Date: 2026-09-22 Status: FOUNDATION CONTEXT-AWARE TRAINING REQUIRED

Reader-selection rationale

The previous pooled/late ContextReadUnit changed hidden state and passed a 31/32 micro-overfit, but produced 0/127 unseen identities. Token-preserving earlier placements also failed the micro gate. A late output residual forced the reader to synthesize an answer representation instead of exposing a Context stream to Foundation computation.

Native architecture

The prototype added a dedicated block-local Context attention branch. At the selected block, normal self-attention runs first, then the main hidden stream produces queries over an independent Context token sequence, followed by the existing FFN. Context uses the frozen Foundation embedding table and is never flattened into the main token sequence.

The primary branch is:

x1 = x + SelfAttention(RMSNorm(x)) x2 = x1 + tanh(g) * ContextAttention(RMSNorm(x1), C1) x3 = x2 + FFN(RMSNorm(x2))

Placement: block index 3 (the fourth block). Trainable parameters were only the native Context Q/K/V/O projections and gate: 409,601. Foundation, Memory, AdaptiveWeightUnits, and Context state remained frozen.

Results

Native separate cross-attention

The first run with a near-zero gate produced 2/32 micro-overfit and 0/64 unseen identities. A bounded gate initialization correction produced 1/32 micro-overfit and 0/64 unseen identities. Attention entropy collapsed near zero and the branch did not establish a reliable copy path.

KV extension alternative

The authorized alternative appended Context-derived Keys/Values to one selected block's self-attention. It used 204,800 trainable parameters, 1,200 steps, AdamW, learning rate 5e-3, and the same copy task.

The KV extension can memorize the tiny training set, but it does not perform unseen identity transfer.

Controls and integrity

The original Foundation checkpoint remained unchanged:

[checksum retained in the private evidence record]

No AdaptiveWeightUnit or Persistent Memory state was modified. No candidate Foundation checkpoint was created. Failed native candidates were not retained as model binaries.

Untested conditions

Because both native mechanisms failed the unseen-identity gate, no larger training, 500-example sealed qualification, binding, structural, source, multi-drone, movement, routing, or Global Context work was started.

Evidence

• [retained internal evidence]
• [retained internal evidence]

Disposition

FOUNDATION CONTEXT-AWARE TRAINING REQUIRED

Both a native separate Context-attention branch and the bounded KV-extension alternative can be made mechanically trainable, but neither generalizes arbitrary unseen Context identities with the Foundation frozen. Stop for review before training a new Foundation revision with Context present.