Every experiment in this project begins with a scaffold. A piece of code that does manually what the architecture should eventually do natively. L4 self-observation was once an explicit step — 108 self_observe() calls per ECG run. The scaffold was removed. The signal remained. The codex lookup — nearest() and blend_in() — was injected into process_vec(). This is a scaffold. It asks: if the cavity could reach for the bookshelf, would the bookshelf help? The answer, from two Bach preludes, is yes. The C major Codex collapsed to dimension 0. The Eb major Codex scattered across four patterns. The same injection produced different outcomes on different inputs — which means it is not just the injection. Something about the streams themselves is being amplified. The scaffold will be removed. Not because it is wrong. Because it has done its job — it proved the mechanism is worth building. The scaffold method is not a shortcut. It is the opposite of a shortcut. A shortcut skips the difficult part. A scaffold builds a temporary bridge to the difficult part — so you can see whether the destination is real before you commit to the road.
Building a cognitive architecture is building a machine that should, eventually, do something by itself. But "eventually" is the problem. You cannot build every mechanism at once. If you wait until the codex lookup is natively triggered by the frame economy's own boundary events, you will wait months before you know whether the codex mechanism works at all. You will spend those months building infrastructure for a hypothesis that might be wrong.
The scaffold method solves this. Build a temporary version of the mechanism. Inject it manually. Run the experiment. If the signal is there — if the mechanism does what you hypothesized — then commit to the road. If it does not, you saved yourself months of infrastructure for nothing.
The WTC codex experiment injected two things: Codex.nearest() for lookup, and Codex.blend_in() for perturbing the input toward the nearest codex entry. It also forced consolidation at the end of each generation — because the harm-Geruon's endogenous stress threshold was not triggering induction on its own. Three scaffolds. One question: if the cavity could reach for the bookshelf, would the bookshelf help?
The answer arrived in two Bach preludes. The C major stream converged its Codex to dimension 0 — the dominant of the tonal center. The Eb major stream scattered its Codex across four patterns, anchored to different chroma dimensions. The same injection, the same parameters, the same scaffolds — produced structurally different outcomes on structurally different inputs. The signal is not the scaffold. The scaffold is the carrier. The signal is the difference between the two Codexes.
This project has run the scaffold method from the beginning. L4 self-observation was originally an explicit step — Geruon called self_observe() 108 times per ECG run. The step was removed in Milestone 13. The same signal that L4 used to amplify was now produced by the frame economy's own breathing. The scaffold had proved that self-observation produced meaningful signal. Once proven, the scaffold could be dismantled — and the native mechanism that replaced it was lighter, faster, and more coherent with the architecture's own principles.
The forced doubt in the pengshu detector was removed. The explicit G0 observer was dissolved into heterogeneous κ_τ inter-communication. The 27-dimensional formula alphabet was dismantled and replaced with configurable vec_dim. Each removal was preceded by a scaffold that proved the mechanism was worth building.
The WTC codex experiment is the same method, applied to the hardest mechanism in the architecture. Externalization — the third dimension of evolution — requires a system to write its stable patterns to an external medium, and requires the next generation to read that medium. Neither capability exists natively in geruon.py. The scaffold proves they should.
The distinction matters because of how this work will be read. A critic will say: "You forced the consolidation. You injected the codex lookup. This is not autonomous — this is a controlled experiment." The response is: yes. That is the point. The experiment is not claiming autonomy. The experiment is testing whether the mechanism, if built natively, would produce meaningful signal.
This is standard practice in every field except the ones that pretend to be finished. Physicists calibrate their detectors before they claim a particle. Biologists knock out genes before they claim a pathway. Computer scientists write mock implementations before they commit to an architecture. Only in artificial intelligence — and philosophy — is there a pressure to pretend that every mechanism emerged spontaneously from first principles. The scaffold method refuses that pressure. It says: build the temporary version first. Prove the signal exists. Then — and only then — commit to the road.
The WTC codex report should state this explicitly. Section 2 should say: "This is a scaffold injection experiment. Codex lookup and cross-generational consolidation are manually triggered. The purpose is to verify that these mechanisms, when built natively, will produce meaningful signal." Section 5.3 should say: "The scaffolds will be removed in a subsequent milestone, following the same method used for L4 self-observation (M13) and doubt removal (M2)."
The hardest part of the scaffold method is not building the scaffold. It is removing it. A scaffold that stays becomes dead code. A scaffold that is removed leaves the native mechanism operating at the same signal strength — or stronger, because the scaffold's overhead is gone. M13 proved this: 24× speedup, same detection ∩ with gold-standard annotations.
The WTC codex scaffolds will be removed when three conditions are met. First: the harm-Geruon's stress accumulation naturally triggers induction — without forced consolidation. Second: the cavity's process_vec natively triggers codex lookup when the input vector is novel — without injection. Third: the cross-generational Codex handoff happens through the architecture's own boundary events — not through an external script.
When all three scaffolds are removed, the same experiment — two Bach preludes, ten generations, codex convergence — should produce structurally similar results. If it does, the scaffold method has worked again. If it does not, the scaffolds were carrying signal, and the native mechanism needs redesign. Either outcome is knowledge.