title: Experiment 13: Second-slot portfolio policy¶

Experiment 13: Second-slot portfolio policy¶

Part of Inverse Innovation with the Encyclopedia of Abstractions · Experiment 13: Second-slot portfolio policy · Last revised August 2026

Result in one sentence¶

The frozen operational verdict is USE AS BOUNDED PORTFOLIO COMPLEMENT: ordinary diversity produced 41/60 (68.3%) incremental survivors and alternative-substrate diversity produced 32/60 (53.3%), with 10 substrate-only and 19 ordinary-only additions.

This result selects a search policy under this prompt–model–corpus bundle. It does not certify any proposal as a novel or effective real-world invention.

What was tested¶

Twelve probability-sampled, previously untested generated solution archetypes were crossed with five probability-sampled domains not used in E6/E9/E12. Each of the 60 cells received a common first proposal, then two isolated second proposals conditioned on that exact P1: one ordinary-diverse and one required to use a non-governance, non-computational primary causal substrate.

An additional proposal counted only when blinded assessment found it complete, structurally faithful, and independent of P1; its four-source public-web light screen survived; and, for the substrate arm, blinded classification confirmed an allowed primary substrate. See METHOD.md for the frozen design.

Primary results¶

Outcome	Ordinary-diverse P2	Substrate-diverse P2
Generated proposals	60	60
Policy-compliant proposals	not required	60
Incremental survivors	41 (68.3%)	32 (53.3%)
Rescues among P1 failures	5	8
Blinded quality wins	32	28

The paired table was 22 both, 10 substrate-only, 19 ordinary-only, and 9 neither. The rate difference was -15.0% for substrate minus ordinary; the archetype-clustered 95% bootstrap interval was [-31.7%, +3.3%]. The exact 12-archetype sign-flip sensitivity test was p=0.1953.

P1 itself survived in 43/60 (71.7%) cells, leaving 17 P1 failures. That confirms why rescue alone was prespecified as secondary: it provides a smaller denominator than counting nonredundant additional opportunities whether or not P1 survived.

Frozen policy gate¶

Prefer substrate diversity outright required a gain of at least five points and more substrate-only than ordinary-only survivors.
A bounded-complement verdict required at least 48 compliant substrate proposals, at least six substrate-only survivors, and no more than a 15-point yield deficit.
The observed checks are recorded exactly in analysis/results.json.

The verdict therefore means that a resource-constrained search should follow the frozen rule above. It should not be generalized into a claim that one substrate is intrinsically more inventive.

Sample¶

Selected archetypes: Agency / Structure Attribution Balance, Ensemble and Population-Level Equilibrium versus Individual-Level Heterogeneity, Abductive Explanation Selection, Event-Log-Centered Modeling, Multiple-Testing Discipline, Continuity–Rupture Regime Diagnosis and Transition Design, Discrete–Continuous Model Selection, Deterioration Monitoring, Stock–Flow Accumulation Control, Predictive Precommitment Correction, Second-System Complexity Restraint, Fast–Slow Store Coupling.

Selected domains: Engineering & Design, Law & Governance, Sport Science & Kinesiology, Literature & Literary Theory, Library & Information Science.

Resource use¶

E13 retained 458 scientific call attempts, with 0 failed transports, 14.3 summed call-hours, and a 13.3-hour first-to-last telemetry span. Token telemetry reports 80,429,441 input tokens (54,841,600 cached) and 1,715,925 output tokens, including 618,283 reasoning tokens.

What remains unmeasured¶

The light screen is bounded and can miss prior art. All generation and most screening used one model family; blinded measurement used another model from the same provider. Domain experts did not judge feasibility or novelty, and no candidate was implemented. The experiment therefore estimates model-mediated search yield, not realized innovation.

The complete cell table is in analysis/cell_level.csv; all incremental survivor summaries are in analysis/incremental_survivors.jsonl; raw prompts, responses, web sources, blind keys, adjudications, and telemetry are retained for audit.