Release qwen38-27b-12gb v1.0.1 [plan b49b55647aa0b786dfe13cdcd6380e951cb5dab88140ac01327bf78302a44030] (#3)

- Release qwen38-27b-12gb v1.0.1 [plan b49b55647aa0b786dfe13cdcd6380e951cb5dab88140ac01327bf78302a44030] (1c4335b692d56e7a1602d69385338f13f788f233)

CITATION.cff

@@ -6,12 +6,12 @@ authors:
   - family-names: Schwartz
     given-names: Matthew
     orcid: "https://orcid.org/0009-0009-4171-7247"
-version: 1.0.0
-doi: "10.5281/zenodo.22166977"
-date-released: 2026-08-29
-repository-code: "https://github.com/matthematics1137/research-artifacts/tree/qwen38-27b-12gb-v1.0.0/papers/2026-qwen38-27b-12gb"
-repository-artifact: "https://doi.org/10.5281/zenodo.22166977"
-url: "https://doi.org/10.5281/zenodo.22166977"
+version: 1.0.1
+doi: "10.5281/zenodo.22314395"
+repository-artifact: "https://doi.org/10.5281/zenodo.22314395"
+date-released: 2026-09-04
+repository-code: "https://github.com/matthematics1137/research-artifacts/tree/qwen38-27b-12gb-v1.0.1/papers/2026-qwen38-27b-12gb"
+url: "https://github.com/matthematics1137/research-artifacts/tree/qwen38-27b-12gb-v1.0.1/papers/2026-qwen38-27b-12gb"
 keywords:
   - large language model inference
   - post-training quantization
@@ -27,6 +27,6 @@ preferred-citation:
       given-names: Matthew
       orcid: "https://orcid.org/0009-0009-4171-7247"
   year: 2026
-  url: "https://github.com/matthematics1137/research-artifacts/blob/qwen38-27b-12gb-v1.0.0/papers/2026-qwen38-27b-12gb/paper/main.pdf"
+  url: "https://github.com/matthematics1137/research-artifacts/blob/qwen38-27b-12gb-v1.0.1/papers/2026-qwen38-27b-12gb/paper/main.pdf"
   institution:
     name: Independent researcher

CLAIM_EVIDENCE_MAP.md

@@ -19,3 +19,11 @@ table. The release inventory and `SHA256SUMS` freeze every other source row.
 
 The release contains response excerpts, not full transcripts. It contains no
 Q4 per-item GSM8K/HumanEval rows because those files were not retained.
+# Correction 1.0.1
+
+`CORRECTIONS.md` records the versioned reporting changes. `correction_metrics.py`
+checks the recovered MTP log against all 45 retained completion-token counts,
+its four distinct MATH rate summaries, Q4 error IDs and elapsed retry time,
+the exact historical GSM selection, and the server/tg128 caption bound.
+`test_corrections.py` includes mutation tests for these checks and stricter
+MATH ID/boolean validation. Original headline results remain unchanged.

CORRECTIONS.md

@@ -0,0 +1,52 @@
+# Version 1.0.1 correction — 2026-09-04
+
+This version corrects reporting in the original [1.0.0 artifact](https://doi.org/10.5281/zenodo.22166977).
+No inference was rerun. Original measurement and result rows, task scores,
+statistical tests, and the central deployment conclusions are unchanged.
+The published 1.0.0 files remain intact.
+
+1. **Request errors and retries:** the MATH table now separates two Q4 request
+   time-outs from output-length truncations. Both errors were already scored
+   wrong. Their 2404.36 seconds remain in the 7313.90-second goodput denominator
+   (0.1887 correct/min). The client could automatically retry once; one retained
+   response does not mean one attempted stochastic request. Attempt counts
+   were not recorded and are not reconstructed as new observations.
+2. **GSM8K selection:** the evaluated 50 items were the first half of a sorted
+   seed-42 sample of 100, not a direct random sample of 50. Their source indices
+   span 13–563. This is disclosed without changing the sample or its scores.
+3. **MTP timing evidence:** the contemporaneous server log, already present in
+   the original evidence-freeze commit, is now included with model paths
+   sanitized. Its 45 request completion-token counts match the MATH-25 and
+   HumanEval+-20 rows in sequence. That retrospective alignment is not the
+   same as run-time benchmark-ID instrumentation. The MATH per-request decode
+   range is 41.09–50.84 tok/s; aggregate decode is 45.35 tok/s. Aggregate logged
+   HTTP-response rate is 44.96 tok/s, a distinct metric. The narrower old
+   45–50 prose range and 48–50 figure annotation are replaced. There is still
+   no identical-prompt draft-off comparison establishing a causal MTP effect.
+4. **Server/tg128 caption:** the matching draft-off server measurements differ
+   from tg128 by at most 0.73 tok/s, not 0.4.
+5. **Effective-bpw figure labels:** labels now derive from checkpoint bytes and
+   parameter count; Q4 is 5.22 and bartowski IQ2_S is 3.06 effective bpw.
+6. **Needle context sizes:** 8K/16K/32K are nominal estimates based on word count,
+   not actual model-tokenizer measurements.
+7. **Offload fit:** the two-bandwidth fit is explicitly descriptive, assessed
+   on its eight fitting configurations. It does not isolate CPU dequantization
+   costs or provide held-out predictive validation.
+
+The added `correction_metrics.py` checks log alignment, MTP rates, error IDs,
+retry-time accounting, and GSM selection. `check_claims.py` also rejects missing,
+duplicate, unexpected, or incorrectly typed MATH result IDs/booleans and prints
+the paired McNemar result as the primary MATH comparison. The public TeX no
+longer has a dormant dependency on a private dark-mode style. Hugging Face
+retains explicit viewer configurations for its summary and throughput tables.
+
+## What did not change
+
+The 1.64× matched-footprint MATH generation-rate comparison, fully resident
+versus partially offloaded deployment conclusion, paired MATH result, and MMLU
+non-resolution all reproduce. Quantizer and engine remain confounded; the
+study is still a one-machine, small-sample deployment characterization.
+
+Run `python3 check_claims.py` from this artifact root. Additional evidence
+provenance is in `environment/correction-provenance.json`; complete release
+hashes are supplied by the generated artifact inventory.

README.md

@@ -27,16 +27,17 @@ This dataset mirrors the redistribution-cleared measurements for “Deploying
 Qwen3.8-27B in 12 GB of VRAM: Accuracy and Throughput Across Quantized
 Inference Stacks.”
 
-The immutable version-of-record artifact is
-[doi:10.5281/zenodo.22166977](https://doi.org/10.5281/zenodo.22166977). The
-matching tagged GitHub source and analysis instructions are at
-<https://github.com/matthematics1137/research-artifacts/tree/qwen38-27b-12gb-v1.0.0/papers/2026-qwen38-27b-12gb>.
-The matching GitHub release is
-<https://github.com/matthematics1137/research-artifacts/releases/tag/qwen38-27b-12gb-v1.0.0>.
-This discovery dataset is
-<https://huggingface.co/datasets/mv1137/qwen38-27b-12gb-results>. The DOI is
-assigned to the reviewed Zenodo deposition. These links identify version 1.0.0,
-dated 2026-08-29.
+This is the 1.0.1 correction, dated 2026-09-04. Original measurement
+rows are unchanged; see [CORRECTIONS.md](CORRECTIONS.md). The previous 1.0.0
+artifact remains at [doi:10.5281/zenodo.22166977](https://doi.org/10.5281/zenodo.22166977).
+The correction's version DOI is
+[doi:10.5281/zenodo.22314395](https://doi.org/10.5281/zenodo.22314395).
+Version-specific source and analysis instructions are at
+<https://github.com/matthematics1137/research-artifacts/tree/qwen38-27b-12gb-v1.0.1/papers/2026-qwen38-27b-12gb>.
+The result dataset is
+<https://huggingface.co/datasets/mv1137/qwen38-27b-12gb-results>.
+In prepublication review copies, the reserved DOI and planned tag need not
+resolve until release; this card alone is not a publication receipt.
 
 This repository contains measurement records under mixed terms rather than a
 single blanket license. Original measurements and documentation are CC BY 4.0;

environment/correction-provenance.json

@@ -0,0 +1,16 @@
+{
+  "version": "1.0.1",
+  "previous_version": "1.0.0",
+  "previous_version_doi": "10.5281/zenodo.22166977",
+  "evidence_freeze_commit": "009859e4af3a923732be09ee60b52383827db0f7",
+  "previous_paper_pdf_sha256": "3d54edc454e9f1c6166ea1d61ceea59d9eb6bb1838df22833b912eb349a14bc6",
+  "new_inference": false,
+  "original_measurement_rows_changed": false,
+  "additional_existing_evidence": {
+    "path": "testsuite/results/server_UD-IQ2_XXS-mtpgpu-effmed.log",
+    "source_git_blob": "645c26c623cb993baa194526257fb6346a29ffd6",
+    "source_sha256": "42b40f1e8e2cc744b36e5423fcbecd459a15cfe2dcc0498755f3292b73fbeada",
+    "sanitization": "Replace the absolute model-directory prefix with <MODEL_ROOT>/; all timing lines are unchanged.",
+    "alignment": "The 45 final request timings align in order with all completion-token counts in MATH-25 followed by HumanEval+-20. Requests do not embed benchmark item IDs; alignment is retrospective, not run-time identity instrumentation."
+  }
+}