[initial-d]

Hash-bucketed n-grams share row updates, so changing bucket capacity does not directly test independent learned identities. This change adds a training-only canonical n-gram registry and independent rows, with a matched hash control that has identical row parameters, effective initial vectors, training exposures and frozen Qwen backbone. Both variants use zero embeddings for unseen keys in the primary comparison; the same trained hash checkpoint is separately evaluated with ordinary open-vocabulary lookup.

The experiment is complete: 24 final models cover Qwen3-0.6B/Qwen3.5-4B, 1,024/20,877 edits and three seeds. At 20,877 edits, exact rows improve the seed/case-pooled harmonic ES/PS/NS score by 0.41/1.24 percentage points versus matched-coverage hashing and 2.29/4.05 versus native open-vocabulary hashing. All four pointwise 95% intervals are positive and locality improves. Two small-grid comparisons have intervals crossing zero and remain in the report. This is frozen-backbone MultiCounterFact editing; it does not establish pretraining, native serving/compression superiority or naturally occurring 99% cosine similarity.

The source-qualified archive includes all 24 reports, four vocabulary reports, two real-model preflights, runtime snapshots, paired test arrays, exact commands and complete bundle costs. A dependency-free verifier checks provenance, budgets and point estimates; an optional NumPy-only CPU check reproduces all eight bootstrap intervals. CI runs both. Model deltas, canonical token-key dictionaries and raw loss-level files remain outside the repository, with hashes retained.

Validation:

• Both real-model preflights passed, including every training key's initial vectors, 16 question prefixes per model, masked replay/gradients, gradient pooling and unchanged backbone hashes.
• All 24 training/evaluation runs completed successfully.
• python3 -m unittest discover -s tests: 250 tests, 125 passed and 125 optional-dependency skips locally.
• python3 -S docs/ngram_identity_evidence_20261107/verify.py: passed.
• python docs/ngram_identity_evidence_20261107/verify.py --bootstrap: all eight intervals reproduced on local CPU.
• Four archive-corruption, unmatched-budget and unsupported-positive-claim checks passed; git diff --check passed.