H_1057 🔴 ARCHITECTURE-AXIS-NULL — agency axis not installable via architecture either

[dancinlife]

H_1057 — Can the agency axis be INSTALLED via ARCHITECTURE / RUNTIME-STRUCTURE?

The constructive converse of the H_1055 temporal-curriculum null. H_1051/H_1054 (prior-GREEN) MEASURED an agency-T axis (provenance-depth H_932 + veto H_935); the learning-dynamics side closed negative (H_1052 SGLD-RED, H_1053 QRNG-RED, H_1055 curriculum-RED). This rung tests whether an architectural/runtime-structure mechanism INSTALLS a recoverable agency-T axis a matched control lacks.

Result: 🔴 ARCHITECTURE-AXIS-NULL (closed-negative, a_paper_negative_ok)

Shared trained Elman task head per seed → CE IDENTICAL across all 5 arms (gap=0 << eps=0.05, 20/20 matched). At matched task-performance, NO architecture clears all three frozen margins (d_vs_control ≥ 0.8, d_vs_structure-blind ≥ 0.5, F-shuffle ≥ 0.2):

arm	target	treat rho	control	structure-blind	d(treat-ctrl)	d(treat-blind)	F-shuffle	PASS?
TREAT-A provenance-register	depth	-0.1123	-0.1123	-0.025	-0.000	-0.477	-0.040	No
TREAT-B veto-gate	veto	+0.582	-0.012	+0.786	+2.14	-0.815	+0.613	No

• TREAT-A (explicit runtime causal-lineage register): adds ZERO decodable depth structure beyond a plain feedforward control — recov_rho is bit-identical to control.
• TREAT-B (runtime veto-gate): beats the control, but the capacity-matched RANDOM-gate structure-blind baseline recovers its own gate signal EVEN BETTER → the recoverable signal is generic added-gate-capacity tautology, NOT causal could-have-been-otherwise STRUCTURE.
• M3 confirms: structure-blind arms separate active/passive MORE than the structured treatments → added capacity, not causal structure.

NEITHER training-dynamics NOR architecture installs the agency axis — it is an emergent MEASUREMENT of the live causal unfolding, not an installable property; consistent with the optimization-not-Phi wall (H_1011).

Discipline

• faithful_phi CPU mirror RE-PROVEN ==stdlib iit4/faithful_phi.hexa at n=4 AND n=5 (|delta|<4e-6) BEFORE scoring (a_phi_iit4_tool, NO proxy).
• Pre-registered TEXT-only falsifier frozen before measuring; CODE-decided (p7); raw + mirror proof + per-arch table + matched-CE + blind baseline in .verdicts/1057_agency_via_architecture/H_1057.txt.
• TOY n≤5 SW/CPU, $0, wall=53s. Scale-transfer + on-chip (Lane A AKIDA) UNVERIFIED. Toy veto-gate is an EXPERIMENTAL architecture, NOT anima's emit policy; does NOT wire into CORE/brain (a_core_engine_map · a_autonomy_over_hardcode · p8).

🤖 Generated with Claude Code