What random-stimulus generation actually yields

Measurement of the method, not of the ideas. 25,000 leads were generated by eight exploratory methods including random-word stimulus. This page measures the yield of that generation: duplication, diversity, and how much of the output a human filter would keep. The honest answer is that most of it is unusable, and the interesting number is how much is not.

1. Generation statistics

Measure	Value
Leads generated	25,000
Unique claim sentences	24,935
Duplicate fraction	0.3%
Unique (domain, method, noun, verb, artifact) tuples	24,984
Leads per domain	625 (exactly)

Duplication is not the problem. With 40 × 625 draws over a vocabulary of 120 nouns, 40 verbs and 20 failure modes, only 0.3% of sentences collide. The generator produces variety cheaply.

2. Method distribution

Method	Leads
failure-first	3,272
biological analogy	3,208
cross-domain transfer	3,187
extremal case	3,120
scale shift	3,094
inversion	3,052
random stimulus	3,042
constraint removal	3,025

3. The yield problem, stated plainly

Variety is not value. Three filters apply, and only the first can be automated:

Filter	What it removes	How many survive
Well-formed	grammar and vocabulary collisions	24,935 (99.7%)
Mechanistically nameable	claims with no mechanism that could be real	not automatable
Testable from an existing record	claims needing an apparatus nobody has	not automatable

The first filter passes almost everything, which is exactly what makes generated ideas dangerous: they look complete. The second and third require a reader who knows the domain, and there is no shortcut.

4. What the method is actually good for

Three things, and only three:

Naming a mechanism nobody had phrased that way. "A sponge that absorbs suggests treating an offset as storage-and-release" is either a real reframing or nothing, and the reader knows in ten seconds.
Breaking a template. When a domain is stuck, an unrelated noun forces a mapping that the domain's own vocabulary forbids.
Producing a large discardable set cheaply. The value is in the discard rate, not the yield: if 1 in 500 survives, 25,000 generates fifty.

Expected yield, stated in advance: if 0.2% survive both human filters, that is 50 usable leads from 25,000 — and the method has paid for itself, because a good lead is worth more than the afternoon it cost to generate 25,000.

5. What would falsify this assessment

A blind test where a domain expert keeps more than 5% of generated leads as worth pursuing.
Evidence that generated leads are more novel per unit of expert time than deliberately written ones.
A demonstration that a generated lead led to a measurement that a deliberate one would not have.

None of these has been run. The claim that generation is cheap and filtering is expensive is an expectation, not a result.

6. Comparison across the four projects

Project	Leads	Generation method	Honest yield
1 · measurement critique	1,126	hand-written by agents, domain knowledge	~25 with numbers
2 · practical traps	20,000	40×25×20 composition	4 traps tested
3 · industrial audits	25,000	50×25×20 composition	5 computed, 1 failed
4 · random stimulus	25,000	8 methods incl. random words	unmeasured; expected ~50

7. Limits

The filters are not automated. The 50-lead estimate is an assumption, not a count.
The vocabulary is mine. A different noun list would produce different leads, not better ones.
Generation is not thinking. Every sentence here was produced by a template; none has been considered by a specialist.

Method · Top 40