Measured from the public record alone, one chain at a time: what got researched, what got funded, what got tested, what got regulated, and what still stands in the way. Four chains are published so far. The model layer has not run; every barrier here was read by hand.
§ 01
The claim
Raven measures how far an idea has moved toward a product, from the public record alone. A chain is a root, the thing that was there first, and the products the record can be checked for. Every number on this site sits beside its denominator and the rule that produced it; that pair is what the site calls a receipt.
Three layers, each measured by a stated rule. A trajectory is a count per year from one search, printed beside the whole index so growth is never mistaken for the index growing. A dependency is the share of one literature's references that land in another's, measured both ways, with a negative control beside it. A barrier is a sentence in a document, hand-checked and typed by whether it can lift: a physical barrier is permanent, an economic one is a price, a political one is a vote away. Every word on a chain's page is computed from that chain's atlas; the drafts that were once written about these chains are kept unpublished, in docs/authored-drafts.md.
Each product carries a score: growth positioned between two controls fixed before any chain was run, dependency, precedence, stages, and a barrier penalty, every component printed with the sentence behind it. It is an evidence score over one chain, anchored to two controls. It is not a probability, and nothing on this site calls it one.
Chains are added one at a time, each with its own hand-checked truth, and published only when the record behind it is complete.
§ 02
The two controls
Every score is anchored to two searches fixed before any chain was run. The smoke detector literature, a field that matured long ago, sits at 0; the large language model literature, the boom of the decade, sits at 1. A product's growth is placed between them after the growth of the whole index is divided out, so 0 means "grew as the smoke detector literature did", not "flat". A product that grew less than that is held at 0, which this site calls the floor; one that outgrew the large language model literature would be held at 1, the ceiling.
From 2000 to 2025, research on large language model grew about 16,000-fold, smoke detector about 3.0-fold, and all published work about 2.6-fold. A product's growth is placed on that range: 0 grew like smoke detector, 1 like large language model.
The chart, the exact multiples and the formula
The multiple beside each is the sum of its last 5 years plus one over the sum of its first 5 years plus one, and "works" counts every work the search returned in the year band.
Fig. 1 · the two controls and the index
Three series on one log axis, 2000–2025; each anchor's multiple is printed below.
Not a ranking: rows follow the chain, never a number. An evidence score over one chain, anchored to two controls; not a calibrated probability. Weak is below 0.40, moderate from 0.40, strong from 0.70. Growth, dependency, precedence and stages are averaged with equal weight, leaving out any the window could not settle; the barrier penalty is then taken off, and the result held between 0 and 1.
Humanoid robots · social robots, care robots, teleoperated robots
Fiction crossed a tenth of its peak (168 works of fiction in 2015) in 1981, research a tenth of its (1,917 works in 2025) in 2002: fiction led by 21 years.
Fig. 3 · Humanoid robots and their downstream products, product by product
Not a ranking: rows follow the chain, never a number. An evidence score over one chain, anchored to two controls; not a calibrated probability. Weak is below 0.40, moderate from 0.40, strong from 0.70. Growth, dependency, precedence and stages are averaged with equal weight, leaving out any the window could not settle; the barrier penalty is then taken off, and the result held between 0 and 1.
Fiction crossed a tenth of its peak (20 works of fiction in 2014) in 1970, research a tenth of its (4,127 works in 2025) in 2002: fiction led by 32 years.
Fig. 4 · Quantum entanglement and its downstream products, product by product
Not a ranking: rows follow the chain, never a number. An evidence score over one chain, anchored to two controls; not a calibrated probability. Weak is below 0.40, moderate from 0.40, strong from 0.70. Growth, dependency, precedence and stages are averaged with equal weight, leaving out any the window could not settle; the barrier penalty is then taken off, and the result held between 0 and 1.
Fiction crossed a tenth of its peak (43 works of fiction in 2017) in 1986, research a tenth of its (3,584 works in 2025) in 2014: fiction led by 28 years.
Fig. 5 · Wearable computing and its downstream products, product by product
Not a ranking: rows follow the chain, never a number. An evidence score over one chain, anchored to two controls; not a calibrated probability. Weak is below 0.40, moderate from 0.40, strong from 0.70. Growth, dependency, precedence and stages are averaged with equal weight, leaving out any the window could not settle; the barrier penalty is then taken off, and the result held between 0 and 1.
§ 04
Growth over time
Each line is how many times more works a literature drew than at the start of the window; the smoke detector and large language model lines are the two yardsticks every score is read between.
Fig. 6 · each literature's growth, window by window
Each point is how many times more works a literature drew in a 5-year window than in 2000–2004, both plus one, on a log scale. Where a product's line ends is the growth its score uses; the ruler reads that end as the growth component, 0 at the smoke detector literature's 3.0x and 1 at the large language model literature's 16,491x. A filled end means the product's barriers were read, a hollow one that they were not. 2 products sit at the floor, having grown less than the smoke detector literature did (3.0x). * No works in 2000–2004, so the line starts from the smoothing alone: Active debris removal (Orbital debris and its downstream products); Full-dive (Wearable computing and its downstream products).
§ 05
The comparison
A product whose barriers were never read takes a penalty of zero, which is the same number a clean sample takes, so these two groups are never ranked against each other.
Every product of every published chain · 23 products across 4 chains
The kinds are what the hand-checked sample found, not a census. Not a ranking: rows follow the chain, never a number. An evidence score over one chain, anchored to two controls; not a calibrated probability.
§ 06
Limitations of the comparison
Only growth shares a scale across chains, because every chain is anchored to the same two controls.
Every other component measures a product against its own root, so two products' dependency or precedence numbers say nothing about each other.
The score is an evidence score over one chain, and this table is not a ranking; it is ordered by chain and never by any number.
Where a product has an arrival year, the page prints what the research record did after it: held at 80% or more of its post-arrival peak, faded down to 40%, refused below that, on a 5-year trailing mean. A dated edge is read as crossed in the first year one citing work in 10 touches the cited literature, counted over a 5-year window; an edge that crosses inside the band's first 4 years is left undetermined. No edge measured so far has reached the tenth; each verdict prints the peak reach it was read against. The after-arrival word reads the research literature, which rarely shrinks; the index and the mature control both read held. Neither reading enters a score.