Status

Research sequence complete. Replacement pipeline qualified for the locked structured semantic task.

The original raw neural-access candidates failed their predeclared micro gate. The replacement succeeded. Both facts are part of the result.

Experimental scope

V4 proved query-conditioned cross-Drone path selection and exact payload return. V5 removed the copy shortcut. The required answer was absent from every source and had to be derived from four simultaneous local observations plus the policy held in G.

The qualified runtime is:

Implementation code and detailed machine records are retained separately. This public edition presents the study methods, aggregate results, and qualification boundaries.

Locked data

The largest development class contains 281/576 examples (48.78%).

Run 1: raw byte-level semantic fusion — failed

Two matched mechanisms trained around the frozen Foundation:

Training loss fell below 0.1 while unseen performance stayed close to the majority baseline. The mechanisms learned the small training surface but did not learn the semantic rule across unseen entity strings. The run stopped at 200 steps. No full training was launched and the sealed set was not opened.

This failure rejects these two V5 byte-level access formulations. It does not reject parallel Context or cross-attention in general.

Run 2: typed evidence and semantic Decision MicroModel — qualified

The replacement uses deterministic readers for the known telemetry schema, then a 57,699-parameter, two-block contextual feature Decision MicroModel with independent weights.

The prediction distributions exactly matched the target distributions. Because each split used new entity identities, this proves identity generalization within the fixed telemetry schema and rule family. All truth-table patterns occur in training, so it does not prove unseen-rule or combinatorial extrapolation.

Run 3: semantic state to frozen Foundation — qualified

The qualified Decision MicroModel was frozen. A 355,491-parameter gated bridge was trained at Foundation blocks 4, 6, and 8. The qualified Foundation remained unchanged.

The learning trace was informative:

• step 1: 0/192, malformed output;
• step 50: 0/192, label characters emerging but unstable;
• step 100: 139/192 = 72.40%, with AMBER commonly rendered as AMBRA;
• step 150: 139/192 = 72.40%;
• step 200: 192/192 = 100%.

This shows the bridge learned to make the frozen Foundation realize the semantic state rather than receiving a hardwired renderer.

Persistence and integrity

A new process reconstructed Foundation, the Decision MicroModel, and the bridge from saved weights:

• representative sealed sample: 192/192;
• complete groups: 64/64;
• missing-source rejection: C1, C2, C3, C4, and G all passed;
• status: PASS.

Unit tests cover answer non-leakage, simultaneous independent gates in the failed baseline, typed provenance, and missing-evidence rejection.

Findings

Within the locked structured task:

1. C1-C4 and G can all contribute to one derived semantic decision.
2. G changes the interpretation of unchanged local evidence.
3. A small independent Decision MicroModel can compose the five evidence tokens perfectly on unseen identities.
4. A small replaceable bridge can communicate that state to a frozen Foundation without changing Foundation weights.
5. The composed capability survives restart.

Qualification limits

• unstructured or mixed natural-language Drone reading;
• unseen semantic rules and deeper multi-hop extrapolation;
• free-form explanations grounded in the semantic state;
• learned reader selection across heterogeneous readers;
• Context Contribution decisions under variable source costs;
• G updates and persistent learned Global Context;
• K>1 raw neural attention as a general mechanism;
• interaction with Memory, AdaptiveWeight composition, teacher/student learning, and self-development.

Architectural decision

The report recommends preserving both experimental paths:

Implementation code and detailed machine records are retained separately. This public edition presents the study methods, aggregate results, and qualification boundaries.

The failed byte-level fusion variants do not justify repetition without a changed hypothesis. The next Context run should broaden the reader contract with mixed structured/natural-language evidence and require a short explanation whose supporting source provenance is scored separately.

The researched implementation sequence, interfaces, matched controls, holdout design, and stop gates are specified in the EMMA V6 heterogeneous-evidence plan.