Final status: CONTEXT QUERY-COPY READER V1 IMPLEMENTED — UNSEEN GENERALIZATION FAILED

This is a bounded prototype report. It is not Context Drone qualification and does not authorize routing, multiple Drones, Global Context, movement, or Memory integration.

Question and reference basis

The experiment asked whether a small separately trainable module could use a query to select and copy an arbitrary unseen value from one manually selected Context Drone C1, while the Context-native candidate and Foundation remained frozen.

The implementation was written from scratch using research ideas, not source code. BiDAF informed token-level bidirectional query/context interaction; Pointer Networks informed position-based outputs over variable-length input; pointer-generator research informed copying and coverage. The public AllenNLP Models repository includes BiDAF and is Apache-2.0, but is archived and in maintenance mode. No repository code, weights, or runtime dependency were imported. References and the exact design mapping are in context-query-copy-reader-reference.md.

Starting integrity

The active Context-native candidate restored strictly at 52,265,546 bytes and matched SHA-256 [checksum retained in the private evidence record]. Direct Context Transfer was 499/512 (97.46%) on the established seed. The immutable qualified Foundation parent matched [checksum retained in the private evidence record].

The reader was isolated from the Foundation and Context-native model. After training, the active candidate hash and Foundation parent hash remained unchanged. No Foundation, Context Encoder, Context Access, Memory, or AdaptiveWeightUnit parameters were trained.

Reader design

The input was frozen, token-level query and C1 state from the shared Context encoder, each with 320 hidden dimensions, plus the original UTF-8 byte IDs. The reader projected the streams separately, computed bidirectional query/context attention, modeled the query-aware C1 sequence, and used a small GRU pointer decoder to select C1 positions or EOS. The V1.1 candidate added a learned bounded exact-byte-match feature; the region-supervised variant also received a training-only loss encouraging attention mass in the correct record.

The V1.1 reader had 511,490 trainable parameters. Its byte output could only copy from C1; it had no answer vocabulary or string/dictionary lookup. The answer-only and region-supervised candidates used the same initialization seed, examples, AdamW optimizer, learning rate, batch size, and stopping budget. The region loss weight was 0.5.

Data and gates

Before training, the run locked a fresh sealed set of 1,501 examples across 2-, 3-, and 4-record counterfactual groups. The sealed file and its hash-only identity firewall were verified after the experiment; sealed examples and labels were never opened. Development, training, micro-overfit, and unseen identity pools were disjoint from those identities. All 832 training value spans decoded exactly from their recorded UTF-8 byte offsets.

The immediate gates were:

The reader could fit the training examples, but it did not transfer the query-to-record/value relation to unseen keys and values. No larger-data training, development qualification, sealed scoring, or Context Access integration was performed.

Failure localization

On the region-supervised unseen diagnostics, query-to-context attention's highest-mass record was correct for 40/64 (62.5%) examples. Mean attention mass on the requested key was 0.485 and on its answer value was 0.088. The pointer decoder emitted 94 of 213 byte steps (44.1%) from the correct value span, while producing 0/64 exact answers. The V1.1 answer-only diagnostic had 84/186 emitted pointer steps (45.2%) inside the target value; the V1.0 baseline had 97/226 (42.9%).

This is not a raw copy-capacity failure: the module learned to emit bytes from C1 and reached perfect micro-overfit. The failure is query-conditioned addressing and the transfer from a matched key to its associated value. The region loss did not make that relation generalize. This is consistent with the earlier retrieval-oracle result: frozen Context Access plus Foundation scored 0/256 even under correct relevance, so the existing generated-answer path is not a substitute for a capable learned reader.

The evidence does not show that all soft retrieval, pointer, or Context-native designs are impossible. It rejects only these bounded reader variants under this training and representation contract.

Controls and qualification not reached

The immediate unseen gate failed, so the experiment stopped before the development and sealed gates. Wrong-C1, absent-C1, query-swap, fresh-process restore, final answer integration, and post-reader Direct Context Transfer controls were not run. The copy-only architecture makes its output provenance observable, but that property is not behavioral qualification.

No reader checkpoint was retained. There is no promoted artifact. The new sealed set remains locked and unopened for a later candidate. The active Context-native candidate and qualified Foundation parent remain bit-identical to their recorded hashes.

Machine-readable evidence

• Starting-state and Direct Transfer check
• Fresh sealed-set manifest
• Split and label integrity
• V1.0 baseline micro/unseen result
• V1.0 pointer diagnostics
• V1.1 answer-only result
• V1.1 query/copy diagnostics
• V1.1 region-supervised result and diagnostics
• Final status evidence

Research priorities

The report does not support further identity-training expansion of the unchanged pointer decoder. A proposed comparison would make key-to-value relations explicit and evaluate record selection separately from value transfer. Candidate-record scoring and a jointly scored span or contiguous pointer path are proposed alternatives. The follow-up would train only the reader or address component, exclude literal key/value parsing, and keep sealed labels closed until the declared development gates pass.