Framed-vs-Structural Retrieval Sweep — a refutation and a reframing (2026-05-22)¶
Predicted (from the structural/framed theory + the due_process diagnostic): framed primes embed far from neutral structural language (their frame holds them at a distance), so they'd be hard to retrieve from a domain-stripped query; structural primes embed near neutral structural language and retrieve easily.
Design. For 7 framed primes and 8 structural primes, an independent agent — blind to the framed/structural labels and to the hypothesis, given each prime's definition — wrote a ~40-word domain-neutral paraphrase of the prime's own structure (no naming the concept, no field jargon). Each paraphrase was embedded (bge-small) as a query and run against the prime catalog; we recorded the self-retrieval rank of the prime it described, against both the structural_signature corpus (SIG) and the core_idea prose corpus (PROSE).
Result — the prediction is refuted¶
Self-retrieval rank (1 = perfect), SIG corpus:
| ranks | median | in top-10 | |
|---|---|---|---|
| Framed (7) | 1, 1, 2, 3, 4, 5, 7 | 3 | 7/7 |
| Structural (8) | 1, 2, 4, 10, 11, 61, 64, 120 | 10 | 4/8 |
Framed primes were retrieved better, not worse. due_process, sovereignty rank 1; authority 3; rights_vs_freedoms 4; legitimacy 5; moral_relativism 7; epistemic_justice 2. The big misses were all structural: recursion 61, modularity 64, symmetry 120.
Why — a confound, and a reframing¶
1. The metric measures distinctiveness, not portability. Structural primes sit in dense near-synonym neighborhoods in the catalog: recursion competes with iteration / self-similarity / decomposition; modularity with decomposition / encapsulation / loose-coupling; symmetry with invariance / conservation / gauge-symmetry. A generic neutral paraphrase ("a thing defined by smaller versions of itself"; "self-contained replaceable pieces with boundaries"; "unchanged under a family of transformations") matches the whole cluster, so the exact target ranks behind its near-duplicates. Framed primes are more isolated (few near-duplicates), so they self-retrieve at rank 1-7. The framed paraphrases also retained distinctive fingerprints the instructions didn't fully strip ("a neutral judge", "stated reasons", "the last word", "boundary"), while the structural paraphrases were genuinely generic. So this sweep mostly measured concept distinctiveness, not frame-distance. It is the wrong instrument for the cross-domain-transfer question.
2. It reframes the due_process forward-case finding — and this is the more important point. due_process self-retrieves at rank 1 from a neutral description of its own structure. So the earlier immunology failure (due_process at rank 45 from the immunology scenario) was not "framed primes are far from neutral structural language." It was a genuine cross-domain instance→abstraction gap: the immunology problem foregrounds irreversibility, uncertainty, and asymmetric harm, whereas due_process foregrounds procedure, impartiality, and reasons. Those overlap only partially in embedding space. The hard problem is recognizing that a far-domain instance instantiates the abstraction — the structural analogy itself — which off-the-shelf embeddings don't capture, and this is a different (and harder) problem than "find the prime from a description of it."
3. Signature-embedding specifically helps framed primes. SIG beat PROSE markedly for several framed primes (epistemic_justice 2 vs 19; moral_relativism 7 vs 102; legitimacy 5 vs 11), consistent with structural_signature being the domain-stripped, structure-focused field; structural primes did slightly better on PROSE. So embedding the signature is the right call for framed content.
What we actually learned¶
- The simple "frame = embedding distance" claim is false as stated; framed primes are highly findable from neutral descriptions of themselves (and are more distinctive than structural primes, which blur into near-synonym clusters).
- The real retrieval barrier exposed by the forward case is cross-domain analogical retrieval (instance→abstraction across a domain gap), which applies regardless of framed/structural and is not fixed by meta-model phrasing alone.
- To properly test framed-vs-structural cross-domain transfer, we need the expensive version: for each prime, a far-domain instance/scenario, then measure retrieval rank of the prime. The cheap self-paraphrase sweep cannot answer it.
Bonus: harness validated¶
semantic_catalog.py (numpy + optional ChromaDB, bge-small ONNX) works. On the immunology meta-model query it returns, among 621 archetypes: irreversible_commitment_management (#1, 0.78), fail_safe_default, final_override_prevention, autonomous_action_zone_protection, physical_constraint_design_for_impossibility — i.e., semantic+meta retrieval reliably surfaces the safeguard/irreversibility cousin cluster that lexical search missed. The Layer-1 fix is real and usable now.
Caveats¶
N=15 primes, one (modest) embedding model, self-paraphrase metric (confounded by distinctiveness as noted). The blind paraphraser couldn't fully strip distinctive structural fingerprints. Treat the framed>structural ordering here as an artifact of distinctiveness, not a finding about portability.
Artifacts: _models/primes_{prose,sig}.npy, _models/archetypes_emb.npy, _models/*_meta.json; harness semantic_catalog.py.