Upload folder using huggingface_hub

CHANGELOG.md

@@ -1,5 +1,22 @@
 # Changelog
 
+## 1.2.0 - 2026-07-21
+
+- Added two fresh Q25 graph-adapter, interaction-audit, mixed-integer selection, and physical-export campaigns over the frozen 118-head candidate pool.
+- Added the cached prefill/decode benchmark and fixed-size global/local KV-cache implementation.
+- Corrected the dense H0 comparator to use exactly the graph-head/program capacity of each Q25 export.
+- Recorded that all three corrected marginal graph-effect intervals include zero; localization/PPL preservation and typed graph computation are therefore reported as separate findings.
+- Updated the manuscript title and generated tables to avoid attributing all 25% localization to the 14--15 graph-enabled heads.
+
+## 1.1.0 - 2026-07-21
+
+- Added a post-review confirmation over 470 document-disjoint 8192-token windows with paired document-bootstrap confidence intervals.
+- Added an untouched 2000-case semantic confirmation split beyond all adapter-training, interaction-audit, and frontier-selection indices.
+- Corrected the architecture description from grouped-query attention to standard 16-head multi-head attention.
+- Separated the 81 local-only heads from the 15 final-layer graph-enabled heads and documented that ordinary PPL runs with graph reads disabled.
+- Added exact token-KV accounting: 25% fewer full-history heads and 21.875% fewer token-KV bytes before graph state.
+- Narrowed the article title and claims to functional localization with typed semantic handoff.
+
 ## 1.0.1 - 2026-07-21
 
 - Flattened the 6903 interaction records to CSV for correct Hugging Face Dataset Viewer inference.

CITATION.cff

@@ -2,7 +2,7 @@ cff-version: 1.2.0
 message: "If you use these artifacts, cite the accompanying article and this repository release."
 title: "STRATA HEADQUOTIENT Q25 Reproducibility Artifacts"
 type: dataset
-version: 1.0.1
+version: 1.2.0
 date-released: 2026-07-21
 license: CC-BY-4.0
 authors:
@@ -16,14 +16,14 @@ repository-code: "https://huggingface.co/datasets/nur-dev/strata-headquotient-q2
 url: "https://huggingface.co/datasets/nur-dev/strata-headquotient-q25"
 keywords:
   - attention substitution
-  - grouped-query attention
+  - multi-head attention
   - knowledge graphs
   - long-context language models
   - neuro-symbolic learning
   - reproducibility
 preferred-citation:
   type: article
-  title: "STRATA: Typed Predicate-Graph Programs Enable Physical Replacement of One Quarter of Global KV Groups at 8k Context"
+  title: "STRATA-HeadQuotient: Functional Localization of One Quarter of Global KV Heads with Typed Predicate-Graph Computation at 8k Context"
   authors:
     - family-names: Kadyrbek
       given-names: Nurgali

MANIFEST.sha256

@@ -1,16 +1,16 @@
 ce29a99003918cb7996582b22985d768809d5591a741cadf9baefb573b1b67d8  .gitattributes
-971020c1a10315e4a8e3fb675aa99f992b2436a378d99e3f3d60f4177557aa4e  CHANGELOG.md
-50fa2cb096caa9eb2146af5518ced52107d268a8e87d3ff81664598a6613ef59  CITATION.cff
+bc01afac4955df90ee25802da4f353c9b8909497b95206648ac8cd9dc25c8ef5  CHANGELOG.md
+cf0e7509796a78f7bee2bbee9314ae498212356786d71c9c9e8c21817493dee0  CITATION.cff
 506798da85f78c2abe11c01492633ca1b3e880f0e83211c1b70be885c1c23825  LICENSE
-c6a84864ce5d0446a81012b2db4bf9cd6db7246ee87d5ed750cc452262c5b098  README.md
-44e161e4495cac2cf7858043e9e6418e9579f0ddcfae826f9a372622968ce066  VERSION
+4177f391b8c1d64fc129a16d0b09d50fe6842104e3331b651e0905eda7212a1a  README.md
+1e5b51cde515396a9fa762909cf8ca6584ccc564b325d2eebeea76175fe95c4d  VERSION
 54bbf7c867415f51366bc8e7aeb29c417ebcb9cd3e38257483ed8663944efeca  config/headquotient_v1.json
 243d5f4e286e2a9ca874dac900f76f1c147d7a79794fe341dfc57a9d1ff965a5  config/headquotient_v1_1.json
-b778715034ee1f6edc77c738422c9f0e83eedc0c6a2017f044b13dee86d81e50  data/computational_taxonomy.csv
+88879d8e9c5cf5904a5d85a16b002c4613afe0c5ef1bb8dc4f84e775ae99e59f  data/computational_taxonomy.csv
 0ced2d36c371d4e3b1b459d6e785cff610c18cbe3071900767accfd33e423378  data/group_classifications.csv
 9172e2383da63ba1a927d3988fb963a74d17002f43465eec6edc09ee706ef137  data/pair_interactions.csv
-d323a0fd84bc8e4dbf41ce9482e5ac12becd084b6d59d0a4196c7b7ed8f2b4dc  data/reported_metrics.csv
-cb269bca3db9f419f43aef109a6b7db316fa833a399ccf78b84ef1d5127edd05  data/use_case_classifications.csv
+2b92bed185c6386851e670599a1ca6581689546e86d340b7c617bdcc85f9e9b0  data/reported_metrics.csv
+7deb4ebea81552f9a3c0c4419dfb55e917cbfc625a6d0df323f00e84a201dbb3  data/use_case_classifications.csv
 e4b8c136dcfe4e7b503b2bc3fac02a0af5eef2f5a24f976db5cd4b29434e30bb  raw/completion_manifest.json
 211b2d42f1b5b6d337e04b7b6065ab9133379d900685df6e6e2e246ee7b2eb1c  raw/external/evidence_decisions.json
 91b23e3453e8b4e138dec5bce424c8c1fc52c4e4f7b86997ce35d1780963d7ca  raw/external/heldout_meta.json
@@ -28,8 +28,17 @@ e6f1b776e6701e6614329472e8fb760680151fccf0df2374e51e7821ab174f8b  raw/negative/d
 183da134902e50fb0c1188fc64ab687dca8301cb46ec6f67cf9f9e3c013547de  raw/negative/sparse_residual.json
 fbf59fa680576de8d988e26db4fdc826a5ec89e9638ef69e33fc510682b5d14b  raw/negative/whole_layer_and_serving.json
 c22a7cec3a5445a7abbc6d777dab7c8a7cb2325b81e592b3abfc15e1cb04d12e  raw/negative/whole_layer_true8k.json
+1ec75c5e2c1873e90379433175ebcd1fef87202effa2b639a995407f3322dedc  raw/postreview/q25_confirmation.json
+00aa763a46ce1572821239bec2721fcd5e4946d09992b36079adb87b1b511c73  raw/replication/q25_replications.json
+78671af18510c466b8a9b8128c0ad348f1eebbb5cfbc8b0c5cf610d4c544a768  raw/reproducibility/model_release.json
+fec2e25194ef49faa5a5d99fabc9c836f9407075c321a16bfc6ef807f1460654  raw/reproducibility/q25_cached_decode_profile.json
+255d14f9399f27c1f7e4a8f8dec93917714162732a3a46e9fef9873d227d1253  raw/reproducibility/q25_matched_h0_reanalysis.json
 b54e87ccd3d7cb2c3eb1659e545fef4a6952995d41b82bbb93545d6e0c8f4a49  raw/selection/q25_plan.json
-dd7a2f4bd592191a4ae05147ef42ef1e8fcf4444daa354f233fbca68b3e5a284  reproduced/summary.json
-4329adcdaa8928e37e96706949be20c17c306d70c3d8ddf508f03e78565466ac  reproduced/tables.md
-89a63edaac7c92e397fe0b2217ed180f58de4c9253415a582acc187a05d2ed10  scripts/reproduce.py
-543f6fae1a6d7f7516079bd84fa934f44dd9cb8cb2d21912b2ab450dcc5d0a45  scripts/verify.py
+426022fb3a7727a8f5a684810b79960b0422fac8848eefedeaae91feafa770d3  reproduced/summary.json
+6b509b38d9bb06caae2648c95cb9954a78745ab27c839861e0531f2d3ae08d2e  reproduced/tables.md
+f8e973e7cfff050c68b2b413efde2388abaf2e260fc03dc6583803d1b24ed807  scripts/__pycache__/reproduce.cpython-311.pyc
+4afbf0fe71e85ef448c09dc01d13be5d426afc8a76b233e12e4593c9c3306e55  scripts/__pycache__/reproduce.cpython-312.pyc
+10e1d93ed057cc9f0777ad6e6379d5776ea24a199913bf96181806185423eea1  scripts/__pycache__/verify.cpython-311.pyc
+e28bdecdcf4492056211c8395bf5feda44016a0070e0ac9fe84be5bf487799cc  scripts/__pycache__/verify.cpython-312.pyc
+9d9a01289136d0f1fe58f889a3e3828f7d8648a2787fcd9231027099c3b5214e  scripts/reproduce.py
+70771e123de64b72dff9a39df740f13416b8ec625000f3449dbb0313f271988f  scripts/verify.py

README.md

@@ -17,7 +17,7 @@ source_datasets:
 tags:
   - tabular
   - long-context-language-modeling
-  - grouped-query-attention
+  - multi-head-attention
   - knowledge-graphs
   - neuro-symbolic-learning
   - causal-intervention
@@ -50,9 +50,9 @@ configs:
 
 This is the compact evidence repository for:
 
-> N. Kadyrbek and M. Mansurova, "STRATA: Typed Predicate-Graph Programs Enable Physical Replacement of One Quarter of Global KV Groups at 8k Context," submitted to *Machine Learning and Knowledge Extraction*, 2026.
+> 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 global key--value (KV) group classifications, all 6903 candidate-pair interactions, the static Q25 assignment, frozen aggregate evaluations, and standard-library scripts that regenerate the article's numerical tables.
+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
 
@@ -60,11 +60,11 @@ 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 registered candidate and Q25 assignment rules;
+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 contain raw Wikipedia text, third-party corpora, or the approximately 1.1 GB exported checkpoint. Those objects are unnecessary for regenerating the published tables and cannot all be redistributed under this repository's license. Their immutable SHA-256 identifiers remain in `raw/completion_manifest.json` and the frontier files.
+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](https://huggingface.co/nur-dev/strata-headquotient-q25/tree/v1.2.0), 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
 
@@ -74,25 +74,29 @@ It is an **analytical reproducibility package**. It does not contain raw Wikiped
 | Q25 groups physically localized | 96 (25%) |
 | `LOCAL` groups | 81 |
 | `LOCAL_GRAPH` groups | 15 |
-| True-8k aggregate perplexity ratio | 0.999939 |
-| Positions 4096--8191 perplexity ratio | 0.999389 |
-| Typed execution | 0.9995 |
-| Untyped execution | 0.1790 |
+| 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 |
-| Global KV state retained | 75% |
+| 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 ratios are relative to the frozen dense reference. Ratios below one are treated as parity, not as evidence of general improvement.
+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 grouped-query KV group, not a complete transformer layer.
+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 historical global KV |
-| `LOCAL_GRAPH` | Local-window KV plus a typed result | Local attention plus an event-scoped exact graph program; no historical global KV |
+| `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/computational_taxonomy.csv). [`data/use_case_classifications.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.
 
@@ -110,7 +114,7 @@ typed_advantage >= 0.20
 
 ### Joint Q25 rule
 
-The selector chooses exactly 96 candidates under the signed pair-interaction objective and these registered constraints:
+The selector chooses exactly 96 candidates under the signed pair-interaction objective and these pre-specified, hash-locked constraints:
 
 ```text
 maximum localized layers:                 10
@@ -134,14 +138,17 @@ The mixed-integer result is feasible and has a reported 11.16% optimality gap; i
 |-- LICENSE
 |-- MANIFEST.sha256
 |-- VERSION
-|-- config/                  # registered experiment contracts
+|-- 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 outputs
+|-- 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
@@ -173,7 +180,7 @@ This writes:
 - [`reproduced/summary.json`](reproduced/summary.json), containing the principal machine-readable summary;
 - [`data/reported_metrics.csv`](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 registered compact-layer constraints, reproduced-file equality, absence of internal absolute paths, and every checksum in `MANIFEST.sha256`.
+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
 
@@ -211,24 +218,29 @@ 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 `v1.0.1` for the article version, and make later corrections in a new tagged release rather than rewriting the cited tag.
+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
 
-- The Q25 adaptation and final selection were run once; this package does not convert evaluation resampling into independent training replication.
-- The confirmation language-model set contains eight complete 8192-token windows. Per-language partitions are reported, but wider replication is required.
+- 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`](https://huggingface.co/nur-dev/primitive-augmented-transformer). PAT-ER retained dense attention and studied typed side-state. This STRATA repository supports physical KV-group substitution and should remain a separate dataset repository. There is no prior PAT-ER Hugging Face collection identifier in the published paper to reuse.
+The prior PAT-ER model artifacts are available at [`nur-dev/primitive-augmented-transformer`](https://huggingface.co/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 sufficient for maintaining this compact release; a separate GitHub repository is not required. If the full Q25 checkpoint is later made public, it can be placed in a linked Hugging Face model repository and grouped with this dataset in a STRATA collection without changing the evidence package.
+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
 
@@ -237,7 +249,7 @@ Until the article DOI is assigned, cite the manuscript and repository as:
 ```bibtex
 @article{kadyrbek2026strata,
   author  = {Kadyrbek, Nurgali and Mansurova, Madina},
-  title   = {STRATA: Typed Predicate-Graph Programs Enable Physical Replacement of One Quarter of Global KV Groups at 8k Context},
+  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}
@@ -247,7 +259,7 @@ Until the article DOI is assigned, cite the manuscript and repository as:
   author    = {Kadyrbek, Nurgali and Mansurova, Madina},
   title     = {STRATA HEADQUOTIENT Q25 Reproducibility Artifacts},
   year      = {2026},
-  version   = {1.0.1},
+  version   = {1.2.0},
   publisher = {Hugging Face},
   url       = {https://huggingface.co/datasets/nur-dev/strata-headquotient-q25}
 }

VERSION

@@ -1 +1 @@
-1.0.1
+1.2.0

data/computational_taxonomy.csv

@@ -1,4 +1,4 @@
 mode,unit_of_assignment,historical_state,admission_rule,exported_computation,stores_global_token_kv,requires_typed_graph_program
-GLOBAL,Global KV group,Full causal token KV history,"Retain when late-context lexical or discourse utility is material, graph substitution is inadequate, or cluster-retention constraints require the group.",Compact global grouped-query projection with historical KV,true,false
-LOCAL,Global KV group before export; local group after export,1024-token local-window KV,"The group passes the registered candidate rule, is selected jointly under the interaction and cluster constraints, and has no event-scoped graph program.",Local causal attention with no historical global KV,false,false
-LOCAL_GRAPH,Global KV group before export; local graph group after export,1024-token local-window KV plus an event-scoped typed result,"The group passes the registered candidate rule, is selected jointly, has a valid typed program assignment, and shows relation-specific causal handoff without wrong-event or random benefit.",Local causal attention plus exact event-scoped typed graph read with no historical global KV,false,true
+GLOBAL,Global KV head,Full causal token KV history,"Retain when late-context lexical or discourse utility is material, graph substitution is inadequate, or cluster-retention constraints require the head.",Compact global multi-head-attention projection with historical KV,true,false
+LOCAL,Global KV head before export; local head after export,1024-token local-window KV,"The head passes the pre-specified candidate rule, is selected jointly under the interaction and cluster constraints, and has no event-scoped graph program.",Local causal attention with no full-history token KV,false,false
+LOCAL_GRAPH,Global KV head before export; local graph head after export,1024-token local-window KV plus an event-scoped typed result,"The head passes the pre-specified candidate rule, is selected jointly, has a valid typed program assignment, and shows relation-specific causal handoff without wrong-event or random benefit.",Local causal attention plus exact event-scoped typed graph read with no full-history token KV,false,true

data/reported_metrics.csv

@@ -23,30 +23,67 @@ frontier,Q30,aggregate_ppl_ratio,1.231681966831007,ratio,<=1.03,raw/frontier/q30
 frontier,Q30,late_ppl_ratio,1.3780701501669157,ratio,<=1.05,raw/frontier/q30.json
 frontier,Q30,typed_execution,1.0,accuracy,>=0.95,raw/frontier/q30.json
 frontier,Q30,core_throughput_ratio,0.9392739975057925,ratio,>=0.95,raw/frontier/q30.json
-q25_ppl,0-2048,ppl_ratio,1.00007699819542,ratio,<=1.03 aggregate; <=1.05 late,raw/frontier/q25.json
-q25_ppl,2048-4096,ppl_ratio,1.0009002031766665,ratio,<=1.03 aggregate; <=1.05 late,raw/frontier/q25.json
-q25_ppl,4096-8192,ppl_ratio,0.9993893766212806,ratio,<=1.03 aggregate; <=1.05 late,raw/frontier/q25.json
-q25_ppl,aggregate,ppl_ratio,0.9999388623918676,ratio,<=1.03 aggregate; <=1.05 late,raw/frontier/q25.json
-q25_language,ar,ppl_ratio,1.0014931180241606,ratio,<=1.05,raw/frontier/q25.json
-q25_language,de,ppl_ratio,1.0014569436456722,ratio,<=1.05,raw/frontier/q25.json
-q25_language,en,ppl_ratio,1.0005086972746728,ratio,<=1.05,raw/frontier/q25.json
-q25_language,es,ppl_ratio,0.9979950617397495,ratio,<=1.05,raw/frontier/q25.json
-q25_language,zh,ppl_ratio,0.9974994495185546,ratio,<=1.05,raw/frontier/q25.json
+q25_ppl_postreview,aggregate,ppl_ratio,1.0014634161587161,ratio,upper 95% CI <1.03,raw/postreview/q25_confirmation.json
+q25_ppl_postreview,aggregate,ppl_ratio_lower_95,1.0013146459959361,ratio,,raw/postreview/q25_confirmation.json
+q25_ppl_postreview,aggregate,ppl_ratio_upper_95,1.0016112211164634,ratio,<1.03,raw/postreview/q25_confirmation.json
+q25_ppl_postreview,0-2048,ppl_ratio,1.000426001591288,ratio,<=1.05 late,raw/postreview/q25_confirmation.json
+q25_ppl_postreview,2048-4096,ppl_ratio,1.0014170876799071,ratio,<=1.05 late,raw/postreview/q25_confirmation.json
+q25_ppl_postreview,4096-8192,ppl_ratio,1.0020058363001705,ratio,<=1.05 late,raw/postreview/q25_confirmation.json
+q25_language_postreview,ar,ppl_ratio,1.0007953605165114,ratio,<=1.05,raw/postreview/q25_confirmation.json
+q25_language_postreview,de,ppl_ratio,1.001623872638894,ratio,<=1.05,raw/postreview/q25_confirmation.json
+q25_language_postreview,en,ppl_ratio,1.0012897886831786,ratio,<=1.05,raw/postreview/q25_confirmation.json
+q25_language_postreview,es,ppl_ratio,1.0012459933881799,ratio,<=1.05,raw/postreview/q25_confirmation.json
+q25_language_postreview,zh,ppl_ratio,1.0015464123814795,ratio,<=1.05,raw/postreview/q25_confirmation.json
 causal,correct,accuracy,0.9995,accuracy,,raw/frontier/q25.json
 causal,correct,mean_margin_vs_zero,0.2550610899925232,margin,,raw/frontier/q25.json
+causal_confirmation,correct,accuracy,0.999,accuracy,,raw/postreview/q25_confirmation.json
 causal,untyped,accuracy,0.179,accuracy,,raw/frontier/q25.json
 causal,untyped,mean_margin_vs_zero,-0.0006964647327549756,margin,,raw/frontier/q25.json
+causal_confirmation,untyped,accuracy,0.187,accuracy,,raw/postreview/q25_confirmation.json
 causal,wrong_role,accuracy,0.0,accuracy,,raw/frontier/q25.json
 causal,wrong_role,mean_margin_vs_zero,-0.2559249997138977,margin,,raw/frontier/q25.json
+causal_confirmation,wrong_role,accuracy,0.0,accuracy,,raw/postreview/q25_confirmation.json
 causal,wrong_event,accuracy,0.1145,accuracy,,raw/frontier/q25.json
 causal,wrong_event,mean_margin_vs_zero,0.0,margin,,raw/frontier/q25.json
+causal_confirmation,wrong_event,accuracy,0.1335,accuracy,,raw/postreview/q25_confirmation.json
 causal,random,accuracy,0.1145,accuracy,,raw/frontier/q25.json
 causal,random,mean_margin_vs_zero,0.0,margin,,raw/frontier/q25.json
+causal_confirmation,random,accuracy,0.1335,accuracy,,raw/postreview/q25_confirmation.json
 causal,zero,accuracy,0.1145,accuracy,,raw/frontier/q25.json
 causal,zero,mean_margin_vs_zero,0.0,margin,,raw/frontier/q25.json
+causal_confirmation,zero,accuracy,0.1335,accuracy,,raw/postreview/q25_confirmation.json
+causal_attribution,correct,matched_capacity_marginal_mean,-8.118145342450589e-05,margin,95% CI must exclude zero,raw/replication/q25_replications.json
+causal_attribution,correct,matched_capacity_marginal_lower_95,-0.00023968867026269436,margin,,raw/replication/q25_replications.json
+causal_attribution,correct,matched_capacity_marginal_upper_95,7.799099694238976e-05,margin,,raw/replication/q25_replications.json
+replication,original,expanded_ppl_ratio,1.0014634161587161,ratio,upper 95% CI <1.03,raw/replication/q25_replications.json
+replication,original,expanded_typed,0.999,accuracy,>=0.95,raw/replication/q25_replications.json
+replication,original,expanded_untyped,0.187,accuracy,,raw/replication/q25_replications.json
+replication,original,matched_capacity_marginal_mean,-8.118145342450589e-05,margin,95% CI must exclude zero,raw/replication/q25_replications.json
+replication,fresh_seed_20260777,expanded_ppl_ratio,1.0013051380076936,ratio,upper 95% CI <1.03,raw/replication/q25_replications.json
+replication,fresh_seed_20260777,expanded_typed,0.9935,accuracy,>=0.95,raw/replication/q25_replications.json
+replication,fresh_seed_20260777,expanded_untyped,0.191,accuracy,,raw/replication/q25_replications.json
+replication,fresh_seed_20260777,matched_capacity_marginal_mean,0.00013299021520651877,margin,95% CI must exclude zero,raw/replication/q25_replications.json
+replication,fresh_seed_20260791,expanded_ppl_ratio,1.0014123009914193,ratio,upper 95% CI <1.03,raw/replication/q25_replications.json
+replication,fresh_seed_20260791,expanded_typed,0.994,accuracy,>=0.95,raw/replication/q25_replications.json
+replication,fresh_seed_20260791,expanded_untyped,0.192,accuracy,,raw/replication/q25_replications.json
+replication,fresh_seed_20260791,matched_capacity_marginal_mean,-3.4817545383702964e-05,margin,95% CI must exclude zero,raw/replication/q25_replications.json
 runtime,Q25,throughput_ratio,0.9843666476918135,ratio,>=0.95,raw/frontier/q25.json
 runtime,Q25,peak_allocation_ratio,0.9997919067191773,ratio,<=1.00,raw/frontier/q25.json
 runtime,Q25,global_kv_retained,0.75,fraction,<=0.80,raw/frontier/q25.json
+runtime,Q25,token_kv_bytes,629145600,bytes,,raw/postreview/q25_confirmation.json
+runtime,Q25,token_kv_reduction,0.21875,fraction,,raw/postreview/q25_confirmation.json
+cached_runtime,batch_1,decode_throughput_ratio,1.0025517041888306,ratio,,raw/reproducibility/q25_cached_decode_profile.json
+cached_runtime,batch_1,prefill_throughput_ratio,0.9263008429955861,ratio,,raw/reproducibility/q25_cached_decode_profile.json
+cached_runtime,batch_1,persistent_kv_ratio,0.78125,ratio,,raw/reproducibility/q25_cached_decode_profile.json
+cached_runtime,batch_4,decode_throughput_ratio,1.1784115948125533,ratio,,raw/reproducibility/q25_cached_decode_profile.json
+cached_runtime,batch_4,prefill_throughput_ratio,1.0179303066424417,ratio,,raw/reproducibility/q25_cached_decode_profile.json
+cached_runtime,batch_4,persistent_kv_ratio,0.78125,ratio,,raw/reproducibility/q25_cached_decode_profile.json
+cached_runtime,batch_8,decode_throughput_ratio,1.2187408554168933,ratio,,raw/reproducibility/q25_cached_decode_profile.json
+cached_runtime,batch_8,prefill_throughput_ratio,1.0150799420381769,ratio,,raw/reproducibility/q25_cached_decode_profile.json
+cached_runtime,batch_8,persistent_kv_ratio,0.78125,ratio,,raw/reproducibility/q25_cached_decode_profile.json
+cached_runtime,batch_16,decode_throughput_ratio,1.2367898816648804,ratio,,raw/reproducibility/q25_cached_decode_profile.json
+cached_runtime,batch_16,prefill_throughput_ratio,0.9938475982700953,ratio,,raw/reproducibility/q25_cached_decode_profile.json
+cached_runtime,batch_16,persistent_kv_ratio,0.78125,ratio,,raw/reproducibility/q25_cached_decode_profile.json
 external_natural,external_v3,raw_execution,0.9105,rate,,raw/external/natural_carrier.json
 external_natural,external_v3,authoritative_execution,0.9826689774696707,rate,,raw/external/natural_carrier.json
 external_natural,external_v3,authoritative_coverage,0.7972814623857511,rate,,raw/external/natural_carrier.json
@@ -58,9 +95,9 @@ external_natural,external_v3,r12_untyped_retrieval,0.49609375,rate,,raw/external
 external_natural,external_v3,r12_wrong_role,0.0078125,rate,,raw/external/natural_carrier.json
 external_natural,external_v3,r12_wrong_event,0.046875,rate,,raw/external/natural_carrier.json
 external_natural,external_v3,r12_random_carrier,0.0,rate,,raw/external/natural_carrier.json
-evidence,registered_800,macro_decision_accuracy,1.0,accuracy,>=0.95,raw/external/evidence_decisions.json
-evidence,registered_800,proof_correctness,1.0,accuracy,=1.00,raw/external/evidence_decisions.json
-evidence,registered_800,paired_identical_agreement,1.0,agreement,=1.00,raw/external/evidence_decisions.json
+evidence,hash_locked_800,macro_decision_accuracy,1.0,accuracy,>=0.95,raw/external/evidence_decisions.json
+evidence,hash_locked_800,proof_correctness,1.0,accuracy,=1.00,raw/external/evidence_decisions.json
+evidence,hash_locked_800,paired_identical_agreement,1.0,agreement,=1.00,raw/external/evidence_decisions.json
 negative_result,Whole-layer graph-only R12,reported_result,aggregate 1.342165; late 1.747375,text,Failed true-8k preservation,raw/negative or raw/frontier/q30.json
 negative_result,Sparse residual oracle,reported_result,aggregate 1.314621; late 1.679730,text,Failed oracle gate,raw/negative or raw/frontier/q30.json
 negative_result,Fixed-basis kernel audit,reported_result,window overlap 0.705307; language overlap 0.697986,text,Rejected before training,raw/negative or raw/frontier/q30.json

data/use_case_classifications.csv

@@ -1,5 +1,5 @@
 use_case_id,scope,use_case,assigned_component,computational_method,kv_mode,typed_program,decision_rule,evaluation_scope
-UC01,Model core,Late-context lexical and discourse continuation,Retained global KV groups,Dense grouped-query causal attention,GLOBAL,NO_GRAPH_READ,"Retain when late-position perplexity utility is material or typed graph recovery is inadequate.",True-8k language-model evaluation
+UC01,Model core,Late-context lexical and discourse continuation,Retained global KV heads,Dense multi-head causal attention,GLOBAL,NO_GRAPH_READ,"Retain when late-position perplexity utility is material or typed graph recovery is inadequate.",True-8k language-model evaluation
 UC02,Model core,Recent token and phrase continuation,Localized non-graph KV groups,Local causal attention,LOCAL,NO_GRAPH_READ,"Localize only as part of a jointly passing subset; no semantic graph claim is attached to this mode.",True-8k language-model evaluation
 UC03,Model core,Agent or ARG0 retrieval,Event-scoped graph group,Exact typed graph execution,LOCAL_GRAPH,READ_EVENT_ARG0,"Use only for the current resolved event and a valid ARG0 address.",Controlled graph-program cases
 UC04,Model core,Patient or ARG1 retrieval,Event-scoped graph group,Exact typed graph execution,LOCAL_GRAPH,READ_EVENT_ARG1,"Use only for the current resolved event and a valid ARG1 address.",Controlled graph-program cases
@@ -7,4 +7,4 @@ UC05,Model core,Recipient or ARG2 retrieval,Event-scoped graph group,Exact typed
 UC06,Model core,Event time retrieval,Event-scoped graph group,Exact typed graph execution,LOCAL_GRAPH,READ_EVENT_TIME,"Use only for the current resolved event and a typed time relation.",Controlled graph-program cases
 UC07,Model core,Event location retrieval,Event-scoped graph group,Exact typed graph execution,LOCAL_GRAPH,READ_EVENT_LOCATION,"Use only for the current resolved event and a typed location relation.",Controlled graph-program cases
 UC08,Preprocessing,Natural text to authoritative carrier facts,Selective carrier compiler,Learned event-relation-argument resolution followed by deterministic typed closure,NOT_APPLICABLE,NOT_APPLICABLE,"Emit a fact only when event, relation, and argument prediction sets are singleton and the frame/valency proof is valid.",External-v3 natural carrier evaluation
-UC09,Postprocessing,Four-way evidence-state decision,Evidence controller,Deterministic proof-state execution,NOT_APPLICABLE,READ_SUPPORT_OR_CONTRADICTION,"ANSWER for support only; REJECT for contradiction only; NOT_SURE for conflict or structural disagreement; IDK for absent required evidence or unresolved structure.",Registered 800-case structured-evidence evaluation
+UC09,Postprocessing,Four-way evidence-state decision,Evidence controller,Deterministic proof-state execution,NOT_APPLICABLE,READ_SUPPORT_OR_CONTRADICTION,"ANSWER for support only; REJECT for contradiction only; NOT_SURE for conflict or structural disagreement; IDK for absent required evidence or unresolved structure.",Hash-locked 800-case structured-evidence evaluation

raw/replication/q25_replications.json

@@ -0,0 +1,229 @@
+{
+  "all_corrected_matched_h0_passed": false,
+  "all_expanded_typed_at_least_0_95": true,
+  "all_expanded_upper_below_1_03": true,
+  "all_invalid_programs_zero": true,
+  "all_localization_and_typed_path_passed": true,
+  "fresh_replications": 2,
+  "program": "STRATA-HEADQUOTIENT-Q25-INDEPENDENT-REPLICATION",
+  "runs": [
+    {
+      "causal_start": 2048,
+      "checkpoint_sha256": "51120c7ecca5234a4e7cee424c199582e444eb04e434a2d7f08263e3ccc74a90",
+      "config_sha256": "c4a3742f240a67380913554dbfc6efd4862a41fd1ce38bc08fd1e6952730a9c9",
+      "confirmation_stage_opened": true,
+      "confirmation_windows": {
+        "count": 8,
+        "start": 112
+      },
+      "corrected_marginal_correct": {
+        "confidence": 0.95,
+        "lower": -0.00023968867026269436,
+        "mean": -8.118145342450589e-05,
+        "samples": 1000,
+        "upper": 7.799099694238976e-05
+      },
+      "corrected_matched_h0_passed": false,
+      "diagnostic_windows": {
+        "count": 8,
+        "start": 96
+      },
+      "expanded_documents": 470,
+      "expanded_ppl_ratio": 1.0014634161587161,
+      "expanded_ppl_ratio_lower": 1.0013146459959361,
+      "expanded_ppl_ratio_upper": 1.0016112211164634,
+      "expanded_random": 0.1335,
+      "expanded_typed": 0.999,
+      "expanded_untyped": 0.187,
+      "expanded_wrong_event": 0.1335,
+      "expanded_wrong_role": 0.0,
+      "expanded_zero": 0.1335,
+      "frontier_late_ppl_ratio": 0.9993893766212806,
+      "frontier_ppl_ratio": 0.9999388623918676,
+      "frontier_throughput_ratio": 0.9843666476918135,
+      "frontier_typed": 0.9995,
+      "frontier_untyped": 0.179,
+      "initialization": "registered_legacy_adapter",
+      "interaction_audit_sha256": "7a844ea72203f0511b99ab52b71d8b2c62dc305cbbb599f144fba84ebd7ecc5c",
+      "interaction_pairs": 6903,
+      "interaction_triples": 128,
+      "interaction_window": 110,
+      "invalid_programs": 0,
+      "local_graph_groups": 15,
+      "local_groups": 81,
+      "localization_and_typed_path_passed": true,
+      "program": "STRATA-HEADQUOTIENT-v1.1",
+      "registered_result_passed": true,
+      "scoped_correct": 0.97,
+      "scoped_untyped": 0.208,
+      "seed": 20260742,
+      "selected_groups": 96,
+      "selected_layers": {
+        "2": 6,
+        "4": 10,
+        "5": 12,
+        "17": 2,
+        "18": 12,
+        "19": 7,
+        "20": 7,
+        "21": 10,
+        "22": 15,
+        "23": 15
+      },
+      "selection_plan_sha256": "cdb8f17ae11322d276b68b96873398257a8936ebb20a007749d4811bce6039e1"
+    },
+    {
+      "causal_start": 20480,
+      "checkpoint_sha256": "bc5ccd9f57643ebb48bab156a671a348b237e2cb1fe0cb3017d0438e869aadee",
+      "config_sha256": "7bdfa6eae5e5ffbbef0e595acec3b58a9091a665460ed3cf778ecba0ce6d4200",
+      "confirmation_stage_opened": false,
+      "confirmation_windows": {
+        "count": 8,
+        "start": 409
+      },
+      "corrected_marginal_correct": {
+        "confidence": 0.95,
+        "lower": -2.0666471755248494e-05,
+        "mean": 0.00013299021520651877,
+        "samples": 1000,
+        "upper": 0.00028320043929852545
+      },
+      "corrected_matched_h0_passed": false,
+      "diagnostic_windows": {
+        "count": 8,
+        "start": 400
+      },
+      "expanded_documents": 470,
+      "expanded_ppl_ratio": 1.0013051380076936,
+      "expanded_ppl_ratio_lower": 1.0011546308526744,
+      "expanded_ppl_ratio_upper": 1.0014523960525095,
+      "expanded_random": 0.1285,
+      "expanded_typed": 0.9935,
+      "expanded_untyped": 0.191,
+      "expanded_wrong_event": 0.1285,
+      "expanded_wrong_role": 0.0015,
+      "expanded_zero": 0.1285,
+      "frontier_late_ppl_ratio": 1.000198165907381,
+      "frontier_ppl_ratio": 0.9998095370934663,
+      "frontier_throughput_ratio": 0.952829708330366,
+      "frontier_typed": 0.998,
+      "frontier_untyped": 0.186,
+      "initialization": "fresh_seeded_grouped_adapter",
+      "interaction_audit_sha256": "498cd1e0af04a7d1d08d4f750a9e689d8131dc6daa3f45ee68d6386a94abb0f1",
+      "interaction_pairs": 6903,
+      "interaction_triples": 128,
+      "interaction_window": 408,
+      "invalid_programs": 0,
+      "local_graph_groups": 14,
+      "local_groups": 82,
+      "localization_and_typed_path_passed": true,
+      "program": "STRATA-HEADQUOTIENT-v1.1-REP3-FRESH",
+      "registered_result_passed": false,
+      "scoped_correct": 0.9975,
+      "scoped_untyped": 0.1935,
+      "seed": 20260777,
+      "selected_groups": 96,
+      "selected_layers": {
+        "2": 6,
+        "4": 9,
+        "5": 12,
+        "6": 4,
+        "18": 12,
+        "19": 7,
+        "20": 7,
+        "21": 11,
+        "22": 14,
+        "23": 14
+      },
+      "selection_plan_sha256": "dd872d27f63c8b9a277afb309ff22be08f37bd0f04a73ed92431b62b8981b190"
+    },
+    {
+      "causal_start": 24576,
+      "checkpoint_sha256": "17cb536d9e74fcae58e28d58eca9aff55484c0490770205670983b935b12dced",
+      "config_sha256": "5f09189d69fbffe1545bc1cd0180e12fbc17c9727509ef5c9381456549de15f2",
+      "confirmation_stage_opened": false,
+      "confirmation_windows": {
+        "count": 8,
+        "start": 509
+      },
+      "corrected_marginal_correct": {
+        "confidence": 0.95,
+        "lower": -0.00018559966702014208,
+        "mean": -3.4817545383702964e-05,
+        "samples": 1000,
+        "upper": 0.00010850727267097682
+      },
+      "corrected_matched_h0_passed": false,
+      "diagnostic_windows": {
+        "count": 8,
+        "start": 500
+      },
+      "expanded_documents": 470,
+      "expanded_ppl_ratio": 1.0014123009914193,
+      "expanded_ppl_ratio_lower": 1.0012669492990858,
+      "expanded_ppl_ratio_upper": 1.0015544143634656,
+      "expanded_random": 0.1205,
+      "expanded_typed": 0.994,
+      "expanded_untyped": 0.192,
+      "expanded_wrong_event": 0.1205,
+      "expanded_wrong_role": 0.0025,
+      "expanded_zero": 0.1205,
+      "frontier_late_ppl_ratio": 1.0010175258666096,
+      "frontier_ppl_ratio": 1.0003239584965344,
+      "frontier_throughput_ratio": 0.9780706254021283,
+      "frontier_typed": 0.9945,
+      "frontier_untyped": 0.186,
+      "initialization": "fresh_seeded_grouped_adapter",
+      "interaction_audit_sha256": "746093b123e1839c38d78e3ce7bbe157a34b9b206fc095e423885a74eaae9822",
+      "interaction_pairs": 6903,
+      "interaction_triples": 128,
+      "interaction_window": 508,
+      "invalid_programs": 0,
+      "local_graph_groups": 14,
+      "local_groups": 82,
+      "localization_and_typed_path_passed": true,
+      "program": "STRATA-HEADQUOTIENT-v1.1-REP4-FRESH",
+      "registered_result_passed": false,
+      "scoped_correct": 0.995,
+      "scoped_untyped": 0.197,
+      "seed": 20260791,
+      "selected_groups": 96,
+      "selected_layers": {
+        "2": 5,
+        "4": 10,
+        "5": 12,
+        "17": 4,
+        "18": 12,
+        "19": 7,
+        "20": 7,
+        "21": 11,
+        "22": 14,
+        "23": 14
+      },
+      "selection_plan_sha256": "73b103008ff57ee18f2c5ff353c5266c9829b12a592f02c7b779dde3a749bdac"
+    }
+  ],
+  "selected_set_similarity": [
+    {
+      "intersection": 91,
+      "jaccard": 0.900990099009901,
+      "left": "STRATA-HEADQUOTIENT-v1.1",
+      "right": "STRATA-HEADQUOTIENT-v1.1-REP3-FRESH",
+      "union": 101
+    },
+    {
+      "intersection": 93,
+      "jaccard": 0.9393939393939394,
+      "left": "STRATA-HEADQUOTIENT-v1.1",
+      "right": "STRATA-HEADQUOTIENT-v1.1-REP4-FRESH",
+      "union": 99
+    },
+    {
+      "intersection": 89,
+      "jaccard": 0.8640776699029126,
+      "left": "STRATA-HEADQUOTIENT-v1.1-REP3-FRESH",
+      "right": "STRATA-HEADQUOTIENT-v1.1-REP4-FRESH",
+      "union": 103
+    }
+  ]
+}

raw/reproducibility/model_release.json

@@ -0,0 +1,29 @@
+{
+  "repository": "https://huggingface.co/nur-dev/strata-headquotient-q25",
+  "immutable_revision": "v1.2.0",
+  "hub_commit": "60b2ea8dc02c1b847faf3770105fecb2e9a74d7d",
+  "dense_checkpoint_sha256": "10e8559713ef1d951c604605f8f3666a027a25a341363d0c17006f628cc38c1f",
+  "q25_checkpoint_sha256": "51120c7ecca5234a4e7cee424c199582e444eb04e434a2d7f08263e3ccc74a90",
+  "tokenizer_sha256": "ac3a38d3522a0f1499f506bb8232704005efa4bbcbb6455f59366fa6181f8301",
+  "physical_modes": {
+    "GLOBAL": 288,
+    "LOCAL": 81,
+    "LOCAL_GRAPH": 15
+  },
+  "included": [
+    "dense reference checkpoint",
+    "physical Q25 checkpoint",
+    "event-scoped adapter",
+    "static selection plan",
+    "470-document tokenized confirmation bundle",
+    "packed semantic-confirmation source",
+    "tokenizer",
+    "pinned environment",
+    "minimal transitive source snapshot",
+    "structural verifier",
+    "one-command evaluation",
+    "two fresh Q25 checkpoints and complete audits",
+    "corrected matched-capacity H0 reanalysis",
+    "cached prefill/decode profile and implementation"
+  ]
+}

raw/reproducibility/q25_cached_decode_profile.json

@@ -0,0 +1,422 @@
+{
+  "created_at": "2026-07-21T02:07:16.839792+00:00",
+  "device": "cuda:6",
+  "graph_enabled": false,
+  "measurement": "inference-only KV cache; BF16 numerical equivalence is checked against full-sequence logits; prefill and autoregressive decode are reported separately",
+  "program": "STRATA-HEADQUOTIENT-Q25-CACHED-DECODE-PROFILE",
+  "rows": [
+    {
+      "batch_size": 1,
+      "dense": {
+        "batch_size": 1,
+        "decode_seconds": 0.23088740557432175,
+        "decode_steps": 32,
+        "decode_tokens_per_second": 138.59569308426106,
+        "first_cached_token_ms": 7.441435630122821,
+        "peak_allocated_bytes": 2767192064,
+        "persistent_kv_bytes": 204472320,
+        "prefill_seconds": 0.02532536784807841,
+        "prefill_tokens_per_second": 80867.53220271169,
+        "prefix_length": 2048
+      },
+      "prefix_length": 2048,
+      "q25": {
+        "batch_size": 1,
+        "decode_seconds": 0.21997203739980856,
+        "decode_steps": 32,
+        "decode_tokens_per_second": 145.47303547422547,
+        "first_cached_token_ms": 7.406104976932208,
+        "peak_allocated_bytes": 2728734208,
+        "persistent_kv_bytes": 178520064,
+        "prefill_seconds": 0.02850375821193059,
+        "prefill_tokens_per_second": 71850.1744497252,
+        "prefix_length": 2048
+      },
+      "ratios": {
+        "decode_throughput": 1.0496216169270371,
+        "persistent_kv": 0.8730769230769231,
+        "prefill_throughput": 0.8884922352968238
+      }
+    },
+    {
+      "batch_size": 4,
+      "dense": {
+        "batch_size": 4,
+        "decode_seconds": 0.23466505793233713,
+        "decode_steps": 32,
+        "decode_tokens_per_second": 545.458284790347,
+        "first_cached_token_ms": 7.920100043217341,
+        "peak_allocated_bytes": 4261779456,
+        "persistent_kv_bytes": 817889280,
+        "prefill_seconds": 0.09082256692151229,
+        "prefill_tokens_per_second": 90197.84705137681,
+        "prefix_length": 2048
+      },
+      "prefix_length": 2048,
+      "q25": {
+        "batch_size": 4,
+        "decode_seconds": 0.2278061956167221,
+        "decode_steps": 32,
+        "decode_tokens_per_second": 561.881118524786,
+        "first_cached_token_ms": 7.969736432035764,
+        "peak_allocated_bytes": 4070479872,
+        "persistent_kv_bytes": 714080256,
+        "prefill_seconds": 0.0923770306011041,
+        "prefill_tokens_per_second": 88680.05332812775,
+        "prefix_length": 2048
+      },
+      "ratios": {
+        "decode_throughput": 1.0301083221070724,
+        "persistent_kv": 0.8730769230769231,
+        "prefill_throughput": 0.9831726169430127
+      }
+    },
+    {
+      "batch_size": 8,
+      "dense": {
+        "batch_size": 8,
+        "decode_seconds": 0.25229039539893466,
+        "decode_steps": 32,
+        "decode_tokens_per_second": 1014.7037091729136,
+        "first_cached_token_ms": 8.601184313495954,
+        "peak_allocated_bytes": 6273153024,
+        "persistent_kv_bytes": 1635778560,
+        "prefill_seconds": 0.19029431355496249,
+        "prefill_tokens_per_second": 86098.2112072825,
+        "prefix_length": 2048
+      },
+      "prefix_length": 2048,
+      "q25": {
+        "batch_size": 8,
+        "decode_seconds": 0.23515919471780458,
+        "decode_steps": 32,
+        "decode_tokens_per_second": 1088.6242415790068,
+        "first_cached_token_ms": 7.94680913289388,
+        "peak_allocated_bytes": 5870467072,
+        "persistent_kv_bytes": 1428160512,
+        "prefill_seconds": 0.19251583764950433,
+        "prefill_tokens_per_second": 85104.6864509341,
+        "prefix_length": 2048
+      },
+      "ratios": {
+        "decode_throughput": 1.072849376362629,
+        "persistent_kv": 0.8730769230769231,
+        "prefill_throughput": 0.9884605644830823
+      }
+    },
+    {
+      "batch_size": 16,
+      "dense": {
+        "batch_size": 16,
+        "decode_seconds": 0.3312743529677391,
+        "decode_steps": 32,
+        "decode_tokens_per_second": 1545.5467512447628,
+        "first_cached_token_ms": 11.202741414308548,
+        "peak_allocated_bytes": 10296591872,
+        "persistent_kv_bytes": 3271557120,
+        "prefill_seconds": 0.40776471979916096,
+        "prefill_tokens_per_second": 80360.06650143603,
+        "prefix_length": 2048
+      },
+      "prefix_length": 2048,
+      "q25": {
+        "batch_size": 16,
+        "decode_seconds": 0.29237379630406696,
+        "decode_steps": 32,
+        "decode_tokens_per_second": 1751.1829256665776,
+        "first_cached_token_ms": 9.928669159611067,
+        "peak_allocated_bytes": 9478374912,
+        "persistent_kv_bytes": 2856321024,
+        "prefill_seconds": 0.41604801764090854,
+        "prefill_tokens_per_second": 78760.13972089658,
+        "prefix_length": 2048
+      },
+      "ratios": {
+        "decode_throughput": 1.1330507629459918,
+        "persistent_kv": 0.8730769230769231,
+        "prefill_throughput": 0.9800905244334155
+      }
+    },
+    {
+      "batch_size": 1,
+      "dense": {
+        "batch_size": 1,
+        "decode_seconds": 0.23166167673965296,
+        "decode_steps": 32,
+        "decode_tokens_per_second": 138.13247167317354,
+        "first_cached_token_ms": 7.798002411921819,
+        "peak_allocated_bytes": 3247719936,
+        "persistent_kv_bytes": 405798912,
+        "prefill_seconds": 0.04814621185262998,
+        "prefill_tokens_per_second": 85074.19052068697,
+        "prefix_length": 4096
+      },
+      "prefix_length": 4096,
+      "q25": {
+        "batch_size": 1,
+        "decode_seconds": 0.22054913143316904,
+        "decode_steps": 32,
+        "decode_tokens_per_second": 145.0923873155069,
+        "first_cached_token_ms": 7.15858240922292,
+        "peak_allocated_bytes": 3113641472,
+        "persistent_kv_bytes": 329515008,
+        "prefill_seconds": 0.05360218447943529,
+        "prefill_tokens_per_second": 76414.79614644154,
+        "prefix_length": 4096
+      },
+      "ratios": {
+        "decode_throughput": 1.0503858039896714,
+        "persistent_kv": 0.812015503875969,
+        "prefill_throughput": 0.8982136142436785
+      }
+    },
+    {
+      "batch_size": 4,
+      "dense": {
+        "batch_size": 4,
+        "decode_seconds": 0.25190934414664906,
+        "decode_steps": 32,
+        "decode_tokens_per_second": 508.1193015431964,
+        "first_cached_token_ms": 8.424771949648857,
+        "peak_allocated_bytes": 6247872512,
+        "persistent_kv_bytes": 1623195648,
+        "prefill_seconds": 0.20185994543135166,
+        "prefill_tokens_per_second": 81165.18591635038,
+        "prefix_length": 4096
+      },
+      "prefix_length": 4096,
+      "q25": {
+        "batch_size": 4,
+        "decode_seconds": 0.231443390250206,
+        "decode_steps": 32,
+        "decode_tokens_per_second": 553.0510068212504,
+        "first_cached_token_ms": 7.94626772403717,
+        "peak_allocated_bytes": 5648562176,
+        "persistent_kv_bytes": 1318060032,
+        "prefill_seconds": 0.20270894095301628,
+        "prefill_tokens_per_second": 80825.2459066296,
+        "prefix_length": 4096
+      },
+      "ratios": {
+        "decode_throughput": 1.0884274719373839,
+        "persistent_kv": 0.812015503875969,
+        "prefill_throughput": 0.9958117509880267
+      }
+    },
+    {
+      "batch_size": 8,
+      "dense": {
+        "batch_size": 8,
+        "decode_seconds": 0.321361793205142,
+        "decode_steps": 32,
+        "decode_tokens_per_second": 796.6099437234028,
+        "first_cached_token_ms": 10.821449880798658,
+        "peak_allocated_bytes": 10243749888,
+        "persistent_kv_bytes": 3246391296,
+        "prefill_seconds": 0.4328818339854479,
+        "prefill_tokens_per_second": 75697.33222184962,
+        "prefix_length": 4096
+      },
+      "prefix_length": 4096,
+      "q25": {
+        "batch_size": 8,
+        "decode_seconds": 0.27276255128284294,
+        "decode_steps": 32,
+        "decode_tokens_per_second": 938.5452614224124,
+        "first_cached_token_ms": 9.249638145168623,
+        "peak_allocated_bytes": 9035757568,
+        "persistent_kv_bytes": 2636120064,
+        "prefill_seconds": 0.43729556041459244,
+        "prefill_tokens_per_second": 74933.30133270327,
+        "prefix_length": 4096
+      },
+      "ratios": {
+        "decode_throughput": 1.1781741727144346,
+        "persistent_kv": 0.812015503875969,
+        "prefill_throughput": 0.9899067659754881
+      }
+    },
+    {
+      "batch_size": 16,
+      "dense": {
+        "batch_size": 16,
+        "decode_seconds": 0.4728281206140916,
+        "decode_steps": 32,
+        "decode_tokens_per_second": 1082.8459173177632,
+        "first_cached_token_ms": 15.76905200878779,
+        "peak_allocated_bytes": 18239866368,
+        "persistent_kv_bytes": 6492782592,
+        "prefill_seconds": 0.8857398132483164,
+        "prefill_tokens_per_second": 73990.12556481644,
+        "prefix_length": 4096
+      },
+      "prefix_length": 4096,
+      "q25": {
+        "batch_size": 16,
+        "decode_seconds": 0.3972581879546245,
+        "decode_steps": 32,
+        "decode_tokens_per_second": 1288.8343538899733,
+        "first_cached_token_ms": 13.64798347155253,
+        "peak_allocated_bytes": 15803827712,
+        "persistent_kv_bytes": 5272240128,
+        "prefill_seconds": 0.8932362391302983,
+        "prefill_tokens_per_second": 73369.1683443221,
+        "prefix_length": 4096
+      },
+      "ratios": {
+        "decode_throughput": 1.1902287604153772,
+        "persistent_kv": 0.812015503875969,
+        "prefill_throughput": 0.9916075663373436
+      }
+    },
+    {
+      "batch_size": 1,
+      "dense": {
+        "batch_size": 1,
+        "decode_seconds": 0.22510927046338716,
+        "decode_steps": 32,
+        "decode_tokens_per_second": 142.1531860244051,
+        "first_cached_token_ms": 7.497106368343036,
+        "peak_allocated_bytes": 4234716672,
+        "persistent_kv_bytes": 805306368,
+        "prefill_seconds": 0.10965904593467712,
+        "prefill_tokens_per_second": 74412.46575189826,
+        "prefix_length": 8160
+      },
+      "prefix_length": 8160,
+      "q25": {
+        "batch_size": 1,
+        "decode_seconds": 0.224536320194602,
+        "decode_steps": 32,
+        "decode_tokens_per_second": 142.5159189046392,
+        "first_cached_token_ms": 7.5872062395016355,
+        "peak_allocated_bytes": 3902450176,
+        "persistent_kv_bytes": 629145600,
+        "prefill_seconds": 0.11838383475939433,
+        "prefill_tokens_per_second": 68928.32975536353,
+        "prefix_length": 8160
+      },
+      "ratios": {
+        "decode_throughput": 1.0025517041888306,
+        "persistent_kv": 0.78125,
+        "prefill_throughput": 0.9263008429955861
+      }
+    },
+    {
+      "batch_size": 4,
+      "dense": {
+        "batch_size": 4,
+        "decode_seconds": 0.31844382857282955,
+        "decode_steps": 32,
+        "decode_tokens_per_second": 401.95472015789375,
+        "first_cached_token_ms": 10.655300070842108,
+        "peak_allocated_bytes": 10188224512,
+        "persistent_kv_bytes": 3221225472,
+        "prefill_seconds": 0.47874748582641286,
+        "prefill_tokens_per_second": 68177.9037307254,
+        "prefix_length": 8160
+      },
+      "prefix_length": 8160,
+      "q25": {
+        "batch_size": 4,
+        "decode_seconds": 0.27023141148189705,
+        "decode_steps": 32,
+        "decode_tokens_per_second": 473.6681028236971,
+        "first_cached_token_ms": 8.966594313581785,
+        "peak_allocated_bytes": 8791914496,
+        "persistent_kv_bytes": 2516582400,
+        "prefill_seconds": 0.4703146008153756,
+        "prefill_tokens_per_second": 69400.35445085618,
+        "prefix_length": 8160
+      },
+      "ratios": {
+        "decode_throughput": 1.1784115948125533,
+        "persistent_kv": 0.78125,
+        "prefill_throughput": 1.0179303066424417
+      }
+    },
+    {
+      "batch_size": 8,
+      "dense": {
+        "batch_size": 8,
+        "decode_seconds": 0.45806841490169364,
+        "decode_steps": 32,
+        "decode_tokens_per_second": 558.8684826805627,
+        "first_cached_token_ms": 15.288629879554113,
+        "peak_allocated_bytes": 18128516096,
+        "persistent_kv_bytes": 6442450944,
+        "prefill_seconds": 0.9819728036721548,
+        "prefill_tokens_per_second": 66478.41952025653,
+        "prefix_length": 8160
+      },
+      "prefix_length": 8160,
+      "q25": {
+        "batch_size": 8,
+        "decode_seconds": 0.3758538272231817,
+        "decode_steps": 32,
+        "decode_tokens_per_second": 681.1158526476502,
+        "first_cached_token_ms": 12.771941100557646,
+        "peak_allocated_bytes": 15313089536,
+        "persistent_kv_bytes": 5033164800,
+        "prefill_seconds": 0.9673846984903017,
+        "prefill_tokens_per_second": 67480.9102334116,
+        "prefix_length": 8160
+      },
+      "ratios": {
+        "decode_throughput": 1.2187408554168933,
+        "persistent_kv": 0.78125,
+        "prefill_throughput": 1.0150799420381769
+      }
+    },
+    {
+      "batch_size": 16,
+      "dense": {
+        "batch_size": 16,
+        "decode_seconds": 0.7506915175666412,
+        "decode_steps": 32,
+        "decode_tokens_per_second": 682.037811829342,
+        "first_cached_token_ms": 25.01120666662852,
+        "peak_allocated_bytes": 34005301248,
+        "persistent_kv_bytes": 12884901888,
+        "prefill_seconds": 2.1065259805570045,
+        "prefill_tokens_per_second": 61978.82257567861,
+        "prefix_length": 8160
+      },
+      "prefix_length": 8160,
+      "q25": {
+        "batch_size": 16,
+        "decode_seconds": 0.606967706233263,
+        "decode_steps": 32,
+        "decode_tokens_per_second": 843.5374645833858,
+        "first_cached_token_ms": 20.8244522412618,
+        "peak_allocated_bytes": 28370909184,
+        "persistent_kv_bytes": 10066329600,
+        "prefill_seconds": 2.119566404571136,
+        "prefill_tokens_per_second": 61597.50396044655,
+        "prefix_length": 8160
+      },
+      "ratios": {
+        "decode_throughput": 1.2367898816648804,
+        "persistent_kv": 0.78125,
+        "prefill_throughput": 0.9938475982700953
+      }
+    }
+  ],
+  "torch_version": "2.5.1+cu124",
+  "verification": {
+    "dense": {
+      "maximum_absolute_logit_difference": 0.0625,
+      "maximum_absolute_probability_difference": 0.0015385262668132782,
+      "mean_absolute_logit_difference": 0.009482888504862785,
+      "numerically_equivalent": true,
+      "top_token_exact": true
+    },
+    "q25": {
+      "maximum_absolute_logit_difference": 0.078125,
+      "maximum_absolute_probability_difference": 0.002120431512594223,
+      "mean_absolute_logit_difference": 0.011962523683905602,
+      "numerically_equivalent": true,
+      "top_token_exact": true
+    }
+  }
+}

raw/reproducibility/q25_matched_h0_reanalysis.json

@@ -0,0 +1,416 @@
+{
+  "checkpoint": {
+    "path": "source-data://strata/outputs/strata-headquotient-v1-1/frontier/Q25/exported_model.pt",
+    "sha256": "42f16ff73107bb218c9cb3d5cd8038f56c03e5f876261af790c5e0362b886841"
+  },
+  "controlled_causal": {
+    "effects_vs_graph_zero": {
+      "correct": {
+        "confidence": 0.95,
+        "lower": 0.2476348578929901,
+        "mean": 0.25487813353538513,
+        "samples": 1000,
+        "upper": 0.26197588443756104
+      },
+      "random": {
+        "confidence": 0.95,
+        "lower": 0.0,
+        "mean": 0.0,
+        "samples": 1000,
+        "upper": 0.0
+      },
+      "untyped": {
+        "confidence": 0.95,
+        "lower": -0.0043089319951832294,
+        "mean": -0.0007616211660206318,
+        "samples": 1000,
+        "upper": 0.002813332946971059
+      },
+      "wrong_event": {
+        "confidence": 0.95,
+        "lower": 0.0,
+        "mean": 0.0,
+        "samples": 1000,
+        "upper": 0.0
+      },
+      "wrong_role": {
+        "confidence": 0.95,
+        "lower": -0.26323336362838745,
+        "mean": -0.2560182511806488,
+        "samples": 1000,
+        "upper": -0.24900954961776733
+      }
+    },
+    "examples": 2000,
+    "marginal_handoff_vs_hq0": {
+      "correct": {
+        "confidence": 0.95,
+        "lower": -0.00023968867026269436,
+        "mean": -8.118145342450589e-05,
+        "samples": 1000,
+        "upper": 7.799099694238976e-05
+      },
+      "random": {
+        "confidence": 0.95,
+        "lower": 0.0,
+        "mean": 0.0,
+        "samples": 1000,
+        "upper": 0.0
+      },
+      "untyped": {
+        "confidence": 0.95,
+        "lower": -9.874618262983859e-05,
+        "mean": 7.627225568285212e-05,
+        "samples": 1000,
+        "upper": 0.00023413376766256988
+      },
+      "wrong_event": {
+        "confidence": 0.95,
+        "lower": 0.0,
+        "mean": 0.0,
+        "samples": 1000,
+        "upper": 0.0
+      },
+      "wrong_role": {
+        "confidence": 0.95,
+        "lower": -0.00010881897469516844,
+        "mean": 6.049078729120083e-05,
+        "samples": 1000,
+        "upper": 0.0002462287957314402
+      }
+    },
+    "matched_h0_graph_capacity": {
+      "attention_modes": "all global",
+      "graph_groups": 15,
+      "program_coverage": {
+        "READ_EVENT_ARG0": 15,
+        "READ_EVENT_ARG1": 7,
+        "READ_EVENT_ARG2": 3,
+        "READ_EVENT_LOCATION": 2,
+        "READ_EVENT_TIME": 3
+      }
+    },
+    "matched_h0_metrics": {
+      "correct": {
+        "accuracy": 0.9995,
+        "mean_margin": 0.254790335893631,
+        "wrong_target_accuracy": 0.0
+      },
+      "random": {
+        "accuracy": 0.125,
+        "mean_margin": -0.000168985512573272,
+        "wrong_target_accuracy": 0.135
+      },
+      "untyped": {
+        "accuracy": 0.1835,
+        "mean_margin": -0.001006878330372274,
+        "wrong_target_accuracy": 0.1885
+      },
+      "wrong_event": {
+        "accuracy": 0.125,
+        "mean_margin": -0.000168985512573272,
+        "wrong_target_accuracy": 0.135
+      },
+      "wrong_role": {
+        "accuracy": 0.0,
+        "mean_margin": -0.2562476694583893,
+        "wrong_target_accuracy": 1.0
+      },
+      "zero": {
+        "accuracy": 0.125,
+        "mean_margin": -0.000168985512573272,
+        "wrong_target_accuracy": 0.135
+      }
+    },
+    "metrics": {
+      "correct": {
+        "accuracy": 1.0,
+        "mean_margin": 0.25481492280960083,
+        "wrong_target_accuracy": 0.0
+      },
+      "random": {
+        "accuracy": 0.1315,
+        "mean_margin": -6.322524131974205e-05,
+        "wrong_target_accuracy": 0.134
+      },
+      "untyped": {
+        "accuracy": 0.184,
+        "mean_margin": -0.0008248474914580584,
+        "wrong_target_accuracy": 0.189
+      },
+      "wrong_event": {
+        "accuracy": 0.1315,
+        "mean_margin": -6.322524131974205e-05,
+        "wrong_target_accuracy": 0.134
+      },
+      "wrong_role": {
+        "accuracy": 0.0,
+        "mean_margin": -0.2560814619064331,
+        "wrong_target_accuracy": 1.0
+      },
+      "zero": {
+        "accuracy": 0.1315,
+        "mean_margin": -6.322524131974205e-05,
+        "wrong_target_accuracy": 0.134
+      }
+    },
+    "null_controls_exact": true
+  },
+  "created_at": "2026-07-21T02:25:27.096612+00:00",
+  "gates": {
+    "causal": false,
+    "ppl": true,
+    "runtime": true
+  },
+  "invalid_programs": 0,
+  "natural_carrier_causal": {
+    "carrier": {
+      "accuracy": 0.9711791212917762,
+      "authoritative_coverage": 0.7972814623857511,
+      "authoritative_execution": 0.9826689774696707,
+      "authoritative_fact_count": 3462,
+      "edge_macro_f1": 0.9209658892559601,
+      "error_categories": {
+        "correct": 1821,
+        "multi-event ambiguity": 47,
+        "valency-signature miss": 7,
+        "wrong carrier relation": 93,
+        "wrong event identity": 32
+      },
+      "examples": 2000,
+      "gold_fact_count": 4267,
+      "invalid_programs": 0,
+      "per_label": {
+        "CCOMP": {
+          "f1": 0.8947368421052632,
+          "precision": 0.952,
+          "recall": 0.8439716312056738,
+          "support": 141
+        },
+        "CLAUSAL_SUBJECT": {
+          "f1": 0.823529411764706,
+          "precision": 0.7777777777777778,
+          "recall": 0.875,
+          "support": 8
+        },
+        "COORDINATED_WITH": {
+          "f1": 0.9640522875816994,
+          "precision": 0.9800664451827242,
+          "recall": 0.9485530546623794,
+          "support": 311
+        },
+        "CORE_OBLIQUE": {
+          "f1": 0.942608695652174,
+          "precision": 0.9492119089316988,
+          "recall": 0.9360967184801382,
+          "support": 579
+        },
+        "DIRECT_OBJECT": {
+          "f1": 0.9611267605633804,
+          "precision": 0.9595050618672666,
+          "recall": 0.9627539503386005,
+          "support": 886
+        },
+        "INDIRECT_OBJECT": {
+          "f1": 0.9483568075117371,
+          "precision": 0.926605504587156,
+          "recall": 0.9711538461538461,
+          "support": 104
+        },
+        "PASSIVE_SUBJECT": {
+          "f1": 0.8480565371024734,
+          "precision": 0.8571428571428571,
+          "recall": 0.8391608391608392,
+          "support": 143
+        },
+        "RECIPIENT_OBLIQUE": {
+          "f1": 0.9651741293532337,
+          "precision": 0.9797979797979798,
+          "recall": 0.9509803921568627,
+          "support": 204
+        },
+        "RELATIVE_CLAUSE": {
+          "f1": 0.8801652892561984,
+          "precision": 0.9301310043668122,
+          "recall": 0.8352941176470589,
+          "support": 255
+        },
+        "SURFACE_SUBJECT": {
+          "f1": 0.9748641304347826,
+          "precision": 0.9755268524813052,
+          "recall": 0.9742023082145281,
+          "support": 1473
+        },
+        "XCOMP": {
+          "f1": 0.9279538904899135,
+          "precision": 0.875,
+          "recall": 0.9877300613496932,
+          "support": 163
+        }
+      },
+      "raw_execution": 0.9105,
+      "split": "external_v3",
+      "wrong_authoritative_rate": 0.01733102253032929
+    },
+    "examples": 2000,
+    "note": "frozen natural carrier/R12 evaluation; not reused for frontier selection",
+    "r12": {
+      "examples": 256,
+      "graph_off_fallback_exact": true,
+      "graph_zero": {
+        "accuracy": 0.046875,
+        "candidate_nll": 3.4653353095054626,
+        "candidate_ppl": 31.987183581840533,
+        "examples": 256
+      },
+      "ordinary_ppl_ratio_to_dense": 1.0019550335910028,
+      "physical_removal": true,
+      "random_carrier": {
+        "accuracy": 0.0,
+        "candidate_nll": 3.4696131348609924,
+        "candidate_ppl": 32.12431226397685,
+        "examples": 256
+      },
+      "typed": {
+        "accuracy": 1.0,
+        "candidate_nll": 3.303746908903122,
+        "candidate_ppl": 27.21441806833827,
+        "examples": 256
+      },
+      "typed_minus_untyped": 0.50390625,
+      "untyped": {
+        "accuracy": 0.49609375,
+        "candidate_nll": 3.3569958806037903,
+        "candidate_ppl": 28.702834489918512,
+        "examples": 256
+      },
+      "wrong_event": {
+        "accuracy": 0.046875,
+        "candidate_nll": 3.4653353095054626,
+        "candidate_ppl": 31.987183581840533,
+        "examples": 256
+      },
+      "wrong_role": {
+        "accuracy": 0.0078125,
+        "candidate_nll": 3.4682549834251404,
+        "candidate_ppl": 32.08071219759812,
+        "examples": 256
+      }
+    },
+    "sha256": "9ef0f47765ee3dd32aee8b34b92b77e1846f685ccd5a2a95722b4c8f04a0c499",
+    "source": "source-data://strata/outputs/strata-carrier-resolve-v1/external_v3/evaluation.json"
+  },
+  "next_stage": "CLOSED",
+  "passed": false,
+  "phase": "diagnostic",
+  "physical_export": {
+    "affected_layers": [
+      2,
+      4,
+      5,
+      17,
+      18,
+      19,
+      20,
+      21,
+      22,
+      23
+    ],
+    "dense_qkv_in_exported_layers": 0,
+    "global_kv_reduction": 1.3333333333333333,
+    "graph_adapter_compaction": {
+      "active_groups_per_program": {
+        "READ_EVENT_ARG0": 15,
+        "READ_EVENT_ARG1": 7,
+        "READ_EVENT_ARG2": 3,
+        "READ_EVENT_LOCATION": 2,
+        "READ_EVENT_TIME": 3
+      },
+      "buffer_bytes": 80,
+      "form": "parameter_free_program_routing",
+      "parameter_bytes": 2
+    },
+    "graph_groups": 15,
+    "graph_program_coverage": {
+      "READ_EVENT_ARG0": 15,
+      "READ_EVENT_ARG1": 7,
+      "READ_EVENT_ARG2": 3,
+      "READ_EVENT_LOCATION": 2,
+      "READ_EVENT_TIME": 3
+    },
+    "groups_removed": 96,
+    "replacement_fraction": 0.25
+  },
+  "plan": {
+    "path": "source-data://strata/outputs/strata-headquotient-v1-1/selection/Q25/plan.json",
+    "sha256": "cdb8f17ae11322d276b68b96873398257a8936ebb20a007749d4811bce6039e1"
+  },
+  "program": "STRATA-HEADQUOTIENT-v1.1",
+  "quality_pass": false,
+  "runtime": {
+    "candidate": {
+      "peak_allocated_bytes": 1488253440,
+      "seconds": 0.11170258559286594,
+      "tokens_per_second": 73337.60410755608
+    },
+    "dense": {
+      "peak_allocated_bytes": 1488563200,
+      "seconds": 0.10870877715448539,
+      "tokens_per_second": 75357.30061941915
+    },
+    "peak_allocation_ratio": 0.9997919067191773,
+    "throughput_ratio": 0.973198396236839
+  },
+  "smoke": false,
+  "stabilization": {
+    "steps": 32,
+    "trajectory": [
+      {
+        "loss": 5.2098612785339355,
+        "step": 1
+      },
+      {
+        "loss": 4.9329328536987305,
+        "step": 32
+      }
+    ]
+  },
+  "stage": "Q25_diagnostic",
+  "true_8k": {
+    "candidate_loss": {
+      "0-2048": 5.627997607458383,
+      "2048-4096": 5.985978780314326,
+      "4096-8192": 5.343196413834307,
+      "aggregate": 5.575120614920034
+    },
+    "dense_loss": {
+      "0-2048": 5.62827554019168,
+      "2048-4096": 5.985017543658614,
+      "4096-8192": 5.345074839236826,
+      "aggregate": 5.575888866121378
+    },
+    "language_ppl_ratio": {
+      "ar": 1.001996138817111,
+      "de": 1.001525510033417,
+      "en": 1.0002027979636123,
+      "es": 0.9981573776510331,
+      "zh": 0.9976745123862902
+    },
+    "ppl_ratio": {
+      "0-2048": 0.9997221058864268,
+      "2048-4096": 1.0009616987917285,
+      "4096-8192": 0.9981233377343309,
+      "aggregate": 0.9992320438280531
+    },
+    "windows": [
+      96,
+      97,
+      98,
+      99,
+      100,
+      101,
+      102,
+      103
+    ]
+  }
+}

reproduced/summary.json

@@ -1,5 +1,61 @@
 {
   "article": "STRATA HEADQUOTIENT Q25",
+  "cached_runtime": {
+    "prefix_8160": [
+      [
+        1,
+        "142.153186",
+        "142.515919",
+        "1.002552",
+        "0.926301",
+        "0.781250",
+        "0.921537"
+      ],
+      [
+        4,
+        "401.954720",
+        "473.668103",
+        "1.178412",
+        "1.017930",
+        "0.781250",
+        "0.862949"
+      ],
+      [
+        8,
+        "558.868483",
+        "681.115853",
+        "1.218741",
+        "1.015080",
+        "0.781250",
+        "0.844696"
+      ],
+      [
+        16,
+        "682.037812",
+        "843.537465",
+        "1.236790",
+        "0.993848",
+        "0.781250",
+        "0.834308"
+      ]
+    ],
+    "verification": {
+      "dense": {
+        "maximum_absolute_logit_difference": 0.0625,
+        "maximum_absolute_probability_difference": 0.0015385262668132782,
+        "mean_absolute_logit_difference": 0.009482888504862785,
+        "numerically_equivalent": true,
+        "top_token_exact": true
+      },
+      "q25": {
+        "maximum_absolute_logit_difference": 0.078125,
+        "maximum_absolute_probability_difference": 0.002120431512594223,
+        "mean_absolute_logit_difference": 0.011962523683905602,
+        "numerically_equivalent": true,
+        "top_token_exact": true
+      }
+    }
+  },
   "evidence": {
     "accuracy": 1.0,
     "decision_coverage": 1.0,
@@ -69,22 +125,115 @@
   "q25": {
     "groups_replaced": 96,
     "language_ppl_ratio": {
-      "ar": 1.0014931180241606,
-      "de": 1.0014569436456722,
-      "en": 1.0005086972746728,
-      "es": 0.9979950617397495,
-      "zh": 0.9974994495185546
+      "ar": {
+        "confidence": 0.95,
+        "documents": 20,
+        "nll_difference": 0.0007950443849503941,
+        "nll_difference_lower": -0.0002508042924294322,
+        "nll_difference_upper": 0.0017058238347448552,
+        "ppl_ratio": 1.0007953605165114,
+        "ppl_ratio_lower": 0.9997492271563378,
+        "ppl_ratio_upper": 1.001707279579853,
+        "samples": 10000
+      },
+      "de": {
+        "confidence": 0.95,
+        "documents": 50,
+        "nll_difference": 0.0016225555833478823,
+        "nll_difference_lower": 0.0012127477743440838,
+        "nll_difference_upper": 0.0020814374370199235,
+        "ppl_ratio": 1.001623872638894,
+        "ppl_ratio_lower": 1.0012134834502926,
+        "ppl_ratio_upper": 1.0020836051316346,
+        "samples": 10000
+      },
+      "en": {
+        "confidence": 0.95,
+        "documents": 50,
+        "nll_difference": 0.0012889576202753262,
+        "nll_difference_lower": 0.000771741110040498,
+        "nll_difference_upper": 0.0018380203796616401,
+        "ppl_ratio": 1.0012897886831786,
+        "ppl_ratio_lower": 1.000772038978832,
+        "ppl_ratio_upper": 1.0018397105744985,
+        "samples": 10000
+      },
+      "es": {
+        "confidence": 0.95,
+        "documents": 50,
+        "nll_difference": 0.0012452177826177114,
+        "nll_difference_lower": 0.0009779897500504543,
+        "nll_difference_upper": 0.0015135438062378857,
+        "ppl_ratio": 1.0012459933881799,
+        "ppl_ratio_lower": 1.000978468137966,
+        "ppl_ratio_upper": 1.0015146897917582,
+        "samples": 10000
+      },
+      "zh": {
+        "confidence": 0.95,
+        "documents": 300,
+        "nll_difference": 0.001545217917117152,
+        "nll_difference_lower": 0.00137042381407093,
+        "nll_difference_upper": 0.0017341674375761192,
+        "ppl_ratio": 1.0015464123814795,
+        "ppl_ratio_lower": 1.0013713632738896,
+        "ppl_ratio_upper": 1.0017356719755082,
+        "samples": 10000
+      }
     },
+    "local_graph_groups": 15,
+    "local_only_groups": 81,
     "peak_allocation_ratio": 0.9997919067191773,
+    "position_ppl_ratio": {
+      "0-2048": {
+        "confidence": 0.95,
+        "documents": 470,
+        "nll_difference": 0.0004259108783716851,
+        "nll_difference_lower": 0.00029588773093642076,
+        "nll_difference_upper": 0.0005560303618140677,
+        "ppl_ratio": 1.000426001591288,
+        "ppl_ratio_lower": 1.0002959315100288,
+        "ppl_ratio_upper": 1.000556184975351,
+        "samples": 10000
+      },
+      "2048-4096": {
+        "confidence": 0.95,
+        "documents": 470,
+        "nll_difference": 0.0014160845587228208,
+        "nll_difference_lower": 0.0012086108188204309,
+        "nll_difference_upper": 0.0016230750466360055,
+        "ppl_ratio": 1.0014170876799071,
+        "ppl_ratio_lower": 1.0012093414832095,
+        "ppl_ratio_upper": 1.0016243929458595,
+        "samples": 10000
+      },
+      "4096-8192": {
+        "confidence": 0.95,
+        "documents": 470,
+        "nll_difference": 0.0020038272965785444,
+        "nll_difference_lower": 0.001770213721668351,
+        "nll_difference_upper": 0.0022347243062779704,
+        "ppl_ratio": 1.0020058363001705,
+        "ppl_ratio_lower": 1.001771781474928,
+        "ppl_ratio_upper": 1.0022372231637129,
+        "samples": 10000
+      }
+    },
     "ppl_ratio": {
-      "0-2048": 1.00007699819542,
-      "2048-4096": 1.0009002031766665,
-      "4096-8192": 0.9993893766212806,
-      "aggregate": 0.9999388623918676
+      "confidence": 0.95,
+      "documents": 470,
+      "nll_difference": 0.001462346408821701,
+      "nll_difference_lower": 0.0013137826055078657,
+      "nll_difference_upper": 0.0016099244922990635,
+      "ppl_ratio": 1.0014634161587161,
+      "ppl_ratio_lower": 1.0013146459959361,
+      "ppl_ratio_upper": 1.0016112211164634,
+      "samples": 10000
     },
     "throughput_ratio": 0.9843666476918135,
-    "typed_execution": 0.9995,
-    "untyped_execution": 0.179
+    "token_kv_reduction": 0.21875,
+    "typed_execution": 0.999,
+    "untyped_execution": 0.187
   },
   "q30": {
     "passed": false,
@@ -95,5 +244,234 @@
       "aggregate": 1.231681966831007
     }
   },
-  "version": "1.0.1"
+  "replication": {
+    "all_corrected_matched_h0_passed": false,
+    "all_expanded_typed_at_least_0_95": true,
+    "all_expanded_upper_below_1_03": true,
+    "all_invalid_programs_zero": true,
+    "all_localization_and_typed_path_passed": true,
+    "fresh_replications": 2,
+    "program": "STRATA-HEADQUOTIENT-Q25-INDEPENDENT-REPLICATION",
+    "runs": [
+      {
+        "causal_start": 2048,
+        "checkpoint_sha256": "51120c7ecca5234a4e7cee424c199582e444eb04e434a2d7f08263e3ccc74a90",
+        "config_sha256": "c4a3742f240a67380913554dbfc6efd4862a41fd1ce38bc08fd1e6952730a9c9",
+        "confirmation_stage_opened": true,
+        "confirmation_windows": {
+          "count": 8,
+          "start": 112
+        },
+        "corrected_marginal_correct": {
+          "confidence": 0.95,
+          "lower": -0.00023968867026269436,
+          "mean": -8.118145342450589e-05,
+          "samples": 1000,
+          "upper": 7.799099694238976e-05
+        },
+        "corrected_matched_h0_passed": false,
+        "diagnostic_windows": {
+          "count": 8,
+          "start": 96
+        },
+        "expanded_documents": 470,
+        "expanded_ppl_ratio": 1.0014634161587161,
+        "expanded_ppl_ratio_lower": 1.0013146459959361,
+        "expanded_ppl_ratio_upper": 1.0016112211164634,
+        "expanded_random": 0.1335,
+        "expanded_typed": 0.999,
+        "expanded_untyped": 0.187,
+        "expanded_wrong_event": 0.1335,
+        "expanded_wrong_role": 0.0,
+        "expanded_zero": 0.1335,
+        "frontier_late_ppl_ratio": 0.9993893766212806,
+        "frontier_ppl_ratio": 0.9999388623918676,
+        "frontier_throughput_ratio": 0.9843666476918135,
+        "frontier_typed": 0.9995,
+        "frontier_untyped": 0.179,
+        "initialization": "registered_legacy_adapter",
+        "interaction_audit_sha256": "7a844ea72203f0511b99ab52b71d8b2c62dc305cbbb599f144fba84ebd7ecc5c",
+        "interaction_pairs": 6903,
+        "interaction_triples": 128,
+        "interaction_window": 110,
+        "invalid_programs": 0,
+        "local_graph_groups": 15,
+        "local_groups": 81,
+        "localization_and_typed_path_passed": true,
+        "program": "STRATA-HEADQUOTIENT-v1.1",
+        "registered_result_passed": true,
+        "scoped_correct": 0.97,
+        "scoped_untyped": 0.208,
+        "seed": 20260742,
+        "selected_groups": 96,
+        "selected_layers": {
+          "17": 2,
+          "18": 12,
+          "19": 7,
+          "2": 6,
+          "20": 7,
+          "21": 10,
+          "22": 15,
+          "23": 15,
+          "4": 10,
+          "5": 12
+        },
+        "selection_plan_sha256": "cdb8f17ae11322d276b68b96873398257a8936ebb20a007749d4811bce6039e1"
+      },
+      {
+        "causal_start": 20480,
+        "checkpoint_sha256": "bc5ccd9f57643ebb48bab156a671a348b237e2cb1fe0cb3017d0438e869aadee",
+        "config_sha256": "7bdfa6eae5e5ffbbef0e595acec3b58a9091a665460ed3cf778ecba0ce6d4200",
+        "confirmation_stage_opened": false,
+        "confirmation_windows": {
+          "count": 8,
+          "start": 409
+        },
+        "corrected_marginal_correct": {
+          "confidence": 0.95,
+          "lower": -2.0666471755248494e-05,
+          "mean": 0.00013299021520651877,
+          "samples": 1000,
+          "upper": 0.00028320043929852545
+        },
+        "corrected_matched_h0_passed": false,
+        "diagnostic_windows": {
+          "count": 8,
+          "start": 400
+        },
+        "expanded_documents": 470,
+        "expanded_ppl_ratio": 1.0013051380076936,
+        "expanded_ppl_ratio_lower": 1.0011546308526744,
+        "expanded_ppl_ratio_upper": 1.0014523960525095,
+        "expanded_random": 0.1285,
+        "expanded_typed": 0.9935,
+        "expanded_untyped": 0.191,
+        "expanded_wrong_event": 0.1285,
+        "expanded_wrong_role": 0.0015,
+        "expanded_zero": 0.1285,
+        "frontier_late_ppl_ratio": 1.000198165907381,
+        "frontier_ppl_ratio": 0.9998095370934663,
+        "frontier_throughput_ratio": 0.952829708330366,
+        "frontier_typed": 0.998,
+        "frontier_untyped": 0.186,
+        "initialization": "fresh_seeded_grouped_adapter",
+        "interaction_audit_sha256": "498cd1e0af04a7d1d08d4f750a9e689d8131dc6daa3f45ee68d6386a94abb0f1",
+        "interaction_pairs": 6903,
+        "interaction_triples": 128,
+        "interaction_window": 408,
+        "invalid_programs": 0,
+        "local_graph_groups": 14,
+        "local_groups": 82,
+        "localization_and_typed_path_passed": true,
+        "program": "STRATA-HEADQUOTIENT-v1.1-REP3-FRESH",
+        "registered_result_passed": false,
+        "scoped_correct": 0.9975,
+        "scoped_untyped": 0.1935,
+        "seed": 20260777,
+        "selected_groups": 96,
+        "selected_layers": {
+          "18": 12,
+          "19": 7,
+          "2": 6,
+          "20": 7,
+          "21": 11,
+          "22": 14,
+          "23": 14,
+          "4": 9,
+          "5": 12,
+          "6": 4
+        },
+        "selection_plan_sha256": "dd872d27f63c8b9a277afb309ff22be08f37bd0f04a73ed92431b62b8981b190"
+      },
+      {
+        "causal_start": 24576,
+        "checkpoint_sha256": "17cb536d9e74fcae58e28d58eca9aff55484c0490770205670983b935b12dced",
+        "config_sha256": "5f09189d69fbffe1545bc1cd0180e12fbc17c9727509ef5c9381456549de15f2",
+        "confirmation_stage_opened": false,
+        "confirmation_windows": {
+          "count": 8,
+          "start": 509
+        },
+        "corrected_marginal_correct": {
+          "confidence": 0.95,
+          "lower": -0.00018559966702014208,
+          "mean": -3.4817545383702964e-05,
+          "samples": 1000,
+          "upper": 0.00010850727267097682
+        },
+        "corrected_matched_h0_passed": false,
+        "diagnostic_windows": {
+          "count": 8,
+          "start": 500
+        },
+        "expanded_documents": 470,
+        "expanded_ppl_ratio": 1.0014123009914193,
+        "expanded_ppl_ratio_lower": 1.0012669492990858,
+        "expanded_ppl_ratio_upper": 1.0015544143634656,
+        "expanded_random": 0.1205,
+        "expanded_typed": 0.994,
+        "expanded_untyped": 0.192,
+        "expanded_wrong_event": 0.1205,
+        "expanded_wrong_role": 0.0025,
+        "expanded_zero": 0.1205,
+        "frontier_late_ppl_ratio": 1.0010175258666096,
+        "frontier_ppl_ratio": 1.0003239584965344,
+        "frontier_throughput_ratio": 0.9780706254021283,
+        "frontier_typed": 0.9945,
+        "frontier_untyped": 0.186,
+        "initialization": "fresh_seeded_grouped_adapter",
+        "interaction_audit_sha256": "746093b123e1839c38d78e3ce7bbe157a34b9b206fc095e423885a74eaae9822",
+        "interaction_pairs": 6903,
+        "interaction_triples": 128,
+        "interaction_window": 508,
+        "invalid_programs": 0,
+        "local_graph_groups": 14,
+        "local_groups": 82,
+        "localization_and_typed_path_passed": true,
+        "program": "STRATA-HEADQUOTIENT-v1.1-REP4-FRESH",
+        "registered_result_passed": false,
+        "scoped_correct": 0.995,
+        "scoped_untyped": 0.197,
+        "seed": 20260791,
+        "selected_groups": 96,
+        "selected_layers": {
+          "17": 4,
+          "18": 12,
+          "19": 7,
+          "2": 5,
+          "20": 7,
+          "21": 11,
+          "22": 14,
+          "23": 14,
+          "4": 10,
+          "5": 12
+        },
+        "selection_plan_sha256": "73b103008ff57ee18f2c5ff353c5266c9829b12a592f02c7b779dde3a749bdac"
+      }
+    ],
+    "selected_set_similarity": [
+      {
+        "intersection": 91,
+        "jaccard": 0.900990099009901,
+        "left": "STRATA-HEADQUOTIENT-v1.1",
+        "right": "STRATA-HEADQUOTIENT-v1.1-REP3-FRESH",
+        "union": 101
+      },
+      {
+        "intersection": 93,
+        "jaccard": 0.9393939393939394,
+        "left": "STRATA-HEADQUOTIENT-v1.1",
+        "right": "STRATA-HEADQUOTIENT-v1.1-REP4-FRESH",
+        "union": 99
+      },
+      {
+        "intersection": 89,
+        "jaccard": 0.8640776699029126,
+        "left": "STRATA-HEADQUOTIENT-v1.1-REP3-FRESH",
+        "right": "STRATA-HEADQUOTIENT-v1.1-REP4-FRESH",
+        "union": 103
+      }
+    ]
+  },
+  "version": "1.2.0"
 }

reproduced/tables.md

@@ -6,15 +6,15 @@ Generated by `python scripts/reproduce.py` from the frozen files in `raw/` and `
 
 | Mode | Historical state | Admission rule | Stores global KV | Typed program |
 | --- | --- | --- | --- | --- |
-| GLOBAL | Full causal token KV history | Retain when late-context lexical or discourse utility is material, graph substitution is inadequate, or cluster-retention constraints require the group. | true | false |
-| LOCAL | 1024-token local-window KV | The group passes the registered candidate rule, is selected jointly under the interaction and cluster constraints, and has no event-scoped graph program. | false | false |
-| LOCAL_GRAPH | 1024-token local-window KV plus an event-scoped typed result | The group passes the registered candidate rule, is selected jointly, has a valid typed program assignment, and shows relation-specific causal handoff without wrong-event or random benefit. | false | true |
+| GLOBAL | Full causal token KV history | Retain when late-context lexical or discourse utility is material, graph substitution is inadequate, or cluster-retention constraints require the head. | true | false |
+| LOCAL | 1024-token local-window KV | The head passes the pre-specified candidate rule, is selected jointly under the interaction and cluster constraints, and has no event-scoped graph program. | false | false |
+| LOCAL_GRAPH | 1024-token local-window KV plus an event-scoped typed result | The head passes the pre-specified candidate rule, is selected jointly, has a valid typed program assignment, and shows relation-specific causal handoff without wrong-event or random benefit. | false | true |
 
 ## Analysed Use Cases
 
 | ID | Scope | Use case | Component | Method | KV mode/program |
 | --- | --- | --- | --- | --- | --- |
-| UC01 | Model core | Late-context lexical and discourse continuation | Retained global KV groups | Dense grouped-query causal attention | GLOBAL / NO_GRAPH_READ |
+| UC01 | Model core | Late-context lexical and discourse continuation | Retained global KV heads | Dense multi-head causal attention | GLOBAL / NO_GRAPH_READ |
 | UC02 | Model core | Recent token and phrase continuation | Localized non-graph KV groups | Local causal attention | LOCAL / NO_GRAPH_READ |
 | UC03 | Model core | Agent or ARG0 retrieval | Event-scoped graph group | Exact typed graph execution | LOCAL_GRAPH / READ_EVENT_ARG0 |
 | UC04 | Model core | Patient or ARG1 retrieval | Event-scoped graph group | Exact typed graph execution | LOCAL_GRAPH / READ_EVENT_ARG1 |
@@ -29,7 +29,7 @@ Generated by `python scripts/reproduce.py` from the frozen files in `raw/` and `
 | Quantity | Count |
 | --- | --- |
 | Audited groups | 384 |
-| Registered candidates | 118 |
+| Pre-specified candidates | 118 |
 | GLOBAL | 288 |
 | LOCAL | 81 |
 | LOCAL_GRAPH | 15 |
@@ -43,30 +43,38 @@ Generated by `python scripts/reproduce.py` from the frozen files in `raw/` and `
 | Q25 | 96 | 25.000000 | 15 | 0.999939 | 0.999389 | 0.999500 | 0.984367 | pass |
 | Q30 | 115 | 29.947917 | 15 | 1.231682 | 1.378070 | 1.000000 | 0.939274 | fail |
 
-## Q25 Position and Language Ratios
+## Expanded Q25 Position and Language Ratios
 
-| Partition | PPL ratio |
-| --- | --- |
-| 0-2048 | 1.000077 |
-| 2048-4096 | 1.000900 |
-| 4096-8192 | 0.999389 |
-| aggregate | 0.999939 |
-| ar | 1.001493 |
-| de | 1.001457 |
-| en | 1.000509 |
-| es | 0.997995 |
-| zh | 0.997499 |
-
-## Q25 Causal Controls
-
-| Condition | Accuracy | Mean margin | Marginal handoff vs. H0 (95% CI) |
+| Partition | Documents | PPL ratio | 95% CI |
+| --- | --- | --- | --- |
+| aggregate | 470 | 1.001463 | [1.001315, 1.001611] |
+| 0-2048 | 470 | 1.000426 | [1.000296, 1.000556] |
+| 2048-4096 | 470 | 1.001417 | [1.001209, 1.001624] |
+| 4096-8192 | 470 | 1.002006 | [1.001772, 1.002237] |
+| ar | 20 | 1.000795 | [0.999749, 1.001707] |
+| de | 50 | 1.001624 | [1.001213, 1.002084] |
+| en | 50 | 1.001290 | [1.000772, 1.001840] |
+| es | 50 | 1.001246 | [1.000978, 1.001515] |
+| zh | 300 | 1.001546 | [1.001371, 1.001736] |
+
+## Q25 Typed-Graph Controls and Corrected Attribution
+
+| Condition | Untouched accuracy | Development margin | Corrected matched-capacity effect (95% CI) |
 | --- | --- | --- | --- |
-| correct | 0.999500 | 0.255090 | 0.096378 [0.088277, 0.104493] |
-| untyped | 0.179000 | -0.000668 | 0.002323 [-0.003675, 0.008805] |
-| wrong_role | 0.000000 | -0.255897 | -0.102981 [-0.111355, -0.094596] |
-| wrong_event | 0.114500 | 0.000028 | 0.002597 [-0.000531, 0.005697] |
-| random | 0.114500 | 0.000028 | 0.000086 [-0.002257, 0.002709] |
-| zero | 0.114500 | 0.000028 | reference |
+| correct | 0.999000 | 0.255090 | -0.000081 [-0.000240, 0.000078] |
+| untyped | 0.187000 | -0.000668 | not re-estimated in matched-capacity audit |
+| wrong_role | 0.000000 | -0.255897 | not re-estimated in matched-capacity audit |
+| wrong_event | 0.133500 | 0.000028 | not re-estimated in matched-capacity audit |
+| random | 0.133500 | 0.000028 | not re-estimated in matched-capacity audit |
+| zero | 0.133500 | 0.000028 | not re-estimated in matched-capacity audit |
+
+## Q25 Fresh Adaptation, Interaction-Audit, and Selection Runs
+
+| Run | Localized | Graph heads | PPL ratio | PPL upper 95% | Typed | Untyped | Matched-capacity marginal effect | Strict attribution |
+| --- | --- | --- | --- | --- | --- | --- | --- | --- |
+| original | 96 | 15 | 1.001463 | 1.001611 | 0.999000 | 0.187000 | -0.000081 [-0.000240, 0.000078] | not passed |
+| fresh_seed_20260777 | 96 | 14 | 1.001305 | 1.001452 | 0.993500 | 0.191000 | 0.000133 [-0.000021, 0.000283] | not passed |
+| fresh_seed_20260791 | 96 | 14 | 1.001412 | 1.001554 | 0.994000 | 0.192000 | -0.000035 [-0.000186, 0.000109] | not passed |
 
 ## Q25 Physical and Runtime Summary
 
@@ -76,10 +84,21 @@ Generated by `python scripts/reproduce.py` from the frozen files in `raw/` and `
 | Graph groups | 15 |
 | Global KV retained | 0.750000 |
 | Dense/Q25 global-KV ratio | 1.333333 |
+| Token-KV bytes before graph state | 629145600 |
+| Token-KV reduction | 0.218750 |
 | Core throughput ratio | 0.984367 |
 | Peak allocation ratio | 0.999792 |
 | Dense QKV modules in localized layers | 0 |
 
+## Cached Model-Core Profile at an 8160-Token Prefix
+
+| Batch | Dense decode tok/s | Q25 decode tok/s | Decode ratio | Prefill ratio | Persistent KV ratio | Peak ratio |
+| --- | --- | --- | --- | --- | --- | --- |
+| 1 | 142.153186 | 142.515919 | 1.002552 | 0.926301 | 0.781250 | 0.921537 |
+| 4 | 401.954720 | 473.668103 | 1.178412 | 1.017930 | 0.781250 | 0.862949 |
+| 8 | 558.868483 | 681.115853 | 1.218741 | 1.015080 | 0.781250 | 0.844696 |
+| 16 | 682.037812 | 843.537465 | 1.236790 | 0.993848 | 0.781250 | 0.834308 |
+
 ## External Natural Carrier and Evidence Decisions
 
 | Metric | Value |

scripts/reproduce.py

@@ -42,6 +42,9 @@ def derive() -> tuple[dict, str, str]:
     use_cases = load_csv("data/use_case_classifications.csv")
     frontier = {name: load_json(f"raw/frontier/{name}.json") for name in ("q10", "q20", "q25", "q30")}
     q25 = frontier["q25"]
+    postreview = load_json("raw/postreview/q25_confirmation.json")
+    cached = load_json("raw/reproducibility/q25_cached_decode_profile.json")
+    replications = load_json("raw/replication/q25_replications.json")
     natural = load_json("raw/external/natural_carrier.json")
     evidence = load_json("raw/external/evidence_decisions.json")
     whole = load_json("raw/negative/whole_layer_true8k.json")
@@ -61,11 +64,11 @@ def derive() -> tuple[dict, str, str]:
     modes = Counter(row["q25_mode"] for row in groups)
     selected_modes = Counter(row["q25_mode"] for row in selected)
     if len(candidates) != 118:
-        raise AssertionError(f"expected 118 registered candidates, found {len(candidates)}")
+        raise AssertionError(f"expected 118 pre-specified candidates, found {len(candidates)}")
     if len(selected) != 96 or selected_modes != Counter({"LOCAL": 81, "LOCAL_GRAPH": 15}):
         raise AssertionError(f"unexpected Q25 assignment: {len(selected)} groups, {selected_modes}")
     if any(row["candidate_by_registered_rule"] != "true" for row in selected):
-        raise AssertionError("Q25 includes a group outside the registered candidate set")
+        raise AssertionError("Q25 includes a group outside the pre-specified candidate set")
 
     active_coverage: Counter[str] = Counter()
     for row in selected:
@@ -124,29 +127,82 @@ def derive() -> tuple[dict, str, str]:
         metric("frontier", label, "typed_execution", typed, "accuracy", ">=0.95", source)
         metric("frontier", label, "core_throughput_ratio", runtime, "ratio", ">=0.95", source)
 
-    ppl = q25["true_8k"]["ppl_ratio"]
-    for bucket, value in ppl.items():
-        metric("q25_ppl", bucket, "ppl_ratio", value, "ratio", "<=1.03 aggregate; <=1.05 late", "raw/frontier/q25.json")
-    for language, value in sorted(q25["true_8k"]["language_ppl_ratio"].items()):
-        metric("q25_language", language, "ppl_ratio", value, "ratio", "<=1.05", "raw/frontier/q25.json")
+    ppl = postreview["language_model"]
+    aggregate_ppl = ppl["paired_document_bootstrap"]
+    metric("q25_ppl_postreview", "aggregate", "ppl_ratio", aggregate_ppl["ppl_ratio"], "ratio", "upper 95% CI <1.03", "raw/postreview/q25_confirmation.json")
+    metric("q25_ppl_postreview", "aggregate", "ppl_ratio_lower_95", aggregate_ppl["ppl_ratio_lower"], "ratio", "", "raw/postreview/q25_confirmation.json")
+    metric("q25_ppl_postreview", "aggregate", "ppl_ratio_upper_95", aggregate_ppl["ppl_ratio_upper"], "ratio", "<1.03", "raw/postreview/q25_confirmation.json")
+    for bucket, values in ppl["position_buckets"].items():
+        metric("q25_ppl_postreview", bucket, "ppl_ratio", values["ppl_ratio"], "ratio", "<=1.05 late", "raw/postreview/q25_confirmation.json")
+    for language, values in sorted(ppl["languages"].items()):
+        metric("q25_language_postreview", language, "ppl_ratio", values["ppl_ratio"], "ratio", "<=1.05", "raw/postreview/q25_confirmation.json")
 
     causal = q25["controlled_causal"]
+    semantic_confirmation = postreview["semantic_confirmation"]["metrics"]
+    corrected_original = replications["runs"][0]["corrected_marginal_correct"]
     causal_rows = []
     for name in ("correct", "untyped", "wrong_role", "wrong_event", "random", "zero"):
         values = causal["metrics"][name]
-        marginal = causal["marginal_handoff_vs_hq0"].get(name)
-        interval = "reference" if marginal is None else (
-            f"{f6(marginal['mean'])} [{f6(marginal['lower'])}, {f6(marginal['upper'])}]"
+        interval = (
+            f"{f6(corrected_original['mean'])} "
+            f"[{f6(corrected_original['lower'])}, {f6(corrected_original['upper'])}]"
+            if name == "correct" else "not re-estimated in matched-capacity audit"
         )
         causal_rows.append([name, f6(values["accuracy"]), f6(values["mean_margin"]), interval])
         metric("causal", name, "accuracy", values["accuracy"], "accuracy", "", "raw/frontier/q25.json")
         metric("causal", name, "mean_margin_vs_zero", causal["effects_vs_graph_zero"].get(name, {}).get("mean", 0.0), "margin", "", "raw/frontier/q25.json")
+        metric("causal_confirmation", name, "accuracy", semantic_confirmation[name]["accuracy"], "accuracy", "", "raw/postreview/q25_confirmation.json")
+    metric("causal_attribution", "correct", "matched_capacity_marginal_mean", corrected_original["mean"], "margin", "95% CI must exclude zero", "raw/replication/q25_replications.json")
+    metric("causal_attribution", "correct", "matched_capacity_marginal_lower_95", corrected_original["lower"], "margin", "", "raw/replication/q25_replications.json")
+    metric("causal_attribution", "correct", "matched_capacity_marginal_upper_95", corrected_original["upper"], "margin", "", "raw/replication/q25_replications.json")
+
+    replication_rows = []
+    for index, run in enumerate(replications["runs"]):
+        label = "original" if index == 0 else f"fresh_seed_{run['seed']}"
+        interval = run["corrected_marginal_correct"]
+        replication_rows.append([
+            label,
+            run["selected_groups"],
+            run["local_graph_groups"],
+            f6(run["expanded_ppl_ratio"]),
+            f6(run["expanded_ppl_ratio_upper"]),
+            f6(run["expanded_typed"]),
+            f6(run["expanded_untyped"]),
+            f"{f6(interval['mean'])} [{f6(interval['lower'])}, {f6(interval['upper'])}]",
+            "pass" if run["corrected_matched_h0_passed"] else "not passed",
+        ])
+        source = "raw/replication/q25_replications.json"
+        metric("replication", label, "expanded_ppl_ratio", run["expanded_ppl_ratio"], "ratio", "upper 95% CI <1.03", source)
+        metric("replication", label, "expanded_typed", run["expanded_typed"], "accuracy", ">=0.95", source)
+        metric("replication", label, "expanded_untyped", run["expanded_untyped"], "accuracy", "", source)
+        metric("replication", label, "matched_capacity_marginal_mean", interval["mean"], "margin", "95% CI must exclude zero", source)
 
     runtime = q25["runtime"]
     physical = q25["physical_export"]
     metric("runtime", "Q25", "throughput_ratio", runtime["throughput_ratio"], "ratio", ">=0.95", "raw/frontier/q25.json")
     metric("runtime", "Q25", "peak_allocation_ratio", runtime["peak_allocation_ratio"], "ratio", "<=1.00", "raw/frontier/q25.json")
     metric("runtime", "Q25", "global_kv_retained", 1.0 - physical["replacement_fraction"], "fraction", "<=0.80", "raw/frontier/q25.json")
+    persistent = postreview["persistent_state"]
+    metric("runtime", "Q25", "token_kv_bytes", persistent["q25_token_kv_bytes_before_graph_state"], "bytes", "", "raw/postreview/q25_confirmation.json")
+    metric("runtime", "Q25", "token_kv_reduction", persistent["token_kv_reduction"], "fraction", "", "raw/postreview/q25_confirmation.json")
+    cached_rows = []
+    for row in cached["rows"]:
+        if int(row["prefix_length"]) != 8160:
+            continue
+        batch = int(row["batch_size"])
+        cached_rows.append([
+            batch,
+            f6(row["dense"]["decode_tokens_per_second"]),
+            f6(row["q25"]["decode_tokens_per_second"]),
+            f6(row["ratios"]["decode_throughput"]),
+            f6(row["ratios"]["prefill_throughput"]),
+            f6(row["ratios"]["persistent_kv"]),
+            f6(row["q25"]["peak_allocated_bytes"] / row["dense"]["peak_allocated_bytes"]),
+        ])
+        source = "raw/reproducibility/q25_cached_decode_profile.json"
+        metric("cached_runtime", f"batch_{batch}", "decode_throughput_ratio", row["ratios"]["decode_throughput"], "ratio", "", source)
+        metric("cached_runtime", f"batch_{batch}", "prefill_throughput_ratio", row["ratios"]["prefill_throughput"], "ratio", "", source)
+        metric("cached_runtime", f"batch_{batch}", "persistent_kv_ratio", row["ratios"]["persistent_kv"], "ratio", "", source)
 
     carrier = natural["carrier"]
     natural_metrics = [
@@ -165,9 +221,9 @@ def derive() -> tuple[dict, str, str]:
     for name, value in natural_metrics:
         metric("external_natural", "external_v3", name, value, "count" if name == "invalid_programs" else "rate", "", "raw/external/natural_carrier.json")
     ev = evidence["metrics"]
-    metric("evidence", "registered_800", "macro_decision_accuracy", ev["macro_decision_accuracy"], "accuracy", ">=0.95", "raw/external/evidence_decisions.json")
-    metric("evidence", "registered_800", "proof_correctness", ev["proof_correctness"], "accuracy", "=1.00", "raw/external/evidence_decisions.json")
-    metric("evidence", "registered_800", "paired_identical_agreement", ev["paired_identical_agreement"], "agreement", "=1.00", "raw/external/evidence_decisions.json")
+    metric("evidence", "hash_locked_800", "macro_decision_accuracy", ev["macro_decision_accuracy"], "accuracy", ">=0.95", "raw/external/evidence_decisions.json")
+    metric("evidence", "hash_locked_800", "proof_correctness", ev["proof_correctness"], "accuracy", "=1.00", "raw/external/evidence_decisions.json")
+    metric("evidence", "hash_locked_800", "paired_identical_agreement", ev["paired_identical_agreement"], "agreement", "=1.00", "raw/external/evidence_decisions.json")
 
     negative_rows = [
         ["Whole-layer graph-only R12", f"aggregate {f6(whole['ppl_ratio']['aggregate'])}; late {f6(whole['ppl_ratio']['4096-8192'])}", "Failed true-8k preservation"],
@@ -212,20 +268,29 @@ def derive() -> tuple[dict, str, str]:
         "",
         markdown_table(
             ["Quantity", "Count"],
-            [["Audited groups", len(groups)], ["Registered candidates", len(candidates)], ["GLOBAL", modes["GLOBAL"]], ["LOCAL", modes["LOCAL"]], ["LOCAL_GRAPH", modes["LOCAL_GRAPH"]]],
+            [["Audited groups", len(groups)], ["Pre-specified candidates", len(candidates)], ["GLOBAL", modes["GLOBAL"]], ["LOCAL", modes["LOCAL"]], ["LOCAL_GRAPH", modes["LOCAL_GRAPH"]]],
         ),
         "",
         "## Replacement Frontier",
         "",
         markdown_table(["Frontier", "Groups", "Replaced (%)", "Graph groups", "Aggregate PPL", "4-8k PPL", "Typed", "Throughput", "Status"], frontier_rows),
         "",
-        "## Q25 Position and Language Ratios",
+        "## Expanded Q25 Position and Language Ratios",
+        "",
+        markdown_table(
+            ["Partition", "Documents", "PPL ratio", "95% CI"],
+            [["aggregate", aggregate_ppl["documents"], f6(aggregate_ppl["ppl_ratio"]), f"[{f6(aggregate_ppl['ppl_ratio_lower'])}, {f6(aggregate_ppl['ppl_ratio_upper'])}]"]]
+            + [[bucket, values["documents"], f6(values["ppl_ratio"]), f"[{f6(values['ppl_ratio_lower'])}, {f6(values['ppl_ratio_upper'])}]"] for bucket, values in ppl["position_buckets"].items()]
+            + [[language, values["documents"], f6(values["ppl_ratio"]), f"[{f6(values['ppl_ratio_lower'])}, {f6(values['ppl_ratio_upper'])}]"] for language, values in sorted(ppl["languages"].items())],
+        ),
+        "",
+        "## Q25 Typed-Graph Controls and Corrected Attribution",
         "",
-        markdown_table(["Partition", "PPL ratio"], [[name, f6(value)] for name, value in ppl.items()] + [[language, f6(value)] for language, value in sorted(q25["true_8k"]["language_ppl_ratio"].items())]),
+        markdown_table(["Condition", "Untouched accuracy", "Development margin", "Corrected matched-capacity effect (95% CI)"], [[row[0], f6(semantic_confirmation[row[0]]["accuracy"]), row[2], row[3]] for row in causal_rows]),
         "",
-        "## Q25 Causal Controls",
+        "## Q25 Fresh Adaptation, Interaction-Audit, and Selection Runs",
         "",
-        markdown_table(["Condition", "Accuracy", "Mean margin", "Marginal handoff vs. H0 (95% CI)"], causal_rows),
+        markdown_table(["Run", "Localized", "Graph heads", "PPL ratio", "PPL upper 95%", "Typed", "Untyped", "Matched-capacity marginal effect", "Strict attribution"], replication_rows),
         "",
         "## Q25 Physical and Runtime Summary",
         "",
@@ -234,11 +299,20 @@ def derive() -> tuple[dict, str, str]:
             ["Graph groups", physical["graph_groups"]],
             ["Global KV retained", f6(1.0 - physical["replacement_fraction"])],
             ["Dense/Q25 global-KV ratio", f6(physical["global_kv_reduction"])],
+            ["Token-KV bytes before graph state", persistent["q25_token_kv_bytes_before_graph_state"]],
+            ["Token-KV reduction", f6(persistent["token_kv_reduction"])],
             ["Core throughput ratio", f6(runtime["throughput_ratio"])],
             ["Peak allocation ratio", f6(runtime["peak_allocation_ratio"])],
             ["Dense QKV modules in localized layers", physical["dense_qkv_in_exported_layers"]],
         ]),
         "",
+        "## Cached Model-Core Profile at an 8160-Token Prefix",
+        "",
+        markdown_table(
+            ["Batch", "Dense decode tok/s", "Q25 decode tok/s", "Decode ratio", "Prefill ratio", "Persistent KV ratio", "Peak ratio"],
+            cached_rows,
+        ),
+        "",
         "## External Natural Carrier and Evidence Decisions",
         "",
         markdown_table(["Metric", "Value"], [[name, f6(value) if isinstance(value, float) else value] for name, value in natural_metrics] + [["evidence macro accuracy", f6(ev["macro_decision_accuracy"])], ["evidence proof correctness", f6(ev["proof_correctness"])] ]),
@@ -256,7 +330,7 @@ def derive() -> tuple[dict, str, str]:
 
     summary = {
         "article": "STRATA HEADQUOTIENT Q25",
-        "version": "1.0.1",
+        "version": "1.2.0",
         "group_audit": {
             "audited": len(groups),
             "candidates": len(candidates),
@@ -265,15 +339,24 @@ def derive() -> tuple[dict, str, str]:
             "active_program_coverage": dict(sorted(active_coverage.items())),
         },
         "q25": {
-            "ppl_ratio": ppl,
-            "language_ppl_ratio": q25["true_8k"]["language_ppl_ratio"],
-            "typed_execution": causal["metrics"]["correct"]["accuracy"],
-            "untyped_execution": causal["metrics"]["untyped"]["accuracy"],
+            "ppl_ratio": aggregate_ppl,
+            "position_ppl_ratio": ppl["position_buckets"],
+            "language_ppl_ratio": ppl["languages"],
+            "typed_execution": semantic_confirmation["correct"]["accuracy"],
+            "untyped_execution": semantic_confirmation["untyped"]["accuracy"],
             "throughput_ratio": runtime["throughput_ratio"],
             "peak_allocation_ratio": runtime["peak_allocation_ratio"],
             "groups_replaced": physical["groups_removed"],
+            "local_only_groups": postreview["architecture"]["local_only_groups"],
+            "local_graph_groups": postreview["architecture"]["local_graph_groups"],
+            "token_kv_reduction": persistent["token_kv_reduction"],
         },
         "q30": {"passed": frontier["q30"]["passed"], "ppl_ratio": frontier["q30"]["true_8k"]["ppl_ratio"]},
+        "cached_runtime": {
+            "verification": cached["verification"],
+            "prefix_8160": cached_rows,
+        },
+        "replication": replications,
         "external_natural": {name: value for name, value in natural_metrics},
         "evidence": evidence["metrics"],
     }

scripts/verify.py

@@ -1,5 +1,5 @@
 #!/usr/bin/env python3
-"""Verify integrity and registered classification rules for the public release."""
+"""Verify integrity and pre-specified classification rules for the public release."""
 
 from __future__ import annotations
 
@@ -63,7 +63,7 @@ def check_groups() -> tuple[int, int, Counter[str]]:
         if row["q25_selected"] == "true":
             selected.append(row)
             if not expected_candidate:
-                raise AssertionError("selected group is not a registered candidate")
+                raise AssertionError("selected group is not a pre-specified candidate")
 
     modes = Counter(row["q25_mode"] for row in rows)
     if len(candidates) != 118 or len(selected) != 96:
@@ -113,6 +113,85 @@ def check_interactions() -> int:
     return len(pairs)
 
 
+def check_postreview() -> None:
+    payload = json.loads(
+        (ROOT / "raw/postreview/q25_confirmation.json").read_text(encoding="utf-8")
+    )
+    lm = payload["language_model"]
+    if lm["documents"] != 470 or lm["tokens"] != 3_850_240:
+        raise AssertionError("post-review LM support differs")
+    if lm["graph_state"] != "disabled/zero for both selection and evaluation":
+        raise AssertionError("post-review graph/PPL boundary differs")
+    if not lm["noninferiority_pass"]:
+        raise AssertionError("post-review PPL non-inferiority failed")
+    if lm["paired_document_bootstrap"]["ppl_ratio_upper"] >= 1.03:
+        raise AssertionError("post-review PPL interval exceeds the margin")
+    semantic = payload["semantic_confirmation"]
+    if not semantic["untouched_by_training_selection_or_thresholding"]:
+        raise AssertionError("semantic confirmation is not untouched")
+    metrics = semantic["metrics"]
+    expected = {
+        "correct": 0.999,
+        "untyped": 0.187,
+        "wrong_role": 0.0,
+        "wrong_event": 0.1335,
+        "random": 0.1335,
+        "zero": 0.1335,
+    }
+    for name, value in expected.items():
+        if float(metrics[name]["accuracy"]) != value:
+            raise AssertionError(f"semantic confirmation differs for {name}")
+    state = payload["persistent_state"]
+    if state["token_kv_reduction"] != 0.21875:
+        raise AssertionError("token-KV accounting differs")
+    architecture = payload["architecture"]
+    if architecture["query_heads_per_kv_head"] != 1:
+        raise AssertionError("the released backbone is not the verified 1:1 MHA geometry")
+    cached = json.loads(
+        (ROOT / "raw/reproducibility/q25_cached_decode_profile.json").read_text(
+            encoding="utf-8"
+        )
+    )
+    if not all(
+        row["numerically_equivalent"]
+        for row in cached["verification"].values()
+    ):
+        raise AssertionError("cached decode did not match the full-sequence path")
+    prefix_8k = {
+        int(row["batch_size"]): row
+        for row in cached["rows"]
+        if int(row["prefix_length"]) == 8160
+    }
+    if set(prefix_8k) != {1, 4, 8, 16}:
+        raise AssertionError("cached 8k batch matrix is incomplete")
+    if any(row["ratios"]["persistent_kv"] != 0.78125 for row in prefix_8k.values()):
+        raise AssertionError("cached persistent-KV ratio differs")
+
+
+def check_replications() -> None:
+    payload = json.loads(
+        (ROOT / "raw/replication/q25_replications.json").read_text(encoding="utf-8")
+    )
+    if payload["fresh_replications"] != 2:
+        raise AssertionError("expected two fresh Q25 campaigns")
+    if not payload["all_localization_and_typed_path_passed"]:
+        raise AssertionError("localization/typed-path replication failed")
+    if payload["all_corrected_matched_h0_passed"]:
+        raise AssertionError("corrected matched-capacity attribution must remain failed")
+    if len(payload["runs"]) != 3:
+        raise AssertionError("replication summary must contain three campaigns")
+    for run in payload["runs"]:
+        if run["selected_groups"] != 96:
+            raise AssertionError("a replication did not select 96 heads")
+        if run["interaction_pairs"] != 6903 or run["interaction_triples"] != 128:
+            raise AssertionError("a fresh interaction audit is incomplete")
+        if run["expanded_ppl_ratio_upper"] >= 1.03 or run["expanded_typed"] < 0.95:
+            raise AssertionError("a localization/typed-path replication metric failed")
+        interval = run["corrected_marginal_correct"]
+        if not interval["lower"] <= 0 <= interval["upper"]:
+            raise AssertionError("corrected matched-capacity interval unexpectedly excludes zero")
+
+
 def check_public_paths() -> None:
     offenders = []
     for path in ROOT.rglob("*"):
@@ -151,19 +230,29 @@ def check_metadata() -> None:
     ):
         if required not in citation:
             raise AssertionError(f"citation metadata lacks {required}")
+    model = json.loads(
+        (ROOT / "raw/reproducibility/model_release.json").read_text(encoding="utf-8")
+    )
+    if model["hub_commit"] != "60b2ea8dc02c1b847faf3770105fecb2e9a74d7d":
+        raise AssertionError("linked model release commit differs")
+    if model["physical_modes"] != {"GLOBAL": 288, "LOCAL": 81, "LOCAL_GRAPH": 15}:
+        raise AssertionError("linked model release mode counts differ")
 
 
 def main() -> None:
     subprocess.run([sys.executable, str(ROOT / "scripts/reproduce.py"), "--check"], check=True)
     candidates, selected, modes = check_groups()
     pairs = check_interactions()
+    check_postreview()
+    check_replications()
     check_public_paths()
     check_metadata()
     files = check_manifest()
     print(
         f"verified {files} files; 384 groups, {candidates} candidates, "
         f"{selected} Q25 selections ({modes['LOCAL']} LOCAL, "
-        f"{modes['LOCAL_GRAPH']} LOCAL_GRAPH), and {pairs} interactions"
+        f"{modes['LOCAL_GRAPH']} LOCAL_GRAPH), {pairs} interactions, post-review confirmation, "
+        "and two fresh Q25 campaigns"
     )