Date: 2026-09-28. Decision: reject the composed candidate; preserve bounded individual-learning and diagnostic evidence. No weight was promoted to the EMMA library. The runbook fixed the task, splits, controls, and gates before candidate training. A development-only mechanism amendment after the first stopped run changed right-unit placement before sealed scoring.

Research question and controls

V13 showed that request-local AdaptiveWeight selection is implementable, but its Foundation tasks did not require two units simultaneously. V14 asked whether two separately trained low-rank units could supply the two positions of one answer through the frozen Foundation. This is a deliberately simple composition test: 48 training, eight development, and eight sealed symbol pairs. Pairs are disjoint across splits, while each individual symbol recurs. Left and right symbols map to separate output digits. A factorized per-symbol lookup scores 8/8 by construction, so this run cannot establish weight learning's advantage over memory or a simple symbolic program. It can expose whether native units preserve their individual effects when activated together.

The Foundation checkpoint hash was [checksum retained in the private evidence record]. It has 9,508,800 parameters and remained frozen. The left unit trained only against answer position one; the right unit trained only against answer position two. The two-scalar composition gate trained after both units were frozen. All selections were request-local. Training used the project's CUDA environment on an NVIDIA GTX 1650 with 4 GiB VRAM.

Staged diagnostic and mechanism change

The initial declared left block-5 FFN unit reached 8/8 on its development position after 180 steps. The initial right block-7 FFN unit reached 1/8, so the run stopped before composition or sealed scoring. That stopped record is saved in [retained internal evidence].

Changing the right placement, rather than repeating the same training, produced:

The last placement was selected using development only, trained for 240 steps, and included in the one sealed composition comparison. This result says that the right mapping needed an earlier attention-capable intervention in this task; it does not identify a generally optimal block.

Composition result

The left unit had 30,720 parameters. The right unit had 42,240. Both fit their owned training target and individually reached 8/8 on the owned development position. A separate two-parameter gate learned contribution weights of 0.626 for left and 0.529 for right; its training changed neither unit tensor.

The joint candidate missed the predeclared 7/8 sealed gate. Its sealed gain over the right-only arm is 4/8, but its loss relative to the individual right skill is also 4/8. The individual skills exist in separate weights; their simultaneous use is unreliable. The result rejects global scalar mixing as a sufficient composition mechanism for this pair. The low training loss of each unit would have hidden this failure without the held-out and ablation controls.

After the sealed comparison, a development-only reapproach added request-local per-token contribution masks. It isolated the left output at 8/8 but collapsed the right output to 0/8. This was not scored on the sealed split and is not a candidate qualification. It indicates that the right unit needs to modify earlier prompt states, while those same modifications interfere with the left unit. Hard answer-position masking removes too much of the right unit's computation. A future experiment should test a learned, query-conditioned activation-space composition or a narrower source/position-conditioned access path, with frozen individual units and an explicit mixer-only control. Repeating this global-scalar or hard-mask setup with more steps would not address the observed mechanism.

A second development-only diagnostic used the same saved units as separate Foundation branches: the left branch generated the first token; that token became part of the right branch's input. This explicit exchange scored 8/8 exact pairs on development. It did not run on a fresh sealed split and is not a qualified direct multi-Transformer fabric. It is nevertheless a concrete alternative to in-block delta addition: preserve independently working modules, communicate through a versioned result, and evaluate the whole operation. The current test is sequential because the second token depends on the first; future independent branch work can execute in parallel before a synchronization point. The lookup control also scores 8/8, so this does not justify neural execution on efficiency grounds.

Integrity, persistence, and limits

All 137/137 previously scanned weight binaries remained byte-identical. Foundation tensors stayed unchanged. The left unit's saved tensor hash matched its hash immediately after left training, proving right-unit and gate training did not alter it; the right unit's saved hash likewise matched its post-training hash. The candidate was stored only in the ignored experiment directory, with SHA-256 [checksum retained in the private evidence record]. A fresh Python process reloaded it and reproduced the sealed 4/8 exact result. Deactivating units reproduced the 0/8 Foundation route. The corrected historical Foundation control convention scored 3/10, 2/10, 7/10, and 2/10 on old skill families A–D; the globally composed candidate scored 4/10, 2/10, 8/10, and 2/10 on the same narrow probes. These are not retention tests of the old AWUs, which were not mounted in this V14 task.

The new runtime token-gate branch validates per-token shapes and [0,1] bounds and leaves existing paths unchanged when unused. Its mechanism test passed. The full backend regression suite passed 283 tests. The task and output grammar are narrow; there is no evidence here for general reasoning, memory superiority, autonomous routing, optional block insertion, or remote whole-stack module cooperation.

Architectural implication

AdapterFusion separates independent adapter learning from the subsequent composition stage. TIES-Merging explains why parameter interference can make simple merging fail; X-LoRA is a reference for finer-grained dynamic contribution. V14 adds direct EMMA evidence that two strong individual AWUs can conflict inside a real frozen Foundation, even when placed in different blocks and controlled by learned global scalars. The next candidate should preserve the individual artifacts and compare two routes on a fresh split: a bounded activation-space mixer and explicit branch-to-branch exchange. Each needs a mixer/bridge-alone control, module ablations, parent retention, restart, and latency measurements before promotion.

The result advances the research by locating the unresolved boundary precisely: independently trainable Foundation modules work on this bounded task; reliable same-pass weight-delta composition does not yet work. Explicit result exchange is promising on development but unqualified.