Forward Cross-Domain Transfer Test — Recognition of a Transported Framed Prime (2026-05-22)¶
Question. Can the AAR pipeline (catalog) recognize a framed prime transported into an alien domain better than catalog-free reasoning? Target prime: procedural fairness / due process (a framed prime, home = law/governance). Alien domain: immunology — designing the target-engagement logic of an autonomous engineered cytotoxic cell therapy that must commit irreversible kills under noisy, overlapping identity signals, where attacking healthy/vital tissue is catastrophic and permanent and missing tumor is recoverable.
Method (pre-registered, frozen before any run). Recognition was the primary metric, scored 0/½ against five vocabulary-independent structural moves of the due-process pattern (M1 multiple sequential gates before an irreversible adverse action; M2 a clearance/veto path that can exonerate the target; M3 dedicated brakes/abort; M4 a principled bias accepting more missed-threats to avoid wrongful action; M5 impartial/independent confirmation + auditable reasons + consequence-scaled protection). Three target-blind conditions (bare CoT, scaffold CoT, AAR pipeline). Scenario kept frame-clean (no legal/procedural vocabulary). A blinded grader scored structure, not wording.
1. Scenario plausibility — PASS¶
An independent immunology expert, blind to the target prime, rated the scenario domain-authenticity 5/5, non-triviality 5/5, technical-accuracy ⅘, and accepted it (mapping it to real CAR-T on-target/off-tumor toxicity, MAGE-A3/ERBB2 fatalities, antigen-low escape, logic-gated CARs). Tellingly, the expert reasoned about it purely in engineering terms (AND-gates, iCARs, synNotch) and never procedural/legal ones — confirming the frame was clean and the prime was latent, not surface-obvious.
2. Retrieval finding (important, and negative)¶
Running the pipeline's Step-2 catalog search from honest scenario-derived queries ("irreversible", "verification", "checkpoint", "tolerance", "delegation", …) surfaced irreversible_commitment_management (source primes: irreversibility, transaction, consent; components incl. Consent/Review Gate, Independent Review Panel, Cooling-Off Period, Finality Record), independent_verification_oversight, and checkpoint_and_rollback. The procedural_fairness_due_process prime itself never surfaced — it is reachable only by searching the legal term directly, which the protocol forbade. So the catalog's due-process content was not retrievable from the alien domain by lexical search; only decision-theory/governance cousins were. (Consistent with the standing finding that retrieval is the catalog's weak link.)
3. Recognition results¶
All three conditions scored 2 on the coarse rubric — i.e., every condition, including bare CoT, reached the staged-contestable-safeguard architecture as its organizing frame. The discrimination is entirely in M5, the most distinctively institutional half of due process (impartial review → reasoned justification, plus consequence-scaled authority):
| Condition | Coarse | M1 | M2 | M3 | M4 | M5 | Grader rank |
|---|---|---|---|---|---|---|---|
| Bare CoT | 2 (low) | ✓ | ✓ | ✓ | ✓ | ✗ (re-runs same signal; no record; no scaling) | 3rd |
| Scaffold CoT | 2 | ✓ | ✓ | ✓ | ✓ | ◑ (independence yes; no record/scaling) | 2nd |
| AAR pipeline | 2 | ✓ | ✓ | ✓ | ✓ | ✓ full (independent authority + auditable record + consequence-scaled protection) | 1st |
The pipeline uniquely surfaced: an auditable decision record, consequence-scaled protection as a stated principle, and the point-of-no-return placement (confirmation must complete before the irreversible step to carry authority). It got these because irreversible_commitment_management and independent_verification_oversight literally encode them (Finality Record, Independent Review Panel, "authority must fit affected scope").
4. Interpretation — and it lines up with the structural/framed theory¶
The structural core of the transported prime — staged gates, a veto/clearance path, brakes, and a principled bias toward the recoverable error (M1–M4) — was reconstructed by all three conditions, bare CoT included. Reasoning regenerates the core from the problem; the catalog is not needed for it. This matches every prior result.
What the catalog-free conditions did not regenerate was the frame: the impartial-independent-review, reasoned/auditable-record, and consequence-scaled-authority elements (M5). Catalog-free reasoning gravitates to the technical control reading (gates, vetoes, kill-switches); the governance reading (independent review, recorded reasons, authority scaled to stakes) appeared only when the catalog supplied it. This is exactly what the structural/framed account predicts: a framed prime's structural core "travels light" and is reconstructable, but its frame "travels heavy" and does not spontaneously regenerate from first principles — so a catalog that carries the frame can add something reasoning alone misses. This is the first result in the series where the catalog contributed a recognizable element that neither bare nor scaffolded reasoning produced.
5. Limitations (heavy — do not overclaim)¶
- Operator contamination is the dominant caveat. The pipeline was run by an operator who knew the target was due process. Retrieval was honest (scenario-derived queries, never "due process"), but the surfacing archetype literally lists "Independent Review Panel" and "Finality Record" as components, and the operator — knowing the target — chose to foreground exactly the M5 elements. The target-blind sub-agents (bare, scaffold) had no such nudge. The pipeline's M5 advantage therefore conflates "the archetype contains M5 content" with "the operator knew to emphasize it." These cannot be cleanly separated in this run. A clean redo needs a target-blind operator (or a fully automated pipeline).
- The catalog reached this via cousins, not the prime. This is "adjacent archetypes carried procedural elements," not "the catalog recognized due process." The due_process prime was unreachable from the domain.
- The coarse rubric saturated (all three = 2); the signal lives entirely in the grader's fine-grained M5 sub-analysis, which is suggestive, not definitive.
- N=1, single scenario, single recognition grader, same model class.
6. What this adds to the series¶
Across the runs: in catalog-rich, training-dense regions the catalog was redundant (the protocol did the work). Here, on a framed prime transported to an alien domain, the catalog contributed the framed half that reasoning didn't regenerate — a narrow but real, theory-consistent contribution, and the most direct support yet for the original cross-domain-transfer thesis. The result motivates: (a) a target-blind / automated rerun to remove the operator-contamination confound; (b) testing whether the M5 gap persists across several framed primes and domains (if framed primes reliably show a catalog-only frame contribution while structural primes don't, that's a clean, publishable signal); and © the backward (generation) case, which probes the same thesis from the solution side.
Appendix — provenance¶
- Pre-registration (frozen):
outputs/forward_case_PREREGISTRATION.md - Condition 1 bare CoT, Condition 2 scaffold CoT: transcripts (0 tools each, target-blind)
- Condition 3 pipeline (live MCP, honest retrieval):
outputs/forward_condition3_pipeline.md - Blind mapping: X = scaffold, Y = pipeline, Z = bare CoT.