triplett (1983), ["Concepts of Quality in Input and Output Price Measures: A Resolution of the User-Value Resource-Cost Debate"](https://www.nber.org/chapters/c7787.pdf), in Foss (ed.), *The U.S. National Income and Product Accounts: Selected Topics*, NBER. User value for input measures, resource cost for output measures (NBER lists it as 1982).
- Konüs (1924, English translation 1939 in *Econometrica*), "The problem of the true index of the cost of living". Check before quoting.
- Adelman & Griliches (1961), "On an index of quality change", *Journal of the American Statistical Association*. Quoted via WP58.
- Rosen (1974), "Hedonic prices and implicit markets: product differentiation in pure competition", *Journal of Political Economy* 82(1). Check before quoting.
- Pakes (2003), "A reconsideration of hedonic price indexes with an application to PC's", *American Economic Review* 93(5). Check before quoting, including his defence of hedonic indices as bounds.
- Triplett (2006), *Handbook on Hedonic Indexes and Quality Adjustments in Price Indexes*, OECD. The practitioner's reference.

**National accounts method (ONS)**

- [Chain-linking methods used within the UK national accounts](https://www.ons.gov.uk/economy/nationalaccounts/uksectoraccounts/methodologies/chainlinkingmethodsusedwithintheuknationalaccounts)
- [Deflators and how we use them in economic estimates](https://www.ons.gov.uk/economy/inflationandpriceindices/methodologies/deflatorsandhowweusethemineconomicestimates)
- [Double deflation and the supply and use framework in the UK national accounts](https://www.ons.gov.uk/economy/nationalaccounts/supplyandusetables/articles/doubledeflationandthesupplyanduseframeworkintheuknationalaccounts/2022-12-02) (2022)
- ONS consumer prices indices technical guidance (replaced the 2019 technical manual in March 2026). Check its wording on what CPI aims to measure (Step 5).

**To find and check before use:**

- For Steps 4–6: Diewert (1976) on superlative indices; Hausman (1997, 1999) on valuing new goods; the Boskin Commission (1996); Gordon's response to Triplett; the BLS and Eurostat statements of what their consumer price indices aim at; ONS practice for light bulbs and for import prices.
- For the §5 examples: the Atkinson Review (2005); ONS's international comparison of non-market output during Covid; ONS hedonic adjustment in CPI; the RPI formula effect after 2010 (UKSA 2013, Johnson Review 2015); Haldane (2010) "The contribution of the financial sector: miracle or mirage?"; Shapiro, Shapiro & Wilcox on cataract surgery; Cutler et al. on heart attacks; Brynjolfsson, Collis & Eggers (2019) on free goods.

## 15. Build log

### 2026-10-29: M0 and M1 (Steps 0–4)

**What exists:**

- `index.html`: one scrolling page with Steps 0–4 and a "coming next" list.
- `engine/indexnumbers.js`: fixed-base, chained Laspeyres, Paasche and Fisher volumes; overlap linking for new items; implied deflators.
- `engine/world.js`: the chocolate and lighting worlds.
- `engine/lens.js`: the ways of counting, plus the whole economy with imports.
- `engine/engine.test.js`: 15 tests, all passing in Node and in Safari (`test-runner.html`).
- `view/`: one module per step, a shared D3 line chart, a bar chart and an HTML change-bar chart, plus the controls.
- Checked in Safari at phone (390 px) and desktop widths, in dark and light themes. Every lens and the import slider were driven through by script, with every conservation check passing.

**Testing hooks:** `?check` puts "READY" or "ERROR" in the page title, `?open` opens every drawer, and `?theme=light|dark` overrides the system theme. Anchors into drawers (e.g. `#step1-sums`) open and scroll once the models have loaded. The local server must not cache, or Safari serves stale modules.

**Simplifications, also stated on the page:**

- Two bulbs only, with no halogen or CFL.
- Bulbs are bought at the steady replacement rate, so the durables timing of Step 3 is described but not modelled.
- The electricity price is one level for the whole period, set by a dial, not a rising path.
- Total spending is held flat in Step 4.
- Owner decision 5 (electricity in the cost of light) is built as a bulbs-vs-electricity breakdown table under Step 3's chart, rather than a "bulbs only" toggle.

**Finding that changed the plan.** Crediting each LED with the electricity it saves does *not* reproduce "count the light" once chain-linked at market prices. Lighting volume comes out at +46% against +519% over 2008–20, and the result depends on how fast LEDs spread. Corrected in Step 3, Step 4, §6 and §9.

The robust consequence, now Step 4's main lesson, is that a statistical office can only re-price products. Re-pricing bulbs by their lifetime light catches about a tenth of the gain in real consumption: +0.16 points against +1.5 for light itself. The rest is household production, outside GDP. This sharpens the contrast with telecoms (Step 7), which is a service UK firms sell, so re-measuring it could reach GDP.

**Still to check before publishing** (all already flagged above):

- the 0.72% lighting share (Tsao et al., from press summaries);
- bulb prices and specifications;
- ONS pack-size practice;
- that UK bulbs are largely imported.

**Next:** owner review of M1. Then M2: Step 5 (hedonic laptops) and Step 7 (telecoms deflators), with real telecoms revenue data.

### 2026-10-29: M2, M3 and the coda (Steps 5–8)

Owner said "looking good, tweaks later, continue". Built:

- **Step 5:** `engine/hedonic.js` (the time-dummy regression, with β fixed at 0, fixed at 1, or estimated) and the laptop world. Each year's noise is centred, so per box is exactly flat and the hedonic index follows exactly from the estimated β.
- **Step 7:** `engine/data/telecoms.js` holds the six published deflators plus Ofcom revenue and data volumes from WP58 Annex A. That means real revenue is in from the start (£40.5bn in 2010, £38.1bn in 2017), so decision 2's "illustrative first" step was skipped.
- **Step 6:** `engine/welfare.js`, with the history chart from `engine/data/light-history.js` (Fouquet & Pearson via OWID, CC BY, decade by decade).
- **Step 8:** `engine/industries.js`, single vs double deflation.
- **Coda:** narrative text.
- **Tests:** 24, all passing in Node and Safari.
- **Checks:** every control on Steps 5–8 driven by script in Safari, with no errors and every check passing; screenshots at phone and desktop widths.

**Findings and decisions:**

- **Price per byte recomputed.** It matches the published series to rounding. The page uses the recomputed index (2017 = 3.92, not the rounded 4), so that real output under it equals bytes moved exactly. The voice conversion rate barely matters (see the corrected note in Step 7).
- **Step 8 resized.** It was first built with "finance" (£100bn) as the buyer, which made telecoms about a quarter of GDP and gave absurd GDP effects. It is now "other industries" at £2,000bn of output, so telecoms is about 1.5% of GDP. At defaults over 2010–17:
  - the old deflator gives telecoms real value added −9%, close to WP58's −8% for 2010–16;
  - A.1 with double deflation gives telecoms +274%, other industries −1% and GDP +1.1%;
  - A.1 with single deflation gives GDP +1.9%.
- **Step 6 numbers.** True gain +1.31%; chained Laspeyres +1.50%, chained Fisher +1.34%; fixed-base +0.61% to +3.74%; counting stuff −0.05%. This confirms the sweep's corrected claim.
- **Calibration question for the owner.** In Fouquet & Pearson's series the UK price of light falls only about 23% over 2008–2020, against −84% in the toy. Their series covers all UK lighting, much of it already efficient tubes in 2008, and electricity got dearer. The toy is the household bulb swap at a fixed electricity price. This is stated on the page (Step 6, "a history check on the toy"). The defaults could be made more conservative: slower LED adoption, CFLs as an intermediate step, a rising electricity price.
- **Kept off the page until checked**, per §11: ONS's 27% and 1.8→2.0% figures (described qualitatively), Nordhaus's 900–1,600 factor (described qualitatively), Gordon's reply to Triplett, and the BLS vs European framing of CPI. The 0.72% lighting share stays as a labelled toy setting.
- **Step 8's household share** defaults to 50%, illustrative. WP58 Annex A has retail at about 70% of revenue, but retail includes businesses.

**Next (M4):** the §5 presets, plus the owner's tweaks from review.
