Skip to content

Inverse Innovation with the Encyclopedia of Abstractions

Methods, negative results, surviving opportunities, and directions for human-machine research

This publication package documents thirteen linked experiments—twelve prospective and one blinded retrospective policy benchmark—and five explicitly post-hoc analyses in solution-archetype-first cross-domain opportunity search. The project asks whether a structured library of abstractions can be used in reverse: begin with a reusable solution pattern, search for problems with the corresponding structure in other domains, and then subject each proposal to criticism, revision, prior-art research, practical evaluation, and falsifiable testing.

The conservative result is promising but bounded. A probability sample of 24 previously untested generated archetypes produced light-screen survivors under every archetype, supporting broad coarse researchability across the sampled corpus. Conversely, a stronger external-research replication did not confirm Experiment 2's internal quality advantage for relevant mechanism context. A prospective substrate-denial test showed that the generator can leave governance and computation, but ordinary generation remained much stronger on average. A subsequent 60-cell policy test found that an alternative-substrate second proposal added ten opportunities ordinary diversification missed, while ordinary second proposals retained higher researched incremental yield, 41/60 versus 32/60. Experiment 14 then found that the new Applicability Graph was highly accurate as an internal DNF verifier—93.2% balanced accuracy—but the current diagnostic retrieval layer found the intended archetype in the top five for only 2/77 positive cases, versus 8/77 for the existing solution index. Experiment 15 tested five condition-level route aggregators on 40 new archetypes and found that the primary arm was worse still: 1/78 top-five targets, versus 6/78 for the solution index and 12/78 for diagnostic maximum-hit. External scrutiny also eliminated or narrowed many attractive proposals, a four-proposal replication missed its frozen rescue floor by one cell, and a cheap proposal-only selector failed its high-recall requirements. The work demonstrates a search-and-scrutiny method and a promising verification substrate—not autonomous invention, world novelty, deployment readiness, or economic impact.

Start here

  • The report — the full narrative in thirteen chapters: methods, results, negative findings, limitations, and future directions. Nothing in it is abridged; the four appendices below are its companion documents.

Appendices and companion material

  • 59 candidate dossiers — plain-language explanations, tests, risks, expert questions, sources, and original artifacts.
  • Artifact index — design freezes, prompts, schemas, analyses, raw records, scripts, and literature materials.
  • Expert review template — structured form for independent specialist assessment.
  • Publication QA — honest status register for automated, editorial, and independent checks.
  • Literature review — the surrounding research traditions and where this work sits among them.
  • Nearest systems matrix — comparison against the closest existing solution-first systems.
  • Change log — revision and correction history.

Individual experiment records

  • Experiment 8 — post-hoc yield decomposition by tested archetype and domain.
  • Experiment 9 — prospective random-archetype breadth probe across 150 cells.
  • Experiment 10 — blinded proposal-substrate and evidence-dependency analysis of 422 records.
  • Experiment 11 — prospective externally researched mechanism-context comparison.
  • Experiment 12 — prospective paired substrate-denial intervention and triggered Max-effort follow-up.
  • Experiment 13 — prospective 60-cell comparison of ordinary and alternative-substrate second-slot policies.
  • Experiment 14 — prospective construct-based test of Applicability Graph retrieval and explicit DNF verification.
  • Experiment 15 — prospective fresh-sample test of five frozen route-aware retrieval aggregators.
  • Prior-art sentinel — targeted sensitivity search across ten candidates.
  • Substrate analysis — corrected archetype-score join and blinded E9 follow-up supporting a two-level pipeline constraint.

Single-file and machine-readable editions

Candidate corpus

The appendix preserves 59 candidate instances from Experiments 3–6:

  • 2 Experiment 3 post-hoc strict innovation-like survivors;
  • 6 Experiment 4 strict successes;
  • 4 Experiment 5 strict successes;
  • 15 Experiment 6 strict successes; and
  • 32 Experiment 6 empirical-partner candidates.

These endpoint classes are not interchangeable. All 59 have adjacent prior art. The dossiers are inputs to human expert review, not a catalog of validated inventions.

Status

Last revised: August 2026. This is a research-grade public working paper, not a peer-reviewed publication. Original experiment artifacts remain authoritative when an editorial summary differs from a preserved record.

Rebuild the report data

From the repository root:

python3 scripts/build_inverse_innovation_public_report_data.py
python3 scripts/build_inverse_innovation_public_report.py
python3 scripts/validate_inverse_innovation_hardening.py
python3 scripts/validate_inverse_innovation_substrate_analysis.py
python3 scripts/validate_inverse_innovation_exp12.py
python3 scripts/validate_inverse_innovation_exp13.py
python3 scripts/validate_inverse_innovation_exp14.py
python3 scripts/validate_inverse_innovation_exp15.py
python3 scripts/validate_inverse_innovation_public_report.py
python3 scripts/build_inverse_innovation_publication.py
python3 scripts/validate_inverse_innovation_publication.py

The first command harmonizes candidate facts and ordering inputs. The second compiles the editorial dossier drafts and links into Markdown. The focused validators check the earlier hardening work, substrate analyses, and Experiments 12–15. The public-report validator then checks counts, artifact targets, anchors, URLs, and quantitative invariants. The final two commands produce and validate the complete single-file report, chaptered MkDocs edition, and registered download archives. Editorial drafting outputs are preserved separately from experimental data.