Overall status: EXPERIMENTAL MODEL — NEEDS IMPROVEMENT Reader status: QUALIFIED FOR TESTED FORMS; broader generalization needs improvement

Purpose

This is a new, continuing experimental model assembly. It combines the strongest currently preserved Context reader with the active Context-native Foundation candidate so broader EMMA Context development can continue while the reader's uncovered limitations remain visible and independently replaceable.

The assembly is a manifest of versioned components, not a merged checkpoint. Loading it verifies each artifact hash and its Foundation lineage. The qualified Foundation parent, active Context-native candidate, and V1.2 reader artifact remain separate and unchanged.

Included components

The model-level status is NEEDS IMPROVEMENT because the assembled system is not a generally qualified Context model. The V1.2 artifact's narrower successful result remains valid for its tested named-field/query forms; this status does not erase or overstate that component evidence.

Evidence assessment

V1.2 scored 99.75% record selection and exact value copy, 99.26% complete counterfactual groups, and 97.65% final answers on its locked 1,620-example set. That establishes a useful working component for the tested named-record formats and query templates, not unrestricted reading or broad structural generalization.

The subsequent V1.3 structural attempt reused that modular reader design across a harder XML/query-composition holdout. It reached 99.51% record selection/copy and 97.47% final answers, but complete counterfactual groups were 98.52%, below the 99% gate. It was not promoted; its temporary candidate binary was removed after recording its evidence. The result identifies a remaining improvement area without invalidating V1.2's narrower tested-form status.

Direct Context Transfer in the active Context-native candidate remains 499/512 (97.46%) on the recorded reference set. The qualified Foundation parent remains bit-identical. These measurements are carried as component references, not as a claim that this new whole-model assembly has passed qualification.

Composed runtime path

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

Each boundary remains replaceable. The reader, parser, value adapter, manual router, Context-native Foundation candidate, and immutable Foundation parent are represented separately in the model manifest. The loader checks artifact and tensor hashes and refuses mismatched Foundation lineage.

Continuing the general model

The new assembly gives later work a stable base. New modules can be added as separate versioned components while this reader stays frozen. For each component, the model manifest and run evidence should distinguish:

• implementation status;
• component-level validation scope;
• known improvement needs;
• artifact identity and lineage;
• whether the whole assembly has been behaviorally qualified.

The next general-model increments can add interfaces and separately testable modules for broader Context retrieval, multiple independent Drones, a learned Context Router, Global Context, and movement. Those additions do not require changing the V1.2 artifact. When the reader is improved, the new reader should be trained and evaluated as a new component version, then substituted into a new model revision after its own evidence is recorded.

The current version deliberately does not claim those future modules are implemented. Its loader currently enables the declared named-record parser and fixed C1 source. The V1.3 held-out formats remain evidence for follow-up work, not newly qualified runtime scope.

Validation

The assembly loader is designed to:

• verify the active Context-native checkpoint, reader artifact, and immutable Foundation parent by SHA-256;
• restore the active candidate and its Context-native architecture strictly;
• keep the reader, candidate Foundation, Context Encoder/Access, and router frozen in this baseline;
• retain selected slot, source ID, confidence, and original C1 token ranges through the value adapter;
• run a small hand-authored composition smoke test without accessing sealed qualification data.

Validation completed:

• Exact component restoration and hash/lineage checks passed; Foundation and reader remained frozen.
• A two-case non-qualification smoke test passed 2/2 learned slot selections and 2/2 final generated answers (ZM8, K7T). No sealed data was accessed.
• The targeted reader and composition regression suite passed: 17 tests passed.
• PyTorch emitted its environment warning that optional NumPy initialization is unavailable; it did not prevent the load, inference, or test checks.

The smoke check only shows that these two hand-authored examples pass through the assembled interface. It does not qualify the overall model or broaden V1.2's tested scope. Machine-readable snapshots are stored in experiments/models/context-general-experimental-v1/evidence/ and the local run directory [retained internal evidence].

To restore and exercise the assembly from the repository root:

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

Disposition

The report recommends retaining the V1.2 reader as a bounded component of the versioned experimental assembly, with the overall model classified as needing improvement. The reader remains an immutable reference for a separately evaluated upgrade. This recommendation does not qualify full Parallel Context behavior, learned routing, multiple Context Drones, Global Context, movement, or general free-form reading.