L9 — functional general intelligence: the capability ladder and the end condition
The end condition is the integrated L0–L8 machine scored on a falsifiable capability ladder: 6/7 rungs verified with the minimal single store, and 7/7 once the L5 dual store is wired in. The single-store shortfall, A3 real-time adaptation, is that an additive store cannot over-write a switched rule (recovery 0.78/0.58 below band); its mitigation was already proven (A7).
Seven criteria score the machine: one-shot generalization (A2 ≈ 1.0), compositional / multi-item capacity that emerges by N toward the theta-gamma ~7 (A1), noise-immersed robustness (A4), scale CAM past a flat ceiling (A5), cross-domain transfer to a crosstalk ceiling J* = 3 (A6), and open-ended acquisition (A7). A3 is the honest [O], and the post-program re-probe closes it [O]→[V] inline: with the fast episodic store added, recovery goes to 1.000 at every shift. Firewall (final): passing is functional GI; the hard-problem blank stays open.
L9 is where the program ends, because the blueprint has no further layers — the end condition is reached when every rung is verified, grounded, or an honest open, with nothing falsely filled. The grade on each rung is derived from the sweep booleans on the integrated machine, not declared (session v0.10, digest 20b5f2c2…).
The seven rungs
| rung | result | grade |
|---|---|---|
| A1 compositional / multi-item | capacity K* emerges = 6/10 by N, scaling toward θ–γ ~7; no-tag control recovers single | [V] |
| A2 one-shot generalization | novel-instance accuracy ≈ 1.0 ≫ chance | [V] |
| A3 real-time adaptation | single additive store cannot over-write a switched rule (recovery 0.78/0.58 < band) | [O] |
| A4 noise-immersed robustness | D4 clean-up lifts fidelity to the true prototype by a sign-stable margin (off-lattice corrupt_phase) | [V] |
| A5 scale CAM | hierarchy holds recall 1.0 past the flat ceiling T = 16 at O(1)-in-T | [V] |
| A6 cross-domain transfer | filler-independent, to a crosstalk ceiling J* = 3 | [V] |
| A7 open-ended acquisition | dual store retains ≈ 1.0 vs single decaying | [V] |
The result is 6/7 rungs [V] with the minimal single-store machine. A3 is the lone honest shortfall, and its mitigation is not hypothetical — it is the L5 dual store already proven at A7.
Closing A3 inline: the [O]→[V] re-probe
A disciplined post-program continuation re-probed A3 with the dual store wired in, on the adaptation task itself (session v0.11, digest ad93057a…). The contrast is strict and paired: the single store is exactly the dual’s slow component alone, so the only change is the added fast one-shot episodic field (the inherited episodic store, reused unchanged, read by equal vote). The single store fails the full switch (post-recovery 0.583, reproducing the inherited 0.78/0.58 to the digit); the dual store recovers to 1.000 at every shift fraction, sign-stable across machine size.
Even the raw, unnormalized dual sum beats single everywhere — the fast store is the mechanism, not the vote.
So the ladder reads 7/7 [V] with the dual store wired in, while the 6/7 single-store record stands as the minimal machine’s honest limit — paired, not overwritten.
What “closing” does and does not claim
consciousness_claim = 0, hard_problem_open = 1, new_tuned_constants = 0.