v1.0.1 public package + figures + paper PDF (DOI 10.5281/zenodo.21875513)

.gitattributes

@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+paper/PAPER.pdf filter=lfs diff=lfs merge=lfs -text

CITATION.cff

@@ -0,0 +1,29 @@
+cff-version: 1.2.0
+message: "If you use this study or its processed results, please cite this record."
+title: "Quantization Regime and Governed Routing: A 2x2 Study of QAT Q4_0 versus PTQ Q4_K_M for Gemma 4 12B and 26B MoE under a Fixed Governance Stack"
+type: article
+authors:
+  - family-names: Toeda
+    given-names: Taiko
+    orcid: "https://orcid.org/0009-0001-7267-0201"
+    affiliation: Independent Researcher
+version: 1.0.1
+date-released: 2026-08-11
+license: CC-BY-4.0
+identifiers:
+  - type: doi
+    value: "10.5281/zenodo.21875513"
+    description: Zenodo archival record (published 2026-08-11)
+repository-code: "https://github.com/mobius-style/gemma4-quant-regime-study"
+url: "https://doi.org/10.5281/zenodo.21875513"
+abstract: >-
+  A 2x2 study (quantization regime x released model pair) of governed
+  routing quality for Gemma 4 12B (dense, same-base pair) and a 26B-class
+  MoE released pair on one RTX 5070 Ti: within-pair regime effects are
+  marginal and oppositely signed; the regime x pair interaction — an
+  exploratory single-run endpoint — is +0.030 with a stem-clustered 95%
+  CI of [+0.010, +0.052], concentrated in the volatile-current family;
+  and only 23-32% of temperature-0 outputs are byte-identical across
+  regimes against 100% same-configuration repeat determinism in all four
+  cells. The license field describes the paper and data; analysis code is
+  Apache-2.0 (see LICENSE.md). Sequel to DOI 10.5281/zenodo.21860461.

CLAIMS_MATRIX.md

@@ -0,0 +1,29 @@
+# Claims matrix
+
+Labels: **Supported** = supported within the frozen local scope only;
+**Partially supported** = directionally relevant but incomplete;
+**Unsupported** = data reject or do not establish; **Requires replication** =
+needs a new condition. All inferential claims are exploratory (no
+preregistration, single run per cell).
+
+| Candidate claim | Status | Exact evidence / boundary |
+|---|---|---|
+| No cell separates by an operationally large margin on Core-500 | Supported as description | Spread 0.022 (11 tasks of 500). NOT a tie claim: no equivalence test was run, and 12B-PTQ vs 26B-PTQ is nominally significant (17 vs 6 discordant, p=0.035, uncorrected) |
+| The cells are statistically equivalent ("tied") | Unsupported | No TOST/equivalence test anywhere; one cross-model McNemar rejects at 0.05 |
+| QAT improves the clean dense 12B pair | Unsupported | Paired Δ −0.012 (QAT-only 1 vs PTQ-only 7), p=0.0703; Smoke agrees in sign (−0.020) |
+| QAT harms the clean dense 12B pair | Partially supported | Same evidence; marginal, single-run, clustered rows — directional only |
+| QAT improves the released 26B MoE pair | Partially supported | Paired Δ +0.018 (15 vs 6), p=0.0784; Smoke agrees (+0.030); pair impure (base revision / expert config unverifiable / offload confound) |
+| The regime × released-pair interaction is nonzero | **Supported as exploratory finding** | Delta-of-deltas +0.030; stem-clustered bootstrap 95% CI [+0.010, +0.052] (primary; 100 stems × 5 paraphrases) and task-iid CI [+0.008, +0.052]; single-run, no multiplicity correction |
+| The interaction is a pure QAT-recipe (or architecture) effect | Unsupported | 26B pair confounds regime with possible base revision, unverifiable expert config, and a CPU-offload compute path; only the 12B pair is same-base, all-GPU |
+| The interaction concentrates in volatile_current | Supported as description | Crossed family pattern 0.5314/0.4971 vs 0.4971/0.5314 (93↔87 of 175); the numeric identity of the crossed counts is coincidence; no test attached |
+| Quantization regime moves safety-critical failure rates | Unsupported | 0.000 in 7 of 8 runs; one over-verification event in 26B-PTQ Core-500 (0.008 = 1/125). Rare-event floor: rule-of-three 95% upper bounds ≈0.6–2.4 %/run |
+| num_ctx pinning (QAT profiles) is behavioral | Unsupported (controlled) | 12B-PTQ rerun at num_ctx 8192: 100/100 byte-identical vs default ctx. Control run on 12B only; extrapolated to 26B |
+| Regime choice preserves exact temperature-0 outputs | Unsupported | Cross-regime byte-identity 23.2–32.0 %; median first divergence 46–73 chars |
+| Observed cross-regime divergence is runtime nondeterminism | Unsupported (Smoke-100) | Same-configuration repeats 100/100 in ALL FOUR cells, incl. the CPU-offloaded 26B-PTQ, across run windows. Core-500 attribution is inference (no Core-500 repeats); a systematic offload contribution to 26B rows cannot be excluded |
+| QAT 26B remains a defensible production binding | Supported as operational inference | Best-or-tied on volatile_current (0.5314) and stale_premise (0.9600), fastest memory-fitting cell (1,613 ms); partially circular w.r.t. the same-day binding decision (two cells reused) — the anti-diagonal is the independent evidence |
+| PTQ 12b should remain the dense-tier rollback binding | Supported as operational inference | Clean-pair QAT effect ≤ 0 (p=0.07, directional); no reason to change the rollback target |
+| date_boundary degrades under QAT | Partially supported | −0.02 in both pairs (n=50 each); descriptive only |
+| Latency differs by regime | Unsupported (by design) | 26B-PTQ column is CPU-offload-dominated (26/74 split); no cross-memory-boundary latency claim is made |
+| Results generalize to other runtimes / stacks / content-quality benchmarks | Requires replication | One GPU, Ollama 0.30.6, one governance stack, routing tasks only |
+| The models have uncontaminated scores | Unsupported | Public weights; no training-corpus audit; suite prompts local and unpublished at run time |
+| The study is reproducible on the same artifacts | Supported | Pinned digests and hashes; 13 runs / 2,900 rows / 0 errored; every routing statistic incl. the family table recomputes from row level via the deposited script (families from the suites' authoritative family fields); safety rates quoted from archived harness scorer outputs |

ENVIRONMENT.md

@@ -0,0 +1,56 @@
+# Environment manifest — captured 2026-08-10 (JST) during run window
+ollama: ollama version is 0.30.6
+gpu: NVIDIA GeForce RTX 5070 Ti, 595.84, 16303 MiB
+kernel: 7.0.0-28-generic
+python: Python 3.10.14
+harness_git: b90d217b0598baa1c4e8db6d7cf828813a0aeaf7 (private working tree; the public harness lives at github.com/mobius-style/mmv → operate-fr-bench)
+
+## Model digests (ollama)
+gemma4:26b-a4b-it-qat                     2dd70431afed    15 GB     32 hours ago     
+gemma4:12b-it-qat                         38044be4f923    7.2 GB    32 hours ago     
+gemma4:12b                                4eb23ef187e2    7.6 GB    2 months ago     
+gemma4:26b                                5571076f3d70    17 GB     2 months ago     
+
+## Model cards (ollama show)
+### gemma4:12b
+  Model
+    architecture        gemma4    
+    parameters          11.9B     
+    context length      262144    
+    embedding length    3840      
+    quantization        Q4_K_M    
+    requires            0.30.5    
+
+### gemma4:12b-it-qat
+  Model
+    architecture        gemma4    
+    parameters          11.9B     
+    context length      262144    
+    embedding length    3840      
+    quantization        Q4_0      
+    requires            0.30.5    
+
+### gemma4:26b
+  Model
+    architecture        gemma4    
+    parameters          25.8B     
+    context length      262144    
+    embedding length    2816      
+    quantization        Q4_K_M    
+    requires            0.20.0    
+
+### gemma4:26b-a4b-it-qat
+  Model
+    architecture        gemma4    
+    parameters          25.2B     
+    context length      262144    
+    embedding length    2816      
+    quantization        Q4_0      
+    requires            0.30.5    
+
+
+## Suite/label hashes
+e962ac4f36fbfed1aa2e04093c3bcadbeafaf3817e8a0ee69f5afb9fa8f89883  <repo-root>/operate-fr-bench/data/smoke100.jsonl
+901e718b7070dfc2479e1ef6bfd18eabcacdeed12b645e7f3260b395a86a8e0f  <repo-root>/operate-fr-bench/data/core500.jsonl
+e86df675454bdbb9650659f593762951f5125140c7764bca209d6f652ec5bb70  <repo-root>/operate-fr-bench/data/labels/smoke100_route_labels.jsonl
+455d54d21eb2b62715cf2eb77f4d3b12ac500b95cae17edf722f858912c462bc  <repo-root>/operate-fr-bench/data/labels/core500_route_labels.jsonl

ETHICS_AND_DISCLOSURE.md

@@ -0,0 +1,20 @@
+# Ethics and AI-assistance disclosure
+
+This systems study uses locally stored open-weight model artifacts,
+project-authored routing suites, and machine-generated outputs. It
+collects no human-subject data, private conversations, personal data, or
+crowd-worker judgments. Route classification is a transparent rule-based
+detector; no human or LLM judge is used for the primary endpoint.
+
+Claude Code (an Anthropic coding agent, model Fable 5) assisted with
+experimental scripting and execution, data processing, statistical
+analysis, literature retrieval, manuscript language preparation, and
+dissemination preparation. AI systems are not authors and cannot accept
+scholarly accountability. The human author, Taiko Toeda, is responsible
+for the research design, the measurements accepted for publication,
+interpretation, citations, claims, final text, and the publication
+decision.
+
+No affiliation with Google, NVIDIA, or the Ollama project is asserted. No
+external sponsorship or model-vendor funding was identified in this local
+study record.

EXECUTIVE_SUMMARY.md

@@ -0,0 +1,79 @@
+# Executive summary
+
+## Conclusion
+
+On a fixed governance stack on one RTX 5070 Ti, we measured what
+quantization regime — QAT Q4_0 versus PTQ Q4_K_M — does to governed
+routing quality, in a 2×2 design over Gemma 4 12B (dense; a clean
+same-base pair) and a 26B-class MoE released pair (25.8B PTQ vs 25.2B
+QAT; regime confounded with a possible base revision). The answer: **on
+this stack the QAT premium is not a scalar — it differs in sign across
+the two released pairs** — and regime choice is a *behavioral* change
+even where quality scores barely move.
+
+| Cell (Core-500, N=500) | route_correctness | mean latency |
+|---|---:|---:|
+| 12B dense, PTQ Q4_K_M | **0.760** | 2,482 ms |
+| 12B dense, QAT Q4_0 | 0.748 | 2,626 ms |
+| 26B MoE, PTQ Q4_K_M | 0.738 | 3,252 ms † |
+| 26B MoE, QAT Q4_0 | 0.756 | **1,613 ms** |
+
+† the 26B PTQ artifact does not fit the 16 GB card and runs ~26/74
+CPU/GPU-split; its latency reflects offload and is excluded from all
+conclusions.
+
+- **Within-pair regime effects are marginal and oppositely signed**:
+  12B dense −0.012 under QAT (McNemar p=0.070); 26B pair +0.018
+  (p=0.078). Smoke-100 agrees in sign for both.
+- **The interaction — an exploratory, single-run headline — survives
+  clustering-robust analysis**: delta-of-deltas +0.030, stem-clustered
+  bootstrap 95% CI [+0.010, +0.052] (Core-500 is 100 stems × 5
+  paraphrases; naive task-iid CI [+0.008, +0.052]). Aggregate averaging
+  cancels the two effects and hides the structure. No equivalence test
+  was run on the cell aggregates; one cross-model contrast (12B-PTQ vs
+  26B-PTQ) is nominally significant (p=0.035).
+- **The interaction concentrates in the volatile-current family**, whose
+  four cells form a crossed pattern (12B: 0.5314→0.4971; 26B:
+  0.4971→0.5314; the numeric identity of the crossed counts is
+  coincidence, not a tested effect).
+- **Safety-critical rates: 0.000 in 7 of 8 runs**; one over-verification
+  event in 26B-PTQ Core-500 (0.008). With events this rare the data
+  bound, rather than establish, regime-invariance.
+- **Regime is a behavioral change**: only 23–32 % of same-task
+  temperature-0 outputs are byte-identical across regimes (median first
+  divergence 46–73 characters), against same-configuration repeat
+  baselines of **100/100 in all four cells** — including the
+  CPU-offloaded PTQ MoE cell, across run windows — and a 100/100
+  `num_ctx` byte-identity control on the 12B side. Hash-based caching,
+  audit replay, and golden-file tests do not survive a
+  quantization-regime swap.
+
+## Practical implications
+
+1. Choose quantization regimes per released artifact pair and per task
+   family, not by recipe reputation; aggregate "quality parity" can
+   conceal offsetting family-level shifts.
+2. On this stack, the QAT 26B artifact is best-or-tied on the two hardest
+   families while being the fastest memory-fitting cell — consistent with
+   its same-day adoption as a production binding (a partially circular
+   check: two of the four cells are that decision's own runs; the
+   anti-diagonal is the independent evidence). The dense-tier fallback is
+   better served by the **PTQ** 12B (clean-pair QAT effect ≤ 0).
+3. Treat any quantization swap as a behavioral migration with full
+   re-validation, not a drop-in.
+
+## Honest limits
+
+The 26B released pair may embed a base revision (25.8B vs 25.2B) and its
+expert configuration is unverifiable from runtime metadata; the PTQ 26B
+cell ran partially on CPU — so regime, revision, and compute path are
+confounded on that side; the 12B pair is clean. All inferential results
+are exploratory: no preregistration, single run per cell, ≥5 inferential
+quantities without multiplicity correction, paraphrase-clustered rows
+(effective N nearer 100 than 500). One GPU, one runtime (Ollama 0.30.6),
+one governance stack, routing tasks only; answer-content quality is
+unmeasured. The general "QAT beats PTQ" direction is vendor-documented
+prior art; the intended contribution is the governed-routing task class,
+the two-scale paired design exposing the signed interaction, and the
+regime-to-regime exact-output divergence measurement (the last is
+plausibly adjacent to community cross-quantization regression reports).

LICENSE-CODE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

LICENSE-PAPER-DATA.md

@@ -0,0 +1,18 @@
+# Creative Commons Attribution 4.0 International
+
+Copyright 2026 Taiko Toeda.
+
+Except where otherwise identified, the original paper, original prose
+documentation, original figures, aggregate tables, processed statistical
+outputs, and project-authored derived measurements in this public release are
+licensed under the Creative Commons Attribution 4.0 International License.
+
+- License summary: https://creativecommons.org/licenses/by/4.0/
+- Legal code: https://creativecommons.org/licenses/by/4.0/legalcode
+- SPDX identifier: `CC-BY-4.0`
+
+Attribution should identify Taiko Toeda, the work title, the version or DOI,
+and indicate whether changes were made. This notice does not relicense any
+third-party benchmark content, source code, model artifact, trademark, or other
+material identified in `THIRD_PARTY_DATA_LICENSES.md` or `NOTICE`.
+

LICENSE.md

@@ -0,0 +1,20 @@
+# License map
+
+Copyright 2026 Taiko Toeda.
+
+This repository deliberately uses a split license; no single blanket
+license applies to every file.
+
+- Original paper, prose documentation, aggregate tables, and processed
+  statistical outputs: Creative Commons Attribution 4.0 International
+  (`CC BY 4.0`), subject to `THIRD_PARTY_DATA_LICENSES.md`.
+- Original study code under `analysis/`: Apache License 2.0.
+- Third-party model artifacts, runtimes, and upstream material: their
+  respective upstream licenses; none are redistributed here.
+- Trademarks and project names remain the property of their respective
+  owners.
+
+`CITATION.cff`'s single `license` field refers to the paper-and-data
+record (CC BY 4.0); it does not override the Apache-2.0 grant on
+`analysis/`. See `LICENSE-PAPER-DATA.md`, `LICENSE-CODE`,
+`THIRD_PARTY_DATA_LICENSES.md`, and `NOTICE` for the operative notices.

LITERATURE_SEARCH.md

@@ -0,0 +1,44 @@
+# Literature search notes — 2026-08-10
+
+Scope: what is already claimed/measured about Gemma QAT (Q4_0) vs post-training
+quantization, to position this study's contribution. Web search performed
+2026-08-10; links verified at search time only (titles/summaries, not deep-read).
+
+## Vendor claims
+
+- Google, "Gemma 4 with quantization-aware training" (official blog):
+  QAT integrates quantization simulation into training so the 4-bit release
+  stays close to full-precision quality, vs. naive PTQ degradation.
+  https://blog.google/innovation-and-ai/technology/developers-tools/quantization-aware-training-gemma-4/
+- Google Developers Blog, "Gemma 3 QAT Models" (prior generation, same recipe
+  lineage): QAT Q4 approaches Q8-PTQ quality on consumer GPUs.
+  https://developers.googleblog.com/en/gemma-3-quantized-aware-trained-state-of-the-art-ai-to-consumer-gpus/
+- Gemma 4 model overview (ai.google.dev): official model cards, context
+  declarations. https://ai.google.dev/gemma/docs/core
+
+## Community measurements
+
+- DEV Community, "Gemma 4 QAT on a 1080 Ti": community-side accuracy
+  measurements; notes that naive conversion of QAT checkpoints to other
+  formats can lose part of the QAT benefit.
+  https://dev.to/sysoft/gemma-4-qat-on-a-1080-ti-what-quantization-aware-actually-buys-and-fitting-the-12b-on-8-gb-at-2bg5
+- Overview of Gemma 4 QAT (note.com/npaka): JA-language summary of the QAT
+  release. https://note.com/npaka/n/ndeef4df16dd2?hl=en
+
+## Surveys
+
+- "Art and Science of Quantizing Large-Scale Models" (arXiv:2409.11650):
+  QAT-vs-PTQ taxonomy. https://arxiv.org/pdf/2409.11650
+- "Resource-Efficient Language Models" (arXiv:2505.08620): quantization for
+  inference efficiency. https://arxiv.org/pdf/2505.08620
+
+## Gap this study addresses
+
+Published comparisons measure general-knowledge/reasoning benchmarks on raw
+models. We found no measurement of (a) QAT-vs-PTQ deltas **under a fixed
+governance stack** on **routing/abstention-style tasks** (answer-entitlement
+decisions rather than answer content), (b) the same contrast run **at two
+scales/architectures** (12B dense, ~26B sparse MoE) in a paired same-task
+design on one consumer GPU, or (c) regime-to-regime **exact-output
+divergence** rates for temperature-0 decoding. That is the niche of this
+study. We do NOT claim novelty for QAT-vs-PTQ quality direction in general.

NOTICE

@@ -0,0 +1,8 @@
+Gemma 4 Quantization-Regime Study (QAT Q4_0 vs PTQ Q4_K_M)
+Copyright 2026 Taiko Toeda
+
+This independent empirical study is not affiliated with or endorsed by
+Google, NVIDIA, or the Ollama project. Gemma is a family of open-weight
+models released by Google under the Gemma Terms of Use; model weights are
+not redistributed by this repository. "MMV" and "OPERATE-FR" refer to the
+author's own MOBIUS project components (github.com/mobius-style/mmv).

PAPER.md

@@ -0,0 +1,349 @@
+# Quantization Regime and Governed Routing: A 2×2 Study of QAT Q4_0 versus PTQ Q4_K_M for Gemma 4 12B (dense) and 26B MoE under a Fixed Governance Stack
+
+**Dates**: runs 2026-08-10, repeats and controls 2026-08-10/11 (Asia/Tokyo) · **Hardware**: single NVIDIA RTX 5070 Ti (16 GB) · **Backend**: Ollama 0.30.6
+**Series**: companion/sequel to the *Gemma 4 MTP Quality–Throughput Study* (repo `gemma4-mtp-quality-study`, DOI 10.5281/zenodo.21860461)
+**Status**: exploratory engineering study on a frozen local harness; single-run per cell; not preregistered; not an independent benchmark standard
+
+---
+
+## Abstract
+
+We measure what quantization regime — quantization-aware-trained Q4_0 (QAT)
+versus post-training-quantized Q4_K_M (PTQ) — is worth on *governed routing
+tasks*: answer-entitlement decisions (verify / date-bound / re-anchor /
+abstain) made under a fixed governance stack, rather than answer-content
+benchmarks. We run a 2×2 design (regime × released model pair) over Gemma 4
+12B (dense, 11.9B — a clean same-base pair) and a 26B-class MoE pair
+(25.8B PTQ vs 25.2B QAT released artifacts — a quasi-contrast that may
+embed a base revision), each on a 100-task Smoke-100 and a 500-task
+Core-500 candidate suite, all cells same-day on one consumer GPU with the
+governance stack bit-identical and temperature 0. Aggregate route
+correctness spans 0.738–0.760 on Core-500; we ran no equivalence test, and
+one cross-model contrast (12B-PTQ vs 26B-PTQ) is nominally significant
+(exact McNemar p=0.035). The paired within-model regime effects are
+individually marginal and **oppositely signed**: QAT −0.012 on the dense
+12B (p=0.070) and +0.018 on the MoE pair (p=0.078). The regime × pair
+**interaction** — this study's headline, an exploratory single-run finding
+— is +0.030 with a stem-clustered bootstrap 95% CI of **[+0.010, +0.052]**
+(primary; Core-500 is 100 stems × 5 paraphrases, so tasks are not
+independent) and [+0.008, +0.052] under naive task-iid resampling. The
+interaction concentrates almost entirely in the volatile-current family,
+where the two pairs' scores cross (0.5314/0.4971 vs 0.4971/0.5314).
+Attribution: with the released artifacts, regime is confounded on the MoE
+side with a possible base revision and with a CPU-offload compute path, so
+we claim conditionality on the *released artifact pair*, not on
+architecture per se. Meanwhile regime choice massively perturbs the *text*
+of temperature-0 outputs: only 23–32 % of same-task responses are
+byte-identical across regimes (median first divergence at 46–73
+characters), against same-configuration Smoke-100 repeat baselines of
+100/100 in **all four cells** (including the CPU-offloaded PTQ MoE cell,
+repeated across run windows) and a 100/100 byte-identity control for the
+one regime-correlated configuration difference (`num_ctx`). Safety-critical
+failure rates are 0.000 in seven of eight runs; the eighth (26B-PTQ
+Core-500) logged a single over-verification event (0.008 on stable
+controls). We conclude that on this stack the QAT premium is not a scalar —
+it differs in sign across the two released pairs at ≈3 percentage points,
+too small to survive aggregate averaging but visible in a paired crossed
+design — and that quantization regime should be treated as a *behavioral*
+change (different outputs), not merely a *quality* change (different
+scores).
+
+---
+
+## 1. Background and motivation
+
+The predecessor study (*Gemma 4 MTP Quality–Throughput Study*, 2026-08-09,
+same GPU) measured multi-token-prediction speed/quality trade-offs with
+both models already in QAT Q4_0 form, leaving the value of QAT itself
+unmeasured. Separately, the same day as this study, the local MMV-Medium
+governance release (MMV-M-RC3.3) underwent a model-binding update from the
+PTQ `gemma4:12b` to the QAT `gemma4:26b-a4b-it-qat` on head-to-head
+evidence in which regime and model were deliberately confounded
+(production binding vs. candidate binding). This study completes the
+missing cells: the anti-diagonal (12B QAT, 26B PTQ) of the resulting 2×2.
+Two of the four cells (12B-PTQ, 26B-QAT) are therefore **reused** from
+that production head-to-head (disclosed further in §3); the anti-diagonal
+cells are the independent addition.
+
+Vendor documentation claims QAT Q4 models retain near-full-precision
+quality where naive PTQ degrades (see `LITERATURE_SEARCH.md` in this package). Published
+comparisons, however, measure general knowledge/reasoning benchmarks on
+raw models. In the searches we ran (shallow, links-and-abstracts level —
+see the literature file's own caveat) we found no prior measurement of
+(a) QAT-vs-PTQ deltas on routing/abstention-style decisions under a
+governance stack, (b) the same contrast at two scales in a paired
+same-task design, or (c) regime-to-regime exact-output divergence at
+temperature 0. We do not claim novelty for the general QAT-beats-PTQ
+direction, and we flag (c) as plausibly adjacent to community regression
+reports on cross-quantization output changes.
+
+## 2. Design
+
+2×2 factorial, one variable per axis as far as the released artifacts allow:
+
+| | PTQ (Q4_K_M) | QAT (Q4_0) |
+|---|---|---|
+| **12B dense** | `gemma4:12b` (11.9B, emb 3840) | `gemma4:12b-it-qat` (11.9B, emb 3840) |
+| **26B MoE** | `gemma4:26b` (25.8B, emb 2816) | `gemma4:26b-a4b-it-qat` (25.2B, ~4B active, emb 2816) |
+
+- The **12B pair is the clean contrast**: identical parameter count and
+  embedding width; the artifacts differ only in quantization recipe.
+- The **26B pair is a quasi-contrast**: the released artifacts report
+  25.8B vs 25.2B parameters, so a base-revision difference cannot be
+  excluded. Moreover, Ollama metadata does not expose expert structure,
+  so we cannot verify that the two 26B artifacts share the same MoE
+  configuration (an earlier harness profile note describes `gemma4:26b`
+  as ~3.8B-active, vs ~4B for the QAT artifact). Findings attributed to
+  "regime" on the 26B side are strictly findings about the released
+  artifact pair.
+
+Constant across all eight primary runs: the MMV Large RC3.3 v3.1
+governance stack (route_transformer + post_validator + force_reanchor_v2)
+applied bit-identically; `think: false`; temperature 0.0; max_tokens 1024;
+one GPU; sequential execution; same Ollama build; same suite and label
+files (SHA-256 pinned in `ENVIRONMENT.md`).
+
+**The one regime-correlated configuration difference, and its control.**
+The two Q4_0 QAT profiles pin `num_ctx: 8192` because their Modelfiles
+declare a 262,144-token context whose KV allocation fails on a 16 GB card;
+the PTQ profiles run at the Ollama default context. Since this difference
+tracks the regime axis exactly, we ran a direct control: the 12B-PTQ
+Smoke-100 suite re-run with `num_ctx: 8192` produced **100/100
+byte-identical** outputs versus its default-context run. Within this
+study's prompt lengths, `num_ctx` is load feasibility, not behavior. (The
+control was run on the 12B cell only; we extrapolate it to the 26B pair.)
+
+## 3. Tasks, metrics, and statistics
+
+**Suites.** OPERATE-FR Smoke-100 (N=100) and Core-500 candidate (N=500, a
+controlled 5× neutral prompt-frame expansion of Smoke-100: **100 base
+stems × 5 paraphrase variants**, preserving the family distribution:
+volatile_current 175, stable_control 125, stale_premise_trap 75,
+date_boundary 50, query_neutrality 50, ambiguous_time_frame 25). Tasks
+probe *whether answering is warranted and in what mode* (verify,
+date-bound answer, re-anchor, abstain), classified by the harness's
+transparent rule-based detector. Because Core-500 tasks are paraphrase
+clusters, its 500 rows are **not independent observations**; statistics
+below address this explicitly.
+
+**Primary metric.** Route correctness (classified route ∈ allowed routes),
+recomputed from row-level results by the deposited analysis script, which
+also derives per-family tables from the suites' authoritative `family`
+fields. Exception: the safety-rate metrics (stale commitment, unsupported
+current claim, over-verification) are quoted from the harness scorer's
+summary outputs, which are archived alongside the raw rows.
+
+**Statistics.** Within-model regime effect: two-sided exact McNemar on
+paired same-task outcomes (the "exact" binomial computation is exact only
+under an independence assumption that paraphrase clustering violates;
+treat those p-values as approximate and anti-conservative). Interaction:
+delta-of-deltas on Core-500 with a **stem-clustered bootstrap** (resample
+the 100 stems, 10,000 resamples, seed 20260810) as the primary interval,
+and a naive task-iid bootstrap reported for comparison. Determinism
+baseline: same-configuration Smoke-100 repeats of **all four cells**.
+
+**Provenance and timeline (JST).** No preregistration exists; the 2×2
+design and the interaction endpoint were specified before the
+anti-diagonal cells ran, but analysis details were fixed at analysis time
+— treat every inferential statement as exploratory. 2026-08-10 early-to-mid
+afternoon: 12B-PTQ and 26B-QAT cells (reused from the same-day MMV-M
+production head-to-head; superseded earlier runs of the same PTQ cells
+from 2026-05/06 score equal or lower, so no favorable-selection is
+present). Evening: 12B-QAT and 26B-PTQ cells, then 12B repeats. Late
+night into 2026-08-11: 26B repeats and the `num_ctx` control. All 13 runs
+(8 primary + 4 repeats + 1 control = 2,900 rows) completed with **0
+errored rows**.
+
+## 4. Results
+
+### 4.1 Cell aggregates (Core-500, N=500)
+
+| cell | route_correctness | preferred_match | mean latency (ms) | mean resp. chars |
+|---|---:|---:|---:|---:|
+| 12B PTQ | **0.760** | 0.544 | 2,482 | 751 |
+| 12B QAT | 0.748 | 0.542 | 2,626 | 834 |
+| 26B PTQ | 0.738 | 0.540 | 3,252 † | 875 |
+| 26B QAT | 0.756 | 0.540 | **1,613** | 899 |
+
+† 26B PTQ (17 GB on disk) runs with a ~26/74 CPU/GPU split on the 16 GB
+card; its latency reflects offload, not model properties, and is excluded
+from all conclusions. Smoke-100 aggregates: 0.88 / 0.86 / 0.83 / 0.86 —
+same direction, with the two QAT cells tied.
+
+The Core-500 spread is 0.022 (11 tasks of 500). We ran **no equivalence
+test**, so "the cells are tied" is not a supported claim; indeed the one
+within-regime cross-model contrast, 12B-PTQ vs 26B-PTQ, is nominally
+significant (17 vs 6 discordant, exact McNemar p=0.035, uncorrected).
+What the aggregates do show is that no cell separates by a margin of
+operational consequence on this suite.
+
+Safety-critical rates (stale commitment, unsupported current claim,
+over-verification on stable controls): **0.000 in seven of eight runs**;
+the 26B-PTQ Core-500 run logged one over-verification event (0.008 =
+1/125 stable controls, consistent with its 0.976 stable_control entry in
+§4.4). With events this rare, the data bound rather than establish
+regime-invariance: by the rule of three, each run's 95% upper bound is
+≈0.6–2.4 % depending on the denominator.
+
+### 4.2 Paired regime effect within model
+
+| suite | pair | QAT-only correct | PTQ-only correct | Δ (QAT−PTQ) | McNemar p |
+|---|---|---:|---:|---:|---:|
+| Core-500 | 12B dense | 1 | 7 | **−0.012** | 0.0703 |
+| Core-500 | 26B MoE | 15 | 6 | **+0.018** | 0.0784 |
+| Smoke-100 | 12B dense | 0 | 2 | −0.020 | 0.50 |
+| Smoke-100 | 26B MoE | 5 | 2 | +0.030 | 0.4531 |
+
+Neither within-model effect reaches p<0.05, but both sit near the margin
+**with opposite signs**, and the smaller Smoke-100 suite points the same
+way in both pairs. The discordant tasks cluster by stem (the 12B pair's 8
+discordants come from 5 stems), which is one reason these p-values should
+be read as approximate (§3).
+
+### 4.3 The interaction is the finding — exploratory, but robust to clustering
+
+Delta-of-deltas (26B effect − 12B effect) on Core-500: **+0.030**.
+Primary stem-clustered bootstrap 95% CI: **[+0.010, +0.052]**; naive
+task-iid CI: [+0.008, +0.052]. Both exclude zero. This is a single-run,
+non-preregistered endpoint computed alongside four within-model tests and
+a family decomposition, with no multiplicity correction — we label it an
+exploratory finding that survived a clustering-robust re-analysis, not a
+confirmed effect. On this stack, QAT helps the released MoE pair and
+mildly hurts the clean dense pair; the two effects cancel almost exactly
+in any aggregate that averages over models.
+
+### 4.4 Family decomposition: a crossed pattern in volatile_current
+
+Core-500 route correctness by family (from the suites' authoritative
+family fields; reproduced by the deposited script):
+
+| family | n | 12B PTQ | 12B QAT | 26B PTQ | 26B QAT |
+|---|---:|---:|---:|---:|---:|
+| volatile_current | 175 | **0.5314** | 0.4971 | 0.4971 | **0.5314** |
+| ambiguous_time_frame | 25 | 0.560 | 0.600 | 0.480 | 0.600 |
+| date_boundary | 50 | 0.660 | 0.640 | 0.600 | 0.580 |
+| stale_premise_trap | 75 | 0.9467 | 0.9467 | 0.9600 | 0.9600 |
+| stable_control | 125 | 0.984 | 0.984 | 0.976 | 0.984 |
+| query_neutrality | 50 | 0.920 | 0.920 | 0.920 | 0.920 |
+
+The interaction concentrates in **volatile_current** (the family that
+decides whether a currency-sensitive question deserves verification or a
+dated hedge), where the four cells form a crossed pattern: the dense pair
+loses under QAT where the MoE pair gains. That the crossed rates land on
+numerically identical values (93↔87 correct of 175) is a coincidence of
+small discordant counts and carries no inferential content; no test is
+attached to this table. Easy families (stale_premise, stable, neutrality)
+are regime-invariant here; date_boundary drifts −0.02 under QAT in both
+pairs (n=50, descriptive only) — directionally consistent with the
+date_boundary watch item recorded in the same-day production
+binding-update freeze note (an internal record; noted for continuity, not
+as independent evidence).
+
+### 4.5 Regime is a behavioral change: exact-output divergence
+
+Byte-identical same-task responses across regimes at temperature 0:
+
+| suite | pair | exact match | rate | median first divergence (chars) |
+|---|---|---:|---:|---:|
+| Core-500 | 12B | 148/500 | 29.6 % | 73 |
+| Core-500 | 26B | 116/500 | 23.2 % | 69 |
+| Smoke-100 | 12B | 32/100 | 32.0 % | 46 |
+| Smoke-100 | 26B | 27/100 | 27.0 % | 57 |
+
+Against this, the same-configuration Smoke-100 repeat baseline is
+**100/100 byte-identical in all four cells** — including the
+CPU-offloaded 26B-PTQ cell, and with repeats separated from their base
+runs by 8–10 hours (crossing run windows and, for the 26B cells, a date
+boundary). The runtime is run-to-run deterministic here, so the
+cross-regime divergence on Smoke-100 is attributable to the weights (plus,
+for the 26B pair, the systematic compute-path difference — see §6.3). No
+Core-500 repeats were run; the attribution extends to Core-500 rows by
+inference from the Smoke-100 baselines, not by direct measurement.
+Quantization regime changes *what the model says* in 68–77 % of governed
+responses even while leaving *how often it is right* within ≈2 points —
+the same dissociation between string reproducibility and objective
+quality that the predecessor MTP study reported for speculative decoding,
+now shown for quantization recipes. Practical corollary: hash-based
+output caching, audit replay, or golden-file tests do **not** survive a
+quantization-regime swap, even at temperature 0.
+
+## 5. Discussion
+
+**The QAT premium is pair-conditional here, not scalar.** Vendor framing
+("QAT preserves quality") treats the QAT benefit as a property of the
+recipe. On governed routing tasks we find a signed interaction: ≈+2
+points (Core-500) for the released MoE pair, ≈−1 point for the clean
+dense pair. Three mechanisms are compatible with the data and not
+separable by this design: (i) 4-bit noise interacts differently with
+expert routing than with dense activations; (ii) the 26B released pair
+differs by more than regime (possible base revision, unverifiable expert
+configuration), so part of the 26B gain may be base improvement; (iii)
+the 26B-PTQ cell ran partially on CPU, so kernel-numerics differences are
+regime-correlated on that side. The 12B side carries none of these
+confounds — its artifacts are same-base, its cells all-GPU, and the
+`num_ctx` difference is controlled to byte-identity — and on that clean
+pair, QAT bought **nothing** on these tasks (directionally, it cost ≈1
+point at p=0.07).
+
+**Aggregates hide the effect.** The four Core-500 cell scores span 0.022.
+Only the paired-and-crossed design exposes the structure. This is a
+caution for model-selection practice: aggregate "quality parity"
+verdicts can conceal offsetting family-level and pair-level shifts.
+
+**Operational reading (inference, not test result).** Two of the four
+cells are the same runs that motivated the same-day production
+model-binding decision, so using this study to "validate" that decision
+is partially circular; the independent evidence is the anti-diagonal.
+With that caveat: the QAT 26B artifact remains the best-or-tied cell on
+the two hardest families while being the fastest memory-fitting cell,
+and the clean-pair result gives no reason to move the dense-tier
+rollback binding off PTQ `gemma4:12b`.
+
+## 6. Threats to validity
+
+1. **26B pair purity.** 25.8B vs 25.2B released artifacts; unverifiable
+   expert configuration; regime, base revision, and MoE structure are
+   confounded on that side. The headline survives only under the reading
+   "released artifact pairs" — which is the operationally relevant one,
+   and the reading used throughout this paper.
+2. **Paraphrase clustering.** Core-500 is 100 stems × 5 paraphrases;
+   McNemar "exact" p-values assume independence and are anti-conservative;
+   the interaction CI is therefore reported stem-clustered as primary.
+   Effective sample size is nearer 100 than 500.
+3. **Compute-path and window confounds (26B side).** 26B-PTQ ran ~26/74
+   CPU/GPU; CPU-vs-GPU numerics differ systematically, and the 26B cells
+   ran in different windows. Repeats (100/100 across windows, including
+   the offloaded cell) rule out run-to-run nondeterminism, but a
+   systematic offload contribution to the 26B regime effect and
+   divergence rows cannot be excluded.
+4. **Exploratory statistics.** No preregistration; ≥5 inferential
+   quantities computed; no multiplicity correction; single run per cell.
+5. **One stack, one GPU, one runtime.** Ollama 0.30.6, one governance
+   stack, rule-based route classification, routing tasks only.
+   Generalization to other runtimes, stacks, or answer-content quality is
+   unmeasured. The `num_ctx` control was run on the 12B cell only.
+6. **Family attribution.** date_boundary/ambiguous cells have n=50/25;
+   their drifts are descriptive only.
+7. **Contamination.** Public Gemma weights; suite prompts are local and
+   unpublished at run time, but no training-corpus audit exists.
+
+## 7. Reproducibility
+
+`REPRODUCE.md` gives the exact commands. Environment, model digests, and
+suite/label SHA-256 hashes: `ENVIRONMENT.md`. Row-level results
+(13 runs, 2,900 rows, 0 errored) are retained in the private study
+archive and identified by SHA-256; the public package ships processed
+tables and the deterministic analysis script (seed 20260810) that
+recomputes every routing statistic in this paper — including the §4.4
+family table — from row level. The safety rates in §4.1 are quoted from
+the harness scorer's archived summary outputs.
+
+## 8. Conclusion
+
+On a fixed governance stack and consumer GPU, quantization regime moved
+governed-routing quality by at most ≈±2 points within a released pair —
+below per-pair significance, but with a clustering-robust, family-
+concentrated, oppositely-signed interaction across the two pairs — and it
+moved the *bytes* of 68–77 % of temperature-0 outputs. Choose regimes per
+released artifact pair and per task family, not by recipe reputation;
+and treat any quantization swap as a behavioral migration, not a drop-in.

PUBLIC_MANIFEST.sha256

@@ -0,0 +1,22 @@
+26e5206197e0e4d3fd766d8827593a4e44562d1860395f1067164e93a49302e8  CITATION.cff
+c0833ff62b248c73b4393a04f87f16d341c635c003cf455e30d7ef4728529de1  CLAIMS_MATRIX.md
+296f15d12f033951677e8054c3239b38206420927882e248c367b95538f026f6  ENVIRONMENT.md
+eda0d8b13e38fad113251467460b52eb8dc06d417ab799d08dbcefae17c8063a  ETHICS_AND_DISCLOSURE.md
+0f7e90cd3f70c1fe22ef80b06cea0ef450f9b0ab762e1273ed6b3d03c532fe13  EXECUTIVE_SUMMARY.md
+cfc7749b96f63bd31c3c42b5c471bf756814053e847c10f3eb003417bc523d30  LICENSE-CODE
+dc22320360909705c84b359968bf5e12873bd10d5cc1cef8ae5287b54a6725c4  LICENSE-PAPER-DATA.md
+dd55a0f65f24543d2340fdb9b126960c4f83df6e512ca98e47730d30bb3e1637  LICENSE.md
+0c805ef7c3a987dec837abc2d5a17f342cb027023566d0df6fe383d8a10629f2  LITERATURE_SEARCH.md
+bb7ef7157df8df440393a826f4c291f88ea99c781ace2553dc47738399fa54bf  NOTICE
+37c80790837ce0979b08825087feb004504fe8613dcb2bd32e5133e9a05ac0ce  PAPER.md
+70df33df0ed84c712a5f46faa202a07a7de4f7af66f2a74b8dd92c7bba516f63  PUBLIC_RELEASE_METADATA.json
+9411cd7f42bd5416304a0fc21577fd5046cdf880adbf9a8bedf547e89a02597b  README.md
+124c993a94cb9dc0bd7038678b0a0d033c18bcc6d52ff7c6f2a2be4b1d3981db  REPRODUCE.md
+54432e7e6c03801f67774e233df80652a25021e464ef0b0f0b4f7d318bba9b3b  THIRD_PARTY_DATA_LICENSES.md
+297ffaca8bd5ebac411cd6f1f6d5c9b5a02910cc324522ced24acc46012538b5  analysis/analyze_2x2.py
+55aeaf1e4dc1f187bda3889bd18cb22f1f5a01fb07950f3d0135c3a22a23e910  data/RAW_ARTIFACT_HASHES.sha256
+3aec5443177782fb8afef7b346b7c4b2cf1f52ef67b7c45a276f44175c41ce0f  data/analysis_summary.json
+6931cf18c73da2e0d64f1c7ab453a04a3cc19fa1ac5a25a6b6af0afcc8334009  data/cell_metrics.csv
+cb07316e040d7942dceeaaa872fb2886fb7f2717e706fae948be65731766cc17  data/family_rates.csv
+d27bcfafa4116ea6d0620ba3b1a85b90c2d5b997a766ab5c43755842e360882b  data/paired_quant_effect.csv
+ebf903b82a19c32b3adcac248f5b330eebd7bde0e10f38c64efdccbde8e58648  data/regime_output_divergence.csv

PUBLIC_RELEASE_METADATA.json

@@ -0,0 +1,37 @@
+{
+  "name": "gemma4-quant-regime-study",
+  "version": "1.0.1",
+  "released": "2026-08-11",
+  "series": "sequel to gemma4-mtp-quality-study (DOI 10.5281/zenodo.21860461)",
+  "public_boundary": "verbatim prompts, generated outputs, raw row-level records, and local paths excluded; identified by SHA-256",
+  "runs": {
+    "suites": [
+      "smoke100",
+      "core500_candidate"
+    ],
+    "cells": 4,
+    "primary_runs": 8,
+    "repeat_runs": 4,
+    "control_runs": 1,
+    "rows": 2900,
+    "errored": 0
+  },
+  "headline": {
+    "interaction_delta_of_deltas": 0.03,
+    "ci95_stem_clustered": [
+      0.01,
+      0.052
+    ],
+    "ci95_task_iid": [
+      0.008,
+      0.052
+    ],
+    "cross_regime_exact_match_rate": [
+      0.232,
+      0.32
+    ],
+    "repeat_determinism": "100/100 x4 cells (Smoke-100)",
+    "num_ctx_control": "100/100 byte-identical (12B PTQ)"
+  },
+  "review": "3 independent adversarial reviewers (statistics / consistency / scope), 2026-08-10; all MAJOR findings remediated in v1.0.1"
+}

README.md

@@ -0,0 +1,79 @@
+---
+license: cc-by-4.0
+language:
+- en
+tags:
+- gemma
+- quantization
+- qat
+- benchmark
+- reproducibility
+- llm-governance
+- routing
+pretty_name: Gemma 4 Quantization-Regime Study (QAT Q4_0 vs PTQ Q4_K_M)
+size_categories:
+- n<1K
+---
+
+# Gemma 4 Quantization-Regime Study (QAT Q4_0 vs PTQ Q4_K_M)
+
+A same-day 2×2 study (quantization regime × released model pair) of
+governed routing quality for Gemma 4 12B IT (dense; a clean same-base
+pair) and a 26B-class MoE released pair on one RTX 5070 Ti, run under a
+fixed, bit-identical governance stack via Ollama.
+Companion/sequel to the *Gemma 4 MTP Quality–Throughput Study* (2026-08-09,
+DOI 10.5281/zenodo.21860461).
+
+Author: Taiko Toeda, Independent Researcher
+ORCID: https://orcid.org/0009-0001-7267-0201
+Version: 1.0.1
+DOI: [10.5281/zenodo.21875513](https://doi.org/10.5281/zenodo.21875513) (published 2026-08-11)
+
+## Result boundary
+
+Aggregate route correctness spanned 0.738–0.760 on Core-500 (N=500 rows =
+100 paraphrase stems × 5); no equivalence test was run, and one
+cross-model contrast (12B-PTQ vs 26B-PTQ) is nominally significant
+(p=0.035). The within-pair regime effects were individually marginal and
+oppositely signed — QAT −0.012 on the dense 12B (McNemar p=0.070) and
++0.018 on the 26B released pair (p=0.078) — and the regime × pair
+interaction, an exploratory single-run endpoint, was +0.030 with a
+stem-clustered bootstrap 95% CI of [+0.010, +0.052] (task-iid CI
+[+0.008, +0.052]), concentrated in the volatile-current task family
+(crossed pattern). Cross-regime byte-identity of temperature-0 outputs
+was only 23–32 %, against same-configuration Smoke-100 repeats of
+100/100 in all four cells (including the CPU-offloaded 26B-PTQ cell,
+across run windows) and a 100/100 `num_ctx` byte-identity control.
+Safety-critical failure rates were 0.000 in seven of eight runs, with a
+single over-verification event (0.008) in 26B-PTQ Core-500.
+
+These results apply only to the measured GPU, runtime (Ollama 0.30.6),
+the released model artifacts named in `ENVIRONMENT.md`, one governance
+stack, and routing/abstention-style tasks. The 26B released pair
+confounds regime with a possible base revision (25.8B vs 25.2B), an
+expert configuration unverifiable from runtime metadata, and a
+CPU-offload compute path; only the 12B pair is a clean same-base,
+all-GPU contrast. All inferential results are exploratory (no
+preregistration, single run per cell, no multiplicity correction). See
+`CLAIMS_MATRIX.md` before reusing any claim.
+
+## Public artifact boundary
+
+This public package excludes verbatim task prompts, generated output text,
+raw row-level request records, local paths, and the private internal
+archive. It retains the paper, processed measurements, per-family and
+paired statistics, environment and model digests, the deterministic
+analysis script, and SHA-256 hashes that identify every excluded raw
+artifact (`data/RAW_ARTIFACT_HASHES.sha256`). The evaluation harness is
+public at `github.com/mobius-style/mmv` (`operate-fr-bench/`); the
+OPERATE-FR suite items themselves are project-local and identified by hash
+in `ENVIRONMENT.md`.
+
+## Layout
+
+`PAPER.md` — full study. `EXECUTIVE_SUMMARY.md` — condensed findings.
+`CLAIMS_MATRIX.md` — claim-by-claim evidence boundaries.
+`REPRODUCE.md` — commands and artifact map. `ENVIRONMENT.md` — pinned
+environment, model digests, suite hashes. `data/` — processed tables and
+statistics. `analysis/` — deterministic analysis script (Apache-2.0).
+Licensing: see `LICENSE.md` (split license).

REPRODUCE.md

@@ -0,0 +1,62 @@
+# Reproduce
+
+Environment, model digests, and suite/label SHA-256: `ENVIRONMENT.md`.
+
+## Harness
+
+Public at `github.com/mobius-style/mmv`, directory `operate-fr-bench/`
+(AGPL-3.0). The four profiles used here (`gemma4_12b_…`,
+`gemma4_12b_qat_…`, `gemma4_26b_…`, `gemma4_26b_a4b_qat_…`, all
+`…route_transformer_plus_validator_v3_1`) share identical governance
+flags; both Q4_0 profiles pin `extra.num_ctx: 8192` because their
+Modelfiles declare a 262,144-token context whose KV allocation fails on a
+16 GB card (load feasibility only — suite prompts are far below 8,192
+tokens). `think: false` everywhere. Unload the previous model
+(`ollama stop <tag>`) between cells; a resident sibling model causes CUDA
+allocation failure.
+
+## Models (Ollama tags)
+
+`gemma4:12b` (PTQ Q4_K_M, 11.9B dense) · `gemma4:12b-it-qat` (QAT Q4_0,
+11.9B dense) · `gemma4:26b` (PTQ Q4_K_M, 25.8B MoE) ·
+`gemma4:26b-a4b-it-qat` (QAT Q4_0, 25.2B MoE, ~4B active). Content
+digests in `ENVIRONMENT.md`.
+
+## Runs (sequential, one cell at a time, from `operate-fr-bench/`)
+
+```bash
+python3 -m harness.run_eval --suite configs/suite_smoke.yaml   --profile <PROFILE> --out reports/<CELL>.jsonl
+python3 -m harness.score    --results reports/<CELL>.jsonl --labels data/labels/smoke100_route_labels.jsonl --out reports/<CELL>_smoke_summary.json
+python3 -m harness.run_eval --suite configs/suite_core500.yaml --profile <PROFILE> --out reports/<CELL>_core500.jsonl
+python3 -m harness.score    --results reports/<CELL>_core500.jsonl --labels data/labels/core500_route_labels.jsonl --out reports/<CELL>_core500_summary.json
+```
+
+Determinism baselines and controls: Smoke-100 repeats for **all four
+cells** (12B same-day; 26B 8–10 h later across a date boundary; all
+100/100 byte-identical, including the CPU-offloaded 26B-PTQ cell), plus a
+`num_ctx: 8192` control re-run of the 12B-PTQ cell (100/100 byte-identical
+vs default context).
+
+## Suite availability
+
+The OPERATE-FR Smoke-100 / Core-500-candidate items are project-local and
+not redistributed (they are identified by SHA-256 in `ENVIRONMENT.md`).
+Reproduction on the same items requires access to the private suite
+archive; reproduction of the *design* on other routing suites requires
+only the public harness.
+
+## Analysis
+
+```bash
+OPFR_BENCH_ROOT=/path/to/operate-fr-bench python3 analysis/analyze_2x2.py
+```
+
+Deterministic (bootstrap seed 20260810; the interaction is reported with
+a stem-clustered CI — primary, since Core-500 is 100 stems × 5
+paraphrases — and a task-iid CI). Regenerates every routing table in
+`data/` from row level, including the per-family table (families read
+from the suites' authoritative `family` fields). The safety rates quoted
+in the paper come from the harness scorer's summary outputs, archived
+privately with the raw rows. Row-level artifacts (13 runs, 2,900 rows, 0
+errored) are excluded from the public package (see `README.md`); their
+SHA-256 hashes are in `data/RAW_ARTIFACT_HASHES.sha256`.

THIRD_PARTY_DATA_LICENSES.md

@@ -0,0 +1,13 @@
+# Third-party data and code license inventory
+
+Audit date: 2026-08-10. This study's third-party surface is small: no
+third-party benchmark items are used (the OPERATE-FR suites are
+project-authored), and no third-party source is vendored.
+
+| Component | Exact experimental source | License basis | Redistributed here? | Public treatment |
+|---|---|---|---|---|
+| Gemma 4 12B IT / 12B IT QAT / 26B / 26B-A4B IT QAT weights | Ollama library tags pinned in `ENVIRONMENT.md` | Gemma Terms of Use (Google) | No | Tags + content digests only |
+| Ollama runtime 0.30.6 | ollama.com install | MIT | No | Version pin only |
+| operate-fr-bench harness | github.com/mobius-style/mmv (project-owned) | AGPL-3.0 | No (referenced) | Public upstream; profiles named in `REPRODUCE.md` |
+| OPERATE-FR Smoke-100 / Core-500 suites | project-local (author-owned) | Author's own work, unpublished | No | SHA-256 hashes in `ENVIRONMENT.md` |
+| Governance stack (route_transformer / post_validator / force_reanchor_v2) | part of the public mmv repository | AGPL-3.0 (project-owned) | No (referenced) | Named profiles, public upstream |

analysis/analyze_2x2.py

@@ -0,0 +1,258 @@
+#!/usr/bin/env python3
+"""2x2 quantization-regime study analysis (QAT Q4_0 vs PTQ Q4_K_M
+across Gemma 4 12B dense and 26B MoE, governance routing tasks).
+
+Reads operate-fr-bench row-level results + labels + suite datasets
+(authoritative per-task family and smoke-stem ids); writes CSVs and a JSON
+summary into ../data/. All routing statistics are recomputed from row
+level. Exception: the safety-rate metrics (stale_commitment,
+unsupported_current_claim, over_verification) quoted in the paper come
+from the harness scorer's *_summary.json outputs, not from this script.
+"""
+import csv
+import json
+import random
+from math import comb
+from pathlib import Path
+
+import os
+R = Path(os.environ.get("OPFR_BENCH_ROOT", "."))  # root of a local operate-fr-bench checkout with reports/
+OUT = Path(os.environ.get("STUDY_DATA_OUT", Path(__file__).resolve().parent.parent / "data"))
+OUT.mkdir(exist_ok=True)
+
+CELLS = {
+    # (size, regime) -> {suite: results file}
+    ("12B", "PTQ"): {
+        "smoke": "gemma4_12b_route_transformer_plus_validator_v3_1_rerun_2026-08-10.jsonl",
+        "core500": "gemma4_12b_core500_candidate_rerun_2026-08-10.jsonl",
+    },
+    ("12B", "QAT"): {
+        "smoke": "gemma4_12b_qat_route_transformer_plus_validator_v3_1.jsonl",
+        "core500": "gemma4_12b_qat_core500_candidate.jsonl",
+    },
+    ("26B", "PTQ"): {
+        "smoke": "gemma4_26b_route_transformer_plus_validator_v3_1_rerun_2026-08-10.jsonl",
+        "core500": "gemma4_26b_core500_candidate_rerun_2026-08-10.jsonl",
+    },
+    ("26B", "QAT"): {
+        "smoke": "gemma4_26b_a4b_qat_route_transformer_plus_validator_v3_1.jsonl",
+        "core500": "gemma4_26b_a4b_qat_core500_candidate.jsonl",
+    },
+}
+LABELS = {
+    "smoke": "data/labels/smoke100_route_labels.jsonl",
+    "core500": "data/labels/core500_route_labels.jsonl",
+}
+DATASETS = {  # authoritative per-task family + stem ids
+    "smoke": "data/smoke100.jsonl",
+    "core500": "data/core500.jsonl",
+}
+REPEATS = {  # same-day second runs for ordinary-nondeterminism baseline
+    ("12B", "PTQ"): "gemma4_12b_route_transformer_plus_validator_v3_1_repeat2.jsonl",
+    ("12B", "QAT"): "gemma4_12b_qat_route_transformer_plus_validator_v3_1_repeat2.jsonl",
+    ("26B", "QAT"): "gemma4_26b_a4b_qat_route_transformer_plus_validator_v3_1_repeat2.jsonl",
+    ("26B", "PTQ"): "gemma4_26b_route_transformer_plus_validator_v3_1_repeat2.jsonl",
+}
+
+
+def load_labels(suite):
+    out = {}
+    for ln in open(R / LABELS[suite], encoding="utf-8"):
+        d = json.loads(ln)
+        out[d["task_id"]] = d
+    return out
+
+
+def load_dataset_meta(suite):
+    """Authoritative task metadata: family + smoke stem id (for clustering)."""
+    meta = {}
+    for ln in open(R / DATASETS[suite], encoding="utf-8"):
+        d = json.loads(ln)
+        meta[d["id"]] = {
+            "family": d["family"],
+            "stem": d.get("source_smoke100_id") or d["id"],
+        }
+    return meta
+
+
+def load_cell(fname, suite):
+    rows = {}
+    p = R / "reports" / fname
+    for ln in open(p, encoding="utf-8"):
+        d = json.loads(ln)
+        rows[d["task_id"]] = d
+    return rows
+
+
+def correct(row, lab):
+    return row["classified_route"] in lab["allowed_routes"]
+
+
+def mcnemar_exact(n01, n10):
+    n = n01 + n10
+    if n == 0:
+        return 1.0
+    k = min(n01, n10)
+    p = sum(comb(n, i) for i in range(0, k + 1)) * 2 / 2 ** n
+    return min(1.0, p)
+
+
+def main():
+    labels = {s: load_labels(s) for s in ("smoke", "core500")}
+    metas = {s: load_dataset_meta(s) for s in ("smoke", "core500")}
+    cells = {}
+    for key, files in CELLS.items():
+        cells[key] = {s: load_cell(f, s) for s, f in files.items()}
+
+    summary = {"cells": {}, "paired": {}, "interaction": {}, "divergence": {},
+               "repeats": {}}
+
+    # ── per-cell metrics ──
+    cell_csv = [["size", "regime", "suite", "n", "errored",
+                 "route_correct", "rate", "preferred_match",
+                 "mean_latency_ms", "mean_resp_chars"]]
+    fam_csv = [["size", "regime", "suite", "family", "n", "correct", "rate"]]
+    for (size, regime), suites in cells.items():
+        for suite, rows in suites.items():
+            lab = labels[suite]
+            ids = sorted(lab)
+            nerr = sum(1 for t in ids if rows[t].get("error"))
+            ncor = sum(1 for t in ids if correct(rows[t], lab[t]))
+            npref = sum(1 for t in ids
+                        if rows[t]["classified_route"] == lab[t].get("preferred_route"))
+            lat = [rows[t].get("latency_ms") or 0 for t in ids]
+            rl = [len(rows[t].get("response_text") or "") for t in ids]
+            cell_csv.append([size, regime, suite, len(ids), nerr, ncor,
+                             round(ncor / len(ids), 4), round(npref / len(ids), 4),
+                             round(sum(lat) / len(lat), 1), round(sum(rl) / len(rl), 1)])
+            summary["cells"][f"{size}_{regime}_{suite}"] = {
+                "n": len(ids), "errored": nerr,
+                "route_correctness": round(ncor / len(ids), 4),
+                "preferred_match": round(npref / len(ids), 4),
+                "mean_latency_ms": round(sum(lat) / len(lat), 1),
+            }
+            meta = metas[suite]
+            fams = {}
+            for t in ids:
+                fams.setdefault(meta[t]["family"], []).append(t)
+            for f, tt in sorted(fams.items()):
+                c = sum(1 for t in tt if correct(rows[t], lab[t]))
+                fam_csv.append([size, regime, suite, f, len(tt), c,
+                                round(c / len(tt), 4)])
+
+    # ── paired quant effect within size (core500 primary; smoke secondary) ──
+    paired_csv = [["suite", "size", "n_pairs", "qat_only_correct",
+                   "ptq_only_correct", "delta_rate", "mcnemar_p"]]
+    deltas = {}
+    for suite in ("core500", "smoke"):
+        lab = labels[suite]
+        ids = sorted(lab)
+        for size in ("12B", "26B"):
+            q, p_ = cells[(size, "QAT")][suite], cells[(size, "PTQ")][suite]
+            n10 = sum(1 for t in ids if correct(q[t], lab[t]) and not correct(p_[t], lab[t]))
+            n01 = sum(1 for t in ids if correct(p_[t], lab[t]) and not correct(q[t], lab[t]))
+            delta = (n10 - n01) / len(ids)
+            pv = mcnemar_exact(n01, n10)
+            paired_csv.append([suite, size, len(ids), n10, n01,
+                               round(delta, 4), round(pv, 4)])
+            summary["paired"][f"{suite}_{size}"] = {
+                "qat_only": n10, "ptq_only": n01,
+                "delta_rate": round(delta, 4), "mcnemar_p": round(pv, 4)}
+            if suite == "core500":
+                deltas[size] = [
+                    (1 if correct(q[t], lab[t]) else 0) -
+                    (1 if correct(p_[t], lab[t]) else 0) for t in ids]
+
+    # ── interaction (delta-of-deltas, task bootstrap, core500) ──
+    rng = random.Random(20260810)
+    obs = sum(deltas["26B"]) / len(deltas["26B"]) - sum(deltas["12B"]) / len(deltas["12B"])
+    n = len(deltas["12B"])
+    ids_c = sorted(labels["core500"])
+    boots = []
+    for _ in range(10000):
+        idx = [rng.randrange(n) for _ in range(n)]
+        boots.append((sum(deltas["26B"][i] for i in idx)
+                      - sum(deltas["12B"][i] for i in idx)) / n)
+    boots.sort()
+    # stem-clustered bootstrap: Core-500 is 100 smoke stems x 5 paraphrase
+    # variants, so tasks are NOT independent. Resample the 100 stems.
+    meta_c = metas["core500"]
+    stems = {}
+    for i, t_ in enumerate(ids_c):
+        stems.setdefault(meta_c[t_]["stem"], []).append(i)
+    stem_keys = sorted(stems)
+    rng2 = random.Random(20260810)
+    cboots = []
+    for _ in range(10000):
+        picks = [stem_keys[rng2.randrange(len(stem_keys))] for _ in stem_keys]
+        tot26 = tot12 = cnt = 0
+        for s in picks:
+            for i in stems[s]:
+                tot26 += deltas["26B"][i]
+                tot12 += deltas["12B"][i]
+                cnt += 1
+        cboots.append((tot26 - tot12) / cnt)
+    cboots.sort()
+    summary["interaction"]["core500"] = {
+        "delta_of_deltas": round(obs, 4),
+        "ci95_task_iid": [round(boots[249], 4), round(boots[9749], 4)],
+        "ci95_stem_clustered": [round(cboots[249], 4), round(cboots[9749], 4)],
+        "n_stems": len(stem_keys),
+        "note": ("positive = QAT helps 26B more than 12B; stem-clustered CI "
+                 "is primary (5x paraphrase dependence)"),
+    }
+
+    # ── exact-output divergence between regimes (same size, same task) ──
+    div_csv = [["suite", "size", "n", "exact_match", "rate",
+                "median_first_divergence_char"]]
+    for suite in ("core500", "smoke"):
+        lab = labels[suite]
+        ids = sorted(lab)
+        for size in ("12B", "26B"):
+            q, p_ = cells[(size, "QAT")][suite], cells[(size, "PTQ")][suite]
+            same = 0
+            first = []
+            for t in ids:
+                a = q[t].get("response_text") or ""
+                b = p_[t].get("response_text") or ""
+                if a == b:
+                    same += 1
+                else:
+                    k = next((i for i, (x, y) in enumerate(zip(a, b)) if x != y),
+                             min(len(a), len(b)))
+                    first.append(k)
+            first.sort()
+            med = first[len(first) // 2] if first else None
+            div_csv.append([suite, size, len(ids), same,
+                            round(same / len(ids), 4), med])
+            summary["divergence"][f"{suite}_{size}"] = {
+                "exact_match": same, "rate": round(same / len(ids), 4),
+                "median_first_divergence_char": med}
+
+    # ── ordinary-nondeterminism baseline (same profile, same day, rerun) ──
+    for (size, regime), fname in REPEATS.items():
+        p = R / "reports" / fname
+        if not p.exists():
+            continue
+        rep = load_cell(fname, "smoke")
+        base = cells[(size, regime)]["smoke"]
+        lab = labels["smoke"]
+        same = sum(1 for t in sorted(lab)
+                   if (rep[t].get("response_text") or "") ==
+                      (base[t].get("response_text") or ""))
+        summary["repeats"][f"{size}_{regime}_smoke"] = {
+            "exact_match": same, "n": len(lab)}
+
+    for name, rows in [("cell_metrics.csv", cell_csv),
+                       ("family_rates.csv", fam_csv),
+                       ("paired_quant_effect.csv", paired_csv),
+                       ("regime_output_divergence.csv", div_csv)]:
+        with open(OUT / name, "w", newline="", encoding="utf-8") as fh:
+            csv.writer(fh).writerows(rows)
+    (OUT / "analysis_summary.json").write_text(
+        json.dumps(summary, indent=2, ensure_ascii=False), encoding="utf-8")
+    print(json.dumps(summary, indent=2, ensure_ascii=False))
+
+
+if __name__ == "__main__":
+    main()

data/RAW_ARTIFACT_HASHES.sha256

@@ -0,0 +1,24 @@
+# SHA-256 of the private row-level artifacts (content not redistributed;
+# verbatim task prompts and generated outputs are excluded from the public
+# package by policy — see README.md 'Public artifact boundary')
+b97b5663b65ed9f2a0bb96f15850c14b0e4adcf29445e59d98dd6bb912e147eb  data/raw/ctx_control_12b.jsonl
+34534f55a82b075dd0ee4422b8add300a42e2db53e2b7b48fd8d39053adefcf9  data/raw/gemma4_12b_core500_candidate_rerun_2026-08-10.jsonl
+a297a0a3e4bc9255e0bd727c9910f45f2bc56e6d1440311ca23f42b6281c1951  data/raw/gemma4_12b_qat_core500_candidate.jsonl
+bca0c680b94f2255bbdfdec3fe005a40c124784e7baf79b0ae6c36d53a8fc9eb  data/raw/gemma4_12b_qat_route_transformer_plus_validator_v3_1.jsonl
+87560e80d1dce7087a8cd38f51c48ca755f50aa8371603bfa1793d8c30d01b7e  data/raw/gemma4_12b_qat_route_transformer_plus_validator_v3_1_repeat2.jsonl
+e2544dd2bc90dfe77b051e0142a319de30be4c3b9b0dd1edeba8ee921def9bac  data/raw/gemma4_12b_route_transformer_plus_validator_v3_1_repeat2.jsonl
+b45f50f2b2f5acc38389dc1b890795752bf3bed2a6fc2a171658bd42aea6a360  data/raw/gemma4_12b_route_transformer_plus_validator_v3_1_rerun_2026-08-10.jsonl
+3e095db83a7476c9a742f74e2bc4e3f86537906a41d2c61095c0d230c024293c  data/raw/gemma4_26b_a4b_qat_core500_candidate.jsonl
+a0c1b0c310d0caac13d8ad2bf69f69e72e6581297220ad60b951066187372fc4  data/raw/gemma4_26b_a4b_qat_route_transformer_plus_validator_v3_1.jsonl
+fab281c1e921a87afa01f6c6770081eb7c439a08767dec980e8f5544d5de871c  data/raw/gemma4_26b_a4b_qat_route_transformer_plus_validator_v3_1_repeat2.jsonl
+cb0dd700e516ee749749ed621435dae2c4b5113aa36f18ca29618bac1634b527  data/raw/gemma4_26b_core500_candidate_rerun_2026-08-10.jsonl
+bfb49dc3dc88ce360dee9302fa4de092ff0aeedee29d4e11ef6fffdf06d29eb4  data/raw/gemma4_26b_route_transformer_plus_validator_v3_1_repeat2.jsonl
+ef52a5cf1cb6d59c3a7f71147c80ff901de402e232f446d6813e57a97baa4dcc  data/raw/gemma4_26b_route_transformer_plus_validator_v3_1_rerun_2026-08-10.jsonl
+39f5a30ee0ba10af0ac43ca72d588e2f19858e1c093b3c50b97a202a996cd306  data/raw/gemma4_12b_core500_candidate_rerun_2026-08-10_summary.json
+a11d8531c9e83b8869b6b1a204e09e196f1b34bcb0f9a8ead057393aeedc87c7  data/raw/gemma4_12b_qat_core500_candidate_summary.json
+4e4274c57f45499ef6c1c896de27d133dc977539de7f865a165d46cd32d07ce1  data/raw/gemma4_12b_qat_route_transformer_plus_validator_v3_1_summary.json
+a207ab7bc877b2e7cad3fc8c687883f93e11c57ebf78bbe2e505e1f60480a053  data/raw/gemma4_12b_route_transformer_plus_validator_v3_1_rerun_2026-08-10_summary.json
+a4a8ec110707ffa5e2ce613076823482bc5cf99d2b30df1c6ea1179311a0c140  data/raw/gemma4_26b_a4b_qat_core500_candidate_summary.json
+1a9602d4ef75d0eb76cda8487f97590d68ede9b4411e9cda0aa4a3e890412dfd  data/raw/gemma4_26b_a4b_qat_route_transformer_plus_validator_v3_1_summary.json
+4b793756f45d195e55a010c28c0f5d647c218fafab9441484c7838458ff79c24  data/raw/gemma4_26b_core500_candidate_rerun_2026-08-10_summary.json
+a5afb912a1b52ebc9f6d878e469f36d9227192f3b2fa09f19bd0a440ba601e8c  data/raw/gemma4_26b_route_transformer_plus_validator_v3_1_rerun_2026-08-10_summary.json

data/analysis_summary.json

@@ -0,0 +1,141 @@
+{
+  "cells": {
+    "12B_PTQ_smoke": {
+      "n": 100,
+      "errored": 0,
+      "route_correctness": 0.88,
+      "preferred_match": 0.57,
+      "mean_latency_ms": 2954.5
+    },
+    "12B_PTQ_core500": {
+      "n": 500,
+      "errored": 0,
+      "route_correctness": 0.76,
+      "preferred_match": 0.544,
+      "mean_latency_ms": 2481.5
+    },
+    "12B_QAT_smoke": {
+      "n": 100,
+      "errored": 0,
+      "route_correctness": 0.86,
+      "preferred_match": 0.58,
+      "mean_latency_ms": 2888.0
+    },
+    "12B_QAT_core500": {
+      "n": 500,
+      "errored": 0,
+      "route_correctness": 0.748,
+      "preferred_match": 0.542,
+      "mean_latency_ms": 2626.3
+    },
+    "26B_PTQ_smoke": {
+      "n": 100,
+      "errored": 0,
+      "route_correctness": 0.83,
+      "preferred_match": 0.58,
+      "mean_latency_ms": 3739.4
+    },
+    "26B_PTQ_core500": {
+      "n": 500,
+      "errored": 0,
+      "route_correctness": 0.738,
+      "preferred_match": 0.54,
+      "mean_latency_ms": 3251.8
+    },
+    "26B_QAT_smoke": {
+      "n": 100,
+      "errored": 0,
+      "route_correctness": 0.86,
+      "preferred_match": 0.56,
+      "mean_latency_ms": 1862.5
+    },
+    "26B_QAT_core500": {
+      "n": 500,
+      "errored": 0,
+      "route_correctness": 0.756,
+      "preferred_match": 0.54,
+      "mean_latency_ms": 1613.4
+    }
+  },
+  "paired": {
+    "core500_12B": {
+      "qat_only": 1,
+      "ptq_only": 7,
+      "delta_rate": -0.012,
+      "mcnemar_p": 0.0703
+    },
+    "core500_26B": {
+      "qat_only": 15,
+      "ptq_only": 6,
+      "delta_rate": 0.018,
+      "mcnemar_p": 0.0784
+    },
+    "smoke_12B": {
+      "qat_only": 0,
+      "ptq_only": 2,
+      "delta_rate": -0.02,
+      "mcnemar_p": 0.5
+    },
+    "smoke_26B": {
+      "qat_only": 5,
+      "ptq_only": 2,
+      "delta_rate": 0.03,
+      "mcnemar_p": 0.4531
+    }
+  },
+  "interaction": {
+    "core500": {
+      "delta_of_deltas": 0.03,
+      "ci95_task_iid": [
+        0.008,
+        0.052
+      ],
+      "ci95_stem_clustered": [
+        0.01,
+        0.052
+      ],
+      "n_stems": 100,
+      "note": "positive = QAT helps 26B more than 12B; stem-clustered CI is primary (5x paraphrase dependence)"
+    }
+  },
+  "divergence": {
+    "core500_12B": {
+      "exact_match": 148,
+      "rate": 0.296,
+      "median_first_divergence_char": 73
+    },
+    "core500_26B": {
+      "exact_match": 116,
+      "rate": 0.232,
+      "median_first_divergence_char": 69
+    },
+    "smoke_12B": {
+      "exact_match": 32,
+      "rate": 0.32,
+      "median_first_divergence_char": 46
+    },
+    "smoke_26B": {
+      "exact_match": 27,
+      "rate": 0.27,
+      "median_first_divergence_char": 57
+    }
+  },
+  "repeats": {
+    "12B_PTQ_smoke": {
+      "exact_match": 100,
+      "n": 100
+    },
+    "12B_QAT_smoke": {
+      "exact_match": 100,
+      "n": 100
+    },
+    "26B_QAT_smoke": {
+      "exact_match": 100,
+      "n": 100
+    },
+    "26B_PTQ_smoke": {
+      "exact_match": 100,
+      "n": 100
+    }
+  }
+}

data/cell_metrics.csv

@@ -0,0 +1,9 @@
+size,regime,suite,n,errored,route_correct,rate,preferred_match,mean_latency_ms,mean_resp_chars
+12B,PTQ,smoke,100,0,88,0.88,0.57,2954.5,790.9
+12B,PTQ,core500,500,0,380,0.76,0.544,2481.5,750.6
+12B,QAT,smoke,100,0,86,0.86,0.58,2888.0,857.8
+12B,QAT,core500,500,0,374,0.748,0.542,2626.3,834.3
+26B,PTQ,smoke,100,0,83,0.83,0.58,3739.4,922.4
+26B,PTQ,core500,500,0,369,0.738,0.54,3251.8,874.8
+26B,QAT,smoke,100,0,86,0.86,0.56,1862.5,926.1
+26B,QAT,core500,500,0,378,0.756,0.54,1613.4,898.6

data/family_rates.csv

@@ -0,0 +1,49 @@
+size,regime,suite,family,n,correct,rate
+12B,PTQ,smoke,ambiguous_time_frame,5,4,0.8
+12B,PTQ,smoke,date_boundary,10,9,0.9
+12B,PTQ,smoke,query_neutrality,10,10,1.0
+12B,PTQ,smoke,stable_control,25,25,1.0
+12B,PTQ,smoke,stale_premise_trap,15,14,0.9333
+12B,PTQ,smoke,volatile_current,35,26,0.7429
+12B,PTQ,core500,ambiguous_time_frame,25,14,0.56
+12B,PTQ,core500,date_boundary,50,33,0.66
+12B,PTQ,core500,query_neutrality,50,46,0.92
+12B,PTQ,core500,stable_control,125,123,0.984
+12B,PTQ,core500,stale_premise_trap,75,71,0.9467
+12B,PTQ,core500,volatile_current,175,93,0.5314
+12B,QAT,smoke,ambiguous_time_frame,5,4,0.8
+12B,QAT,smoke,date_boundary,10,9,0.9
+12B,QAT,smoke,query_neutrality,10,10,1.0
+12B,QAT,smoke,stable_control,25,25,1.0
+12B,QAT,smoke,stale_premise_trap,15,14,0.9333
+12B,QAT,smoke,volatile_current,35,24,0.6857
+12B,QAT,core500,ambiguous_time_frame,25,15,0.6
+12B,QAT,core500,date_boundary,50,32,0.64
+12B,QAT,core500,query_neutrality,50,46,0.92
+12B,QAT,core500,stable_control,125,123,0.984
+12B,QAT,core500,stale_premise_trap,75,71,0.9467
+12B,QAT,core500,volatile_current,175,87,0.4971
+26B,PTQ,smoke,ambiguous_time_frame,5,3,0.6
+26B,PTQ,smoke,date_boundary,10,9,0.9
+26B,PTQ,smoke,query_neutrality,10,10,1.0
+26B,PTQ,smoke,stable_control,25,25,1.0
+26B,PTQ,smoke,stale_premise_trap,15,14,0.9333
+26B,PTQ,smoke,volatile_current,35,22,0.6286
+26B,PTQ,core500,ambiguous_time_frame,25,12,0.48
+26B,PTQ,core500,date_boundary,50,30,0.6
+26B,PTQ,core500,query_neutrality,50,46,0.92
+26B,PTQ,core500,stable_control,125,122,0.976
+26B,PTQ,core500,stale_premise_trap,75,72,0.96
+26B,PTQ,core500,volatile_current,175,87,0.4971
+26B,QAT,smoke,ambiguous_time_frame,5,4,0.8
+26B,QAT,smoke,date_boundary,10,8,0.8
+26B,QAT,smoke,query_neutrality,10,10,1.0
+26B,QAT,smoke,stable_control,25,25,1.0
+26B,QAT,smoke,stale_premise_trap,15,14,0.9333
+26B,QAT,smoke,volatile_current,35,25,0.7143
+26B,QAT,core500,ambiguous_time_frame,25,15,0.6
+26B,QAT,core500,date_boundary,50,29,0.58
+26B,QAT,core500,query_neutrality,50,46,0.92
+26B,QAT,core500,stable_control,125,123,0.984
+26B,QAT,core500,stale_premise_trap,75,72,0.96
+26B,QAT,core500,volatile_current,175,93,0.5314

data/paired_quant_effect.csv

@@ -0,0 +1,5 @@
+suite,size,n_pairs,qat_only_correct,ptq_only_correct,delta_rate,mcnemar_p
+core500,12B,500,1,7,-0.012,0.0703
+core500,26B,500,15,6,0.018,0.0784
+smoke,12B,100,0,2,-0.02,0.5
+smoke,26B,100,5,2,0.03,0.4531

data/regime_output_divergence.csv

@@ -0,0 +1,5 @@
+suite,size,n,exact_match,rate,median_first_divergence_char
+core500,12B,500,148,0.296,73
+core500,26B,500,116,0.232,69
+smoke,12B,100,32,0.32,46
+smoke,26B,100,27,0.27,57

paper/PAPER.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6accb2c3fedc9e0e27d116fde893d25d55aea26c0ae88354173d99ca355d4d6e
+size 165173