Status: CONTEXT KEY-VALUE SLOT READ V1 QUALIFIED — STRUCTURAL GENERALIZATION PENDING Run: context-kv-slot-reader-v1-20260925 Code base: repository HEAD 4c0252bd227f44000b7b2fecd4f30abe210d80c9, with this experiment's implementation in the working tree Scope: one manually selected C1 using key=value records separated by |

Result

The new 74,369-parameter slot reader selected the requested record and copied its arbitrary unseen payload correctly on 1,498 of 1,501 fresh sealed examples (99.80%). The frozen Context-native candidate then generated the correct final answer on 1,466 of 1,501 examples (97.67%) after the selected value passed through the candidate's shuffled VALUE / NOTE / STATUS Context format. Wrong-source, absent-Context, and blocked-Context-Access controls each scored 0/256 exact answers. The Foundation and Context-native checkpoint hashes remained unchanged.

This result qualifies the tested structured key/value read path. It does not qualify raw Foundation binding over an unadapted multi-record C1 string or general Context Drone behavior.

Method revision

Context Address V2 had nearly qualified whole-record selection, but its independent start/end span head reached only 56.45% exact spans. A separate query-aware pointer-copy reader overfit 32 examples but scored 0/64 on unseen identities. A soft relevance-bias interface also failed its frozen-model oracle gate at 0/256. Those experiments did not show that the model could reliably connect a requested field identifier to its associated value.

For the current procedural key=value|... task, predicting arbitrary span boundaries is unnecessary. The replacement represents each candidate as a key paired with a value. A learned scorer compares the query with key tokens; the selected slot returns its paired value tokens and source coordinates. Visible delimiters define candidate slots but do not select one.

Open-source reference

The architecture is informed by Key-Value Memory Networks for Directly Reading Documents and its public TensorFlow research implementation. The paper's relevant idea is to separate query-to-key addressing from the associated value read. The repository describes itself as an implementation tested on bAbI; it does not establish arbitrary byte copying in EMMA. No external source code, weights, or packages were copied or imported. The EMMA module is an original implementation using that decomposition.

Architecture and training

ContextKeyValueMemoryReader uses the order-aware V2-A late-interaction selector over query tokens and key-only token representations. It has 74,369 trainable parameters, a 64-dimensional address space, and frozen 320-dimensional Foundation byte embeddings. Candidate slots are built from the actual runtime | and = bytes. The selected value is read by its stored token offsets and copied without a learned output vocabulary.

The deleted V2-A trained weights were not recovered. This candidate reinitialized the same selector architecture with the recorded seed and trained it on key-only slots.

Only the address module trained. The Foundation candidate, Context encoder, Context Access layers, and manual Context Router were frozen. Memory and AdaptiveWeightUnits were not loaded or attached to this experiment. Their artifacts were outside the code path and were not modified.

Training used AdamW, learning rate 8e-4, weight decay 1e-4, gradient clipping at 1.0, and batches sampled from six counterfactual groups. The seed was 20261627. The 32-example micro-set reached 32/32 at step 50. The bounded mixed two-to-four-record run stopped at step 400 after 21.03 seconds of recorded training/evaluation time. Peak VRAM and per-query latency were not captured.

Identity splits and sealed protocol

A fresh sealed set was created and hash-locked before model training. It contains 1,501 queries in 542 counterfactual Context groups: 500 two-record, 501 three-record, and 500 four-record examples. Key/value identities were kept disjoint across training, development, the immediate unseen probe, the sealed set, and historical identity hash firewalls. The sealed data file SHA-256 is [checksum retained in the private evidence record]; the locked identity-hash file SHA-256 is [checksum retained in the private evidence record]. Sealed labels were first evaluated only after the developmental gates passed.

The fresh sealed identities were not used to train or tune. A first evaluation invocation was interrupted before it wrote results; no metrics from that attempt were reviewed or used to change the candidate. The same locked data and unchanged candidate were then scored to completion.

Results

The oracle-record diagnostic and generated-answer controls use the same frozen Context Access and Foundation. Oracle record labels are diagnostic only. The qualification path used the learned record selector and its predicted value.

Runtime and restoration

The independent reader API accepts a query token sequence, one live C1 token sequence, and the frozen embedding table. It returns a KVReadResult with source ID, selected slot, probabilities, confidence, entropy, original record, key and value coordinates, and copied value bytes. build_value_access_view creates the selected-value Context using the trained VALUE / NOTE / STATUS record distribution and randomized field order while keeping the original C1 offsets attached as provenance. Its serialized bytes matched the existing frozen-candidate serializer on all 256 development inputs, and the Foundation generated the correct answer on 252/256. A one-field VALUE=... view without the trained distractor layout produced only 19/256 and was rejected.

A new process reloaded the active Context candidate and reader artifact and reproduced the selected slot, exact payload bytes, and generated outputs on a fixed 128-example subset. The runtime single-query API matched the batched reader's selected slot, payload, and original coordinates on all 128 cases. Generated answers and the reader tensor checksum matched exactly.

Integrity and artifacts

The reader file is 304,569 bytes and remains at [retained internal evidence]. The failed V2 span and pointer-reader binaries were not retained. The qualified Foundation control remains unchanged and the active Context-native candidate remains an experimental checkpoint, not a newly promoted Foundation release.

Focused tests: 12 passed across the KV slot reader, previous query-copy reader, and retrieval module test files. py_compile and git diff --check passed.

Qualification limits

The raw, unadapted Foundation path still cannot answer RETURN <arbitrary key> from the multi-record C1; its development result was 0/256. The successful path first selects a key/value slot, then adapts the returned value to the Foundation's existing VALUE Context format. That is a deliberate modular bridge and the current qualification boundary.

A proposed follow-up evaluates unseen serialization families, spacing, record lengths, and distractor layouts while keeping C1 manually selected. Broader structural validation of this single-source path is a prerequisite for subsequent Context Router, multi-Drone, Global Context, or movement studies.