Date: 2026-10-01 Run: query-bound-adaptation-v25 Result: bounded query-bound attachment pass; research candidates retained, not promoted.

Rationale and protocol changes

V24's raw-record attachments failed role-balanced acquisition (80.73% and64.06% development). V25 changed the evidence interface rather than extending those optimizers. A deterministic structured reader resolves the requested field names, retains complete values and exact source spans, and copies only their first characters into canonical left/right operands. It never computes equality. A newly trained native attachment then produces Y/N through the frozen Foundation and its original head.

This deliberately narrows the neural job to the relation between selected operands. It does not demonstrate learned retrieval or broad Foundation reasoning. The original raw reader remains a separate unresolved neural capability. The reduction is a diagnostic mechanism and replaceable evidence boundary, not removal of the intended Context/weight architecture.

Research basis

The NAACL span-supervised attention paper reports improved compositional parsing from explicit span alignment. That motivates separating evidence binding from relation learning; EMMA's deterministic span reader is an EMMA implementation, not a reproduction of the paper's learned attention. Microsoft LoRA supplies the frozen-base/separate-delta mechanism reference. Stanford ReFT supplies an alternative position-specific representation-intervention reference; ReFT was researched but not implemented in V25. No external model weights, dependencies or source code were imported.

Locked experiment

Runbook and source/data hashes locked before fitting. Independent seeds 2501/2502 each own a new 28,800-parameter rank 10 native LoRA attachment at blocks 0/3 attention QKV/output. Foundation, original language head, V23/V24 and all prior modules are immutable. Optimizer scope includes only the new attachment's eight tensors. Zero-initialization and gradient-isolation checks pass.

Explicit request-local operation selects the attachment for the equality contract or the unchanged parent for native tasks. It is deterministic activation, not a trained intent model; every artifact remains resident. Invalid/missing/duplicate roles are rejected. The reader returns original field identity, full value, value coordinates and selected byte coordinates. Gold labels and provenance audit fields are not neural features.

Fresh disjoint unordered leading-symbol pairs define 64 training worlds,16 development worlds and24 sealed worlds. Each world enumerates8 binary assignments and6 ordered queries. Raw source examples are disjoint, role-label balanced and include shuffled layouts/suffixes. Training has 3072 source rows, development768, sealed1152, five-field stress1280 and longer-value stress192.

Canonical diversity is much smaller: training160 unique inputs, development56, sealed73, field stress8 and length stress15. Development shares24 equal-character canonical inputs with training; sealed shares25. Unequal symbol pairs are held out. The scored cases do not all represent novel semantic states, and repeated rows cannot be treated as independent statistical trials. Both fit diagnostics score 48/48 but represent only four unique canonical patterns.

Full fitting uses AdamW 0.001, no decay, batch 32 and clip 1; maximum 1200 updates, checkpoints 200/400/800/1200, selection by the lower development/renamed score and plateau stop after two nonimprovements. Both seeds require 95% development before sealed evaluation. No post-sealed optimization or hyperparameter sweep occurred.

Interruption and recovery

Seed2501 completed800 updates and stopped at the declared plateau; its selected snapshot was update 200. Seed2502 was interrupted after400 recorded updates, losing its in-memory selected weights. Its200/400 development scores were identical. Recovery replayed only the selected update 200 using the same initialization and minibatch seed. Ordinary and renamed predictions matched the original recorded checkpoint exactly before evaluation resumed.

This is an execution deviation: seed 2502 did not complete the originally planned later plateau checkpoints. The two runs therefore cannot be described as uninterrupted, matched training budgets. No data, gate, candidate-selection score or optimizer setting was changed. Original losses/checkpoints, interruption metadata, original locked runner source and recovery-source hash are preserved. Original source reconstruction matched the locked SHA-256 exactly. No completed seed 2501 fitting was repeated.

Controls

On fresh development: parent0/768; raw nearest-neighbor memory422/768 (54.95%); canonical memory384/768 (50%). Frozen V24 seed 2401 scores 616/768 (80.21%) on raw prompts but only 408/768 (53.13%) on canonical prompts. Thus normalization alone does not repair the previous attachment; newly learned weights are needed for this changed interface. These are diagnostic comparisons, not a matched raw-versus-canonical new-training causal study.

A deterministic equality function solves the task100% by definition. The learned module has not established necessity or superiority over that function. Majority/role-only control50%; the old all-three-uniform shortcut is no longer an input available to the relation module. The reader itself supplies selected first characters and therefore owns the query-binding part of the job.

Results

Complete sealed worlds: seed 2501 23/24, seed 2502 24/24.

Both selected snapshots fit 3072/3072 training rows. Development recurring equality inputs:384/384 for each seed. Novel unequal canonical inputs:372/384 for each. Development errors reduce to the ordered D,Q pair. Renamed sealed results match sealed scores. Role invariance largely comes from correct deterministic binding and canonicalization; it is not independent evidence of neural raw-role generalization.

Sealed canonical memory scores 744/1152 (64.58%). Seed2501 exceeds it by34.38 points and seed 2502 by35.42. Required-value interventions flip the target by changing an actual selected source byte; seed 2501 scores 97.92%, seed 2502 scores 100%. Irrelevant source-value changes preserve predictions exactly for both. Missing requested fields fail validation instead of fabricating operands. Field-count and value-length stress pass through the deterministic reader into the same bounded operand interface; they do not demonstrate a neural five-field/long-value reader.

The reader's6464 source cases all pass exact value/byte-span copying audits. Source identities and positions remain in the evidence packet, separate from canonical neural input.

Parent retention, persistence and organization

Inactive requests retain159/160 original curriculum tasks, identical to physically unmounted attachments. Forced activation drops performance to99/160 and138/160; neither is safe as an always-on attachment. Parent tensors are unchanged, so these are activation-interference failures rather than corrupted base weights.

Eight alternating and two concurrent read-only requests per candidate match isolated results. This does not prove simultaneous GPU kernels or whole-system thread safety. All four artifacts restore in a fresh process with identical outputs:48 diagnostic cases and768 development cases per respective artifact.

All 164 pre-existing binaries stayed byte-identical. Four new owned artifacts make168 total; the read-only audit preserves all 168. Focused tests: 22 passed, covering evidence coordinates, absence/renaming/counterfactual behavior, response alignment, task scope, gradients and native attachment behavior.

Sidecars categorize each artifact as a native Foundation attachment and record parent hash, target tensors, parameters, input/output/activation contracts, seed, data hash, optimizer scope, results and limits. Micro artifacts are fit diagnostics; full artifacts are bounded-qualified research candidates. No active registry entry, previous artifact or app route was replaced.

Protected Foundation SHA-256: [checksum retained in the private evidence record]. Training source-data hash: [checksum retained in the private evidence record]. Training canonical-data hash: [checksum retained in the private evidence record]. Sealed canonical hash: [checksum retained in the private evidence record]. All source/split/runbook hashes, candidate tensor hashes, original losses, recovery evidence and per-case predictions are under ignored [retained internal evidence].

Conclusions

The bounded claim passes: a separately owned attachment can acquire a simple relation over query-bound evidence while the Foundation stays frozen, transfer to held-out unequal symbol pairs, pass declared source interventions, and preserve original behavior when deactivated. The result supports making query/evidence interfaces explicit before asking a module to learn a relation.

It does not qualify arbitrary raw-language reading, Foundation reasoning attribution beyond its attached computation path, always-on retention, learned intent routing, sparse/remote weight loading, parallel attachment composition, serial block stacks or whole-Transformer communication. A programmed equality rule remains the stronger minimal solution to this toy relation. No active promotion is justified by this experiment.