Archetype substrate analysis¶
Part of Inverse Innovation with the Encyclopedia of Abstractions · Archetype substrate analysis · Last revised August 2026
Status: complete, explicitly post hoc
Overall verdict: SUPPORTIVE_OF_TWO_LEVEL_SUBSTRATE_CONSTRAINT
Plain-language result¶
The earlier claim that the inverse-innovation experiments were governance-heavy because the archetype catalog was built from the wrong kinds of source material was not supported. Physical and life-science primes are represented among archetype source links at almost exactly their corpus availability. A more precise effect appears later in the system.
The archetype still matters: institutionally framed archetypes tend to produce governance/process interventions, while more structural archetypes tend to produce computational/information interventions. But changing the archetype does not usually move the proposal outside those two families. Across the larger balanced E9 follow-up, 139 of 150 proposals (92.7%) used one of those two primary causal substrates.
That is the two-level finding:
- source modulation: archetype character shifts proposals between governance and computation;
- pipeline constraint: the tested generation-and-screening bundle remains strongly concentrated in governance and computation across archetype types.
This is a finding about the entire tested bundle—the encyclopedia snapshot, archetype text, prompt, model, rubric, and protocol—not proof of an intrinsic limitation in the base language model.
Correction to the motivating note¶
The original internal analysis omitted Deadweight Loss Reduction (score −0.50) because it inferred the experimental archetype list from survivor dossiers. The corrected list contains all ten archetypes and has an unweighted mean structural/framed score of +0.366, close to the corpus reference of +0.374, rather than the originally reported +0.462. The corrected internal note is preserved in ARCHETYPE_SUBSTRATE_REALIZATION_V2_CORRECTED.md.
E10B: existing 422-proposal classifications¶
E10B rejoined the already blinded and adjudicated Experiment 10 substrate labels to the corrected archetype scores. Its primary complete-census universe contained all 340 version-zero proposals from Experiments 5 and 6; the ten archetypes were the inferential units.
- Source score versus governance share within the governance/computational channels: Spearman ρ = −0.666, exact permutation
p = 0.0422. - Leave-one-archetype-out range: −0.822 to −0.601.
- Source score versus escape from those two dominant channels: ρ = −0.058,
p = 0.85. - Governance plus computational share: 95.5% framed, 100% middle, and 89.7% structural.
This was an exploratory post-hoc diagnosis. An initial cross-tab was seen before the formal analysis freeze, and that fact is recorded in the design freeze. The complete E10B report, method, and machine-readable results are available.
E9 follow-up: 50 archetypes × 3 fixed domains¶
The E9 follow-up supplied the stronger test. It retained all 150 E9 proposals regardless of screen outcome and removed archetype, domain, sampling-stratum, score, prior-art, and screen fields before classification. Two fresh passes independently assigned a primary causal substrate. They agreed on 140/150 labels (93.3%; Cohen's κ = 0.885). A third blinded pass adjudicated all ten disagreements, and the classifications were cryptographically sealed before the private archetype key was joined.
| Score band | Archetypes | Governance | Computational | Other | Governance + computational |
|---|---|---|---|---|---|
| Framed | 7 | 18 | 2 | 1 | 95.2% |
| Middle | 11 | 22 | 9 | 2 | 93.9% |
| Structural | 32 | 28 | 60 | 8 | 91.7% |
Across all 50 archetypes, source score versus governance share within the two dominant channels had ρ = −0.624 (seeded two-sided permutation p = 0.00001; leave-one-archetype-out range −0.688 to −0.608). Each of the three fixed domains had the same negative direction.
The most defensible population-facing result uses E9's probability-sampled 24 previously untested generated archetypes rather than the 26 purposive or census comparators. In that stratum the association was ρ = −0.791 (p = 0.00002; leave-one-out range −0.812 to −0.779). The pattern therefore did not depend on the ten reference archetypes or the hand-curated block.
The substrate question was chosen after E9 generation and screening were complete. Prespecifying and blinding the new classifications protects the labeling and analysis step, but cannot convert the question into a prospective confirmatory experiment. See the frozen method, design freeze, classification seal, full report, and machine-readable results.
What this changes¶
- It rejects the simple remedy of mining more physical sources merely to correct a demonstrated source-prime deficit. No such deficit appeared under the source-edge measure.
- It preserves a narrower reason to mine patents and design handbooks: as a source-form experiment asking whether specification- and design-shaped texts yield different archetypes.
- It clarifies that E9's broad opportunity-search result and the substrate result answer different questions. E9 showed that the method could generate researchable proposals across a much broader archetype sample. This follow-up shows that those proposals still have a strong substrate “house style.”
- It motivates a future prospective intervention that varies the prompt, rubric, or generation architecture while holding archetype/domain cells fixed. That would be needed to localize the constraint causally.
Reproducibility and validation¶
The analysis inputs, source hashes, blinded records, two classifier passes, adjudications, private key, joined rows, scripts, and results are retained in this directory. The independent validator checks freeze hashes, forbidden-field absence, row coverage, adjudication completeness, seal hashes, balanced-domain structure, clustered analysis outputs, and frozen-gate consistency. Its result is PASS.
Relevant scripts: