nur-dev/strata-headquotient-q25

STRATA HEADQUOTIENT Q25 Reproducibility Artifacts

This is the compact evidence repository for:

N. Kadyrbek and M. Mansurova, "STRATA-HeadQuotient: Functional Localization of One Quarter of Global KV Heads with Typed Predicate-Graph Computation at 8k Context," submitted to Machine Learning and Knowledge Extraction, 2026.

It contains the functional taxonomy, all 384 audited key--value (KV) head classifications, all 6903 candidate-pair interactions, the static Q25 assignment, the expanded 470-document PPL confirmation, the untouched 2000-case semantic confirmation, frozen aggregate evaluations, and standard-library scripts that regenerate the article's numerical tables.

Scope

This repository supports five reproducibility tasks:

1. inspect the GLOBAL, LOCAL, and LOCAL_GRAPH taxonomy;
2. inspect every analysed KV-group classification and the typed program use cases;
3. verify the pre-specified candidate and Q25 assignment rules;
4. reproduce the reported summary statistics and tables;
5. verify release integrity and cite the artifacts.

It is an analytical reproducibility package. It does not duplicate raw Wikipedia text, third-party corpora, or model weights. The dense reference, three physical Q25 checkpoints, exact tokenized confirmation material, and minimal transitive source snapshot are public in the linked immutable nur-dev/strata-headquotient-q25 v1.2.0 model release, Hub commit 60b2ea8dc02c1b847faf3770105fecb2e9a74d7d. Their immutable SHA-256 identifiers also remain in raw/completion_manifest.json, raw/replication/q25_replications.json, raw/reproducibility/model_release.json, and the frontier files.

Principal Result Encoded by the Artifacts

Quantity	Frozen value
Audited global KV groups	384
Q25 groups physically localized	96 (25%)
LOCAL groups	81
LOCAL_GRAPH groups	15
Expanded true-8k aggregate PPL ratio	1.001463 (95% CI 1.001315--1.001611)
Expanded positions 4096--8191 PPL ratio	1.002006 (95% CI 1.001772--1.002237)
Untouched typed execution	1998/2000 (0.9990)
Untouched untyped execution	374/2000 (0.1870)
Model-core throughput relative to dense	0.9844
Cached decode ratio after 8160-token prefix	1.003x (batch 1) to 1.237x (batch 16)
Fresh Q25 adaptation/audit/selection campaigns	2; both reproduce PPL and typed-path criteria
Corrected matched-capacity attribution	Not passed in all three campaigns; every 95% interval includes zero
Full-history KV heads retained	75%
Token-KV bytes retained before graph state	78.125%
Q30 aggregate perplexity ratio	1.231682 (failed frontier)

The expanded ratios are relative to the frozen dense reference and use a paired document bootstrap over 470 unique held-out documents. Ordinary-text PPL was evaluated with graph reads disabled, so it tests global-to-local localization rather than graph necessity for PPL.

Computational Taxonomy

The unit of classification is one KV head in a standard MHA backbone, not a complete transformer layer. The tested model has 16 query heads and 16 KV heads per layer (one-to-one), so it is not GQA.

Mode	Historical state	Exported operation
GLOBAL	Full causal token KV history	Retained compact global attention
LOCAL	1024-token local-window KV	Local causal attention; no full-history token KV
LOCAL_GRAPH	Local-window KV plus a typed result	Local attention plus an event-scoped exact graph program; no full-history token KV

The machine-readable definitions and admission rules are in data/computational_taxonomy.csv. data/use_case_classifications.csv maps lexical continuation, event-role retrieval, natural carrier compilation, and deterministic evidence decisions to their actual computational components. Preprocessing and evidence control are marked NOT_APPLICABLE for KV mode because they occur outside the attention-group taxonomy.

Classification Rules

Candidate rule

A KV group enters the frozen 118-group candidate set exactly when:

graph_substitutability >= 0.95
and
typed_advantage >= 0.20

Joint Q25 rule

The selector chooses exactly 96 candidates under the signed pair-interaction objective and these pre-specified, hash-locked constraints:

maximum localized layers:                 10
minimum selected groups/localized layer:   2
minimum graph groups in layer 23:         10
minimum readers for each typed program:    2
retain a representative of every
  populated functional cluster:          yes

The mixed-integer result is feasible and has a reported 11.16% optimality gap; it is not claimed to be the unique optimum. A selected group is LOCAL_GRAPH only when it has an exported event-scoped program. Other selected groups are LOCAL. Every non-selected group is GLOBAL.

data/group_classifications.csv contains the complete 24-layer by 16-group grid, audit scores, candidate flag, final mode, primary program, active top-two program set, functional cluster, and selected node objective.

Repository Layout

.
|-- README.md
|-- CITATION.cff
|-- LICENSE
|-- MANIFEST.sha256
|-- VERSION
|-- config/                  # pre-specified experiment contracts
|-- data/
|   |-- computational_taxonomy.csv
|   |-- group_classifications.csv
|   |-- pair_interactions.csv
|   |-- reported_metrics.csv
|   `-- use_case_classifications.csv
|-- raw/                     # compact frozen aggregate and post-review outputs
|   `-- postreview/q25_confirmation.json
|   `-- reproducibility/q25_cached_decode_profile.json
|   `-- replication/q25_replications.json
|-- reproduced/
|   |-- summary.json
|   `-- tables.md
`-- scripts/
    |-- reproduce.py
    `-- verify.py

Absolute source-machine paths in copied manifests are normalized to source-artifact://.... Original byte sizes and SHA-256 identifiers are retained where they form part of the provenance record.

Reproduce the Tables

Python 3.10 or later is sufficient. No third-party package is required.

python scripts/reproduce.py --check
python scripts/verify.py

To regenerate the derived files after intentionally changing a source artifact:

python scripts/reproduce.py

This writes:

• reproduced/tables.md, containing the taxonomy, use cases, frontier, position/language results, causal controls, runtime, natural carrier, evidence, negative-result, and layer-distribution tables;
• reproduced/summary.json, containing the principal machine-readable summary;
• data/reported_metrics.csv, a long-form table suitable for the Hugging Face Dataset Viewer or pandas.

The verifier checks all 384 identities, the 118-group candidate rule, the 96-group Q25 assignment, the 81/15 mode split, active graph-program coverage, all 6903 unique candidate pairs, the hash-locked compact-layer constraints, both fresh Q25 campaigns, the corrected matched-capacity intervals, reproduced-file equality, absence of internal absolute paths, and every checksum in MANIFEST.sha256.

Dataset Viewer

The README metadata defines five Hugging Face configurations:

• group-classifications: complete 384-group audit and Q25 classification;
• pair-interactions: all 6903 signed candidate-pair measurements;
• computational-taxonomy: formal three-mode definitions;
• use-case-classifications: reproducible component/use-case mapping;
• reported-metrics: normalized article results and gates.

Example after publication:

from datasets import load_dataset

groups = load_dataset(
    "nur-dev/strata-headquotient-q25",
    "group-classifications",
    split="audit",
)
selected = groups.filter(lambda row: row["q25_selected"])

The reproduction scripts themselves do not require datasets.

Publish on Hugging Face

After reviewing the repository identifier and citation metadata:

python scripts/verify.py
hf auth login
hf repo create nur-dev/strata-headquotient-q25 --repo-type dataset
hf upload nur-dev/strata-headquotient-q25 . . --repo-type dataset

No GitHub mirror is required for this compact release. Use the immutable Hugging Face tag v1.2.0 for the corrected article version, and make later corrections in a new tagged release rather than rewriting the cited tag.

Evidence Boundaries

• Two fresh Q25 campaigns reinitialized the graph adapter, used disjoint diagnostic/interaction material, recomputed all 6903 candidate-pair and 128 triple interactions, solved new selections, and exported new checkpoints. The original 118-head candidate pre-screen remained fixed, so these are not independent rediscoveries from all 384 heads.
• The original matched-H0 comparison was capacity mismatched. Under the corrected comparator, all three marginal graph-effect confidence intervals include zero. The typed path is causally operative against zero/untyped/wrong-role controls, but graph takeover from localized attention is not established.
• The expanded confirmation contains 470 document-disjoint complete 8192-token windows and reports document-bootstrap intervals; language support is unequal.
• Ordinary PPL runs with graph reads disabled. PPL preservation and controlled typed handoff are separate findings.
• The 25% figure is the reduction in full-history KV heads. Including 1024-token local windows, token-KV bytes fall by 21.875% before graph state.
• All 15 graph-enabled heads are in final layer 23 by design; adapter depth was not searched.
• The result applies to one approximately 554-million-parameter backbone and 8k context.
• Q30 failed. The demonstrated replacement frontier is 25%, not 30% or 50%.
• Model-core throughput excludes synchronous natural carrier compilation. Compiler-inclusive throughput remained approximately 0.19 times dense.
• The cached profile uses one NVIDIA L40, identical prompts within each batch, and a fixed 32-token continuation; it is an implementation-specific model-core measurement rather than a production-serving claim.
• Peak allocation was effectively unchanged.
• The external carrier result is selective, not universal parsing.
• The evidence controller is deterministic given structured proof facts; it does not establish universal truth awareness.

Related Previous Release

The prior PAT-ER model artifacts are available at nur-dev/primitive-augmented-transformer. PAT-ER retained dense attention and studied typed side-state. This STRATA repository supports physical KV-head substitution and remains a separate dataset repository linked to the STRATA model release.

Hugging Face is the sole public maintenance location; a separate GitHub repository is not required. The analytical dataset, physical model release, and PAT-ER precursor are grouped in the authors' typed-semantic-state collection.

Citation

Until the article DOI is assigned, cite the manuscript and repository as:

@article{kadyrbek2026strata,
  author  = {Kadyrbek, Nurgali and Mansurova, Madina},
  title   = {STRATA-HeadQuotient: Functional Localization of One Quarter of Global KV Heads with Typed Predicate-Graph Computation at 8k Context},
  journal = {Machine Learning and Knowledge Extraction},
  year    = {2026},
  note    = {Manuscript submitted for publication}
}

@dataset{kadyrbek2026strata_artifacts,
  author    = {Kadyrbek, Nurgali and Mansurova, Madina},
  title     = {STRATA HEADQUOTIENT Q25 Reproducibility Artifacts},
  year      = {2026},
  version   = {1.2.0},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/nur-dev/strata-headquotient-q25}
}

Update the article entry with the final DOI, volume, issue, and article number after publication. CITATION.cff contains the same citation metadata.

License

The compact release is available under CC BY 4.0. Upstream corpora, omitted checkpoints, and third-party software retain their original terms.