Publish gliclass-std-base-v3-daecore-5facet-qint8-v2 (qualified ONNX export and complete attribution)

MODIFICATIONS.md

@@ -0,0 +1,26 @@
+# Modifications from the upstream model
+
+`gliclass-std-base-v3-daecore-5facet-qint8-v2` is not an unmodified copy of
+`knowledgator/gliclass-base-v3.0`. It is not endorsed by Knowledgator or by
+Microsoft, whose `deberta-v3-base` backbone the upstream model builds on.
+
+What changed:
+
+- **Fine-tune.** The upstream GLiClass single-pass classifier was trained
+  further on 59,886 labeled passages for five fixed facets (`trap`,
+  `decision`, `constraint`, `mechanism`, `procedure`) with the five label
+  prompts recorded in `classifier-metadata.json`. Labels were frontier-model
+  judgments under a frozen protocol, not human annotations.
+- **Export.** The fine-tuned weights were exported to ONNX (opset 17) as a
+  single graph taking `input_ids` and `attention_mask` and emitting five
+  logits.
+- **Quantization.** Only the token-embedding `Gather` tables were quantized
+  to signed INT8 (per-channel off, reduce-range off); every matrix product
+  stays FP32. Twenty-four constant identity nodes were folded. The compression
+  was qualified as equivalent to the FP32 export on a 5,298-row panel.
+- **Calibration.** Per-facet temperature scaling and two frozen threshold
+  tables (`recall_leaning`, `contract`) travel in `classifier-metadata.json`
+  and are part of the artifact's identity.
+
+Unchanged: the tokenizer vocabulary and the `<<LABEL>>` / `<<SEP>>` prompt
+convention of the upstream model.

README.md

@@ -19,8 +19,8 @@ This model tags passages of working notes and documentation with five
 independent facets: `trap`, `decision`, `constraint`, `mechanism`, and
 `procedure`. It is a supervised fine-tune of
 [`knowledgator/gliclass-base-v3.0`](https://huggingface.co/knowledgator/gliclass-base-v3.0)
-trained on 59,886 labeled passages and shipped as a 452.8 MB ONNX bundle
-qualified on CPU, CUDA, and DirectML.
+trained on 59,886 labeled passages and shipped as a 452.8 MB ONNX graph inside a
+461.5 MB bundle, qualified on CPU, CUDA, and DirectML.
 
 On a frozen 1,300-row held-out panel drawn from three unseen document families,
 it reached macro average precision 0.9676 against 0.8526 for the previous
@@ -46,11 +46,16 @@ the model behaves on unrelated real documents has not been measured.
 | Serialized artifact | `model.onnx`, 452,812,018 bytes, opset 17 |
 | Quantization | signed INT8 on the token-embedding matrix only; the transformer body stays FP32 |
 | Threshold tables | two frozen operating views, `contract` and `recall_leaning` |
-| Deployment state | promotion-qualified; activation as the shipped default is a separate step |
+| Deployment state | the installed default since 2026-09-11; published 2026-09-06 |
 
 The artifact size follows from the quantization choice: the 128k-entry
 embedding table, about 98M parameters, is stored at one byte per weight while
-the 88M-parameter body stays at four. Full-graph INT8 was not qualified.
+the 88M-parameter body stays at four. Broader quantization was measured, not
+skipped: on the predecessor fit of the same recipe, six arms were scored
+against the FP32 reference on a 5,298-row panel,
+and every one beyond the embedding table cost quality — embedding plus attention
+−0.0013, full per-channel reduced-range −0.0531, and full per-channel −0.0599
+macro average precision.
 
 ## Intended use
 
@@ -77,9 +82,11 @@ both training and evaluation.
 
 **Label provenance.** Every label was produced by a language model applying the
 same one-sentence facet definitions the classifier sees. Training rows received
-one primary judgment from GPT-5.6 (Sol, medium reasoning), with a frozen
-400-row cross-model audit by GPT-5.6 (Terra, high) and a Sol-high tiebreak only
-where the two disagreed semantically. Calibration and promotion panels used two
+two independent judgments each with a higher-effort tiebreak on semantic
+disagreement, except for 3,750 inherited rows that received one primary judgment
+from GPT-5.6 (Sol, medium reasoning) with a frozen 400-row cross-model audit by
+GPT-5.6 (Terra, high) and a Sol-high tiebreak where the two disagreed.
+Calibration and promotion panels used two
 independent Sol-high sessions per row plus a third Sol-xhigh session that saw
 only the disputed fields. Unresolved fields stay unresolved and are masked per
 facet; nothing unresolved becomes a negative.
@@ -96,20 +103,36 @@ agreement with that instrument, not with human ground truth.
 
 | Lane | Rows | Origin |
 |---|---:|---|
-| Final training set | 59,886 | 54,803 generated, 5,083 from the operator's own project documentation |
+| Final training set | 59,886 | 54,803 generated; 4,583 real rows from the operator's self-owned corpora and 500 real rows of public-domain US Federal Register text |
+
+The 59,886 passages come from 9,376 documents, about 6.4 passages each: 8,212
+generated documents spanning 143 fictional organizations, and 1,164 real
+documents — 1,042 from the operator's two workspaces and 122 from the Federal
+Register set.
 
 Generated documents come from fictional organizations written as complete
 documents (procedures, incident reports, decision records, design notes) and
 then passed through the production parser. They were preferred over the
-operator's real corpus for both training and evaluation because they cover many
-more domains, document shapes, and organizational settings; the real corpus is
-one person's projects and too narrow to generalize from.
+real lane for both training and evaluation because they cover many more domains,
+document shapes, and organizational settings. The real lane is narrower than its
+family count suggests: 315 of its 437 source families are directory groupings inside
+two workspaces belonging to one person, and the other 122 are single Federal
+Register dockets, not 437 organizational settings; the Federal Register rows are
+also regulatory prose from a single genre. Whole-family
+splitting is therefore a strong isolation guarantee in the generated lane, where
+a family is a whole organization, and a weak one in the real lane.
+
+Every real row is redistributable — the operator's corpora by ruling,
+the Federal Register text as public-domain US government work. The Federal
+Register rows are training supervision only: no panel that calibrates, gates, or
+reports on this model contains them, or any other real row.
 
 Splits are by whole source family, never by row. The seven families reserved
 for calibration and promotion were removed from training along with 3,875
 predecessor training rows that shared them. Calibration and promotion families
-are disjoint from each other and from training by family, document lineage,
-exact passage identity, and lexical near-duplicate screening.
+are disjoint from each other and from training by family, document lineage, and
+exact passage identity. Assignment was label-blind and checked for support only
+afterward.
 
 The final fit used weighted binary cross-entropy with a 2.0 multiplier on
 mentions-only hard negatives, learning rate 2e-5, batch 4 with 8 accumulation
@@ -168,8 +191,7 @@ is the classifier this model replaces.
 | procedure | 0.375 | 0.9578 | 0.8167 | 0.886 | 0.528 | 0.822 | 0.489 |
 | macro | 0.631 | 0.9676 | 0.8526 | 0.902 | 0.719 | 0.860 | 0.687 |
 
-The macro gain is 0.1150. The same candidate gained 0.1200 on the separate
-calibration panel, so the two panels agree within 0.005. The one-sided 95%
+The macro gain is 0.1150. The one-sided 95%
 lower bound on the macro precision gain under the frozen contract thresholds is
 0.009; the decision-facet precision gain has a negative lower bound, because the
 previous classifier's decision threshold was so strict that it predicted only 26
@@ -278,8 +300,9 @@ Paired macro average-precision differences from this model, with family-grouped
 [−0.0548, −0.0366]; GPT-5.5 low −0.0588 [−0.0666, −0.0442]; GPT-5.4 low
 −0.0937 [−0.1015, −0.0827]. Every interval excludes parity. Raising reasoning
 effort improved GPT-5.4 by 0.0509 [0.0456, 0.0589] and GPT-5.5 by 0.0080
-[0.0014, 0.0204]; it narrows the gap without closing it. Two unplanned repeats
-of the Spark run spanned 0.8086 to 0.8496, a spread of 0.041 that nearly
+[0.0014, 0.0204]; it narrows the gap without closing it. Three accepted runs
+of the Spark configuration spanned 0.8086 to 0.8496 — the reported 0.8086 plus
+repeats at 0.8180 and 0.8496 — a spread of 0.041 that nearly
 matches the smallest GPT gap and exceeds every one of the four interval widths,
 which is the reason single runs are flagged above.
 
@@ -313,10 +336,13 @@ precision unless stated.
 3. **Context window.** Widening the input from 416 to 768 tokens added +0.0048
    [+0.0028, +0.0068] on development and +0.0033 [+0.0011, +0.0065] on a
    separate held-out set. A counterfactual-projector branch on minimal pairs
-   changed nothing (+0.0001) and was dropped.
-4. **Backbone size.** With the full data and a matched recipe on a 5,298-row
-   eight-family panel, the larger backbone beat the base by 0.0042 average
-   precision and 0.0152 at 90% recall [+0.0015, +0.0451], but cost 2.6 times
+   changed nothing against its own replay control (+0.0001) and was dropped; the
+   run's window also did not match its draft plan, so it could not have
+   authorized a recipe change either way.
+4. **Backbone size.** On the ModernBERT GLiClass pair, with 44,049 training
+   rows and a matched recipe on a 5,298-row eight-family panel, the large arm beat
+   the base by 0.0042 average precision and 0.0152 at 90% recall [+0.0015,
+   +0.0451] — at 95% recall the interval spans zero — but cost 2.6 times
    the artifact bytes, 2.4 times the training time, and 1.7 times peak GPU
    memory, while the base scored 1.8 times as many rows per second. The base was
    selected as the product architecture.
@@ -328,10 +354,18 @@ precision unless stated.
 6. **Final fit.** The frozen recipe was trained once on all 59,886 rows,
    calibrated, compressed, and read once on the promotion panel.
 
-What did not help, in one place: more real documents without generated
-diversity, the data curriculum, the robust loss, the counterfactual projector,
-and the larger backbone at its cost. Better and more diverse examples, boundary
-cases, and a wider context did the work.
+What did not help: more real documents without generated diversity, the data
+curriculum, the robust loss, the counterfactual projector, and the larger
+backbone at its cost. A distribution-balanced loss, R-Drop, SMART smoothness, a
+three-state auxiliary head, and a 512-token window were also measured and
+rejected, and the experiment log carries each with its numbers. Two further
+branches closed without a scored comparison: a frozen large-NLI recipe, stopped
+after 23.2 hours with no completed epoch, and a decision-only rationale
+auxiliary, which trained four epochs and reduced all three training losses but
+failed artifact assembly on incomplete checkpoint-average candidates, so no
+panel was mounted and no per-row scores were emitted. It is retired as
+inconclusive. Better and more diverse examples, boundary cases, and a wider
+context did the work.
 
 ## Runtime
 
@@ -387,7 +421,7 @@ CUDA and 277 MiB on DirectML above a resident session of 657 MiB and 481 MiB res
 | Previous-classifier identity | verified 2026-09-04 against the staged bundle's own config |
 | Frontier-model comparisons | as recorded; single runs on a subsample that is not independent of the promotion read |
 | Serving throughput | 3.74 passages per second, as recorded from the scoring run bound to the compression score receipt |
-| Per-passage latency percentiles | not measured |
+| Per-passage latency percentiles | single-passage p50 per provider only; no p95 or p99 |
 
 ## Limitations, ranked
 
@@ -416,8 +450,16 @@ CUDA and 277 MiB on DirectML above a resident session of 657 MiB and 481 MiB res
 8. **Serving cost is measured on one machine.** The provider table in the runtime
    section comes from a single 8 GB NVIDIA card and a four-core CPU budget; other
    hosts will land elsewhere, and cold-start cost is not separated from warm batches.
-9. **Not yet the shipped default.** The artifact is promotion-qualified; the
-   activation, restamp, and rollback steps are separate operations.
+9. **Second sealed campaign in its family.** An earlier fit was read on a
+   sealed panel, returned NO-GO on a per-origin recall gate, and the contract was
+   then amended so origin slices are diagnostics rather than independent vetoes.
+   That fit's calibration and promotion rows were folded into this model's own
+   training set, and this campaign therefore grants its panel promotion authority
+   while explicitly declining a claim to a pristine sealed alpha-spending draw.
+   The earlier panel carried 333 operator-workspace rows; this one is entirely
+   generated, so the per-origin recall that failed on the earlier read cannot be
+   measured on this model at all. Resolving evidence would be a promotion panel
+   drawn from unspent reserve that includes real rows.
 
 ## Reproducibility
 

publication-manifest.json

@@ -1,12 +1,14 @@
 {
   "schema": "daecore.classifier-publication-manifest",
-  "staged_at": "2026-09-06T22:38:02+00:00",
+  "tool_sha256": "ec0f21b2c945a07b0a1f7be9895c91f30a985fb981a5fb83f409604b33879af4",
   "public_repo": "Daecore/gliclass-std-base-v3-5facet-qint8-v2",
   "classifier_version": "gliclass-std-base-v3-daecore-5facet-qint8-v2",
   "upstream": {
     "model_id": "knowledgator/gliclass-base-v3.0",
     "revision": "77a70e6cd52e602ed18184ef37d18bdd3741e3d5"
   },
+  "nominee_receipt_sha256": "d463fc167b2a26f8f5b56e03c18e4c604b231eb25b5093064dd8a9bd7ffa6373",
+  "staged_at": "2026-09-18T18:58:21+00:00",
   "files": {
     "model.onnx": {
       "sha256": "690e50920e7780db5ecd14e4b209f0d3c86199214063e8bb59d399c93861324c",
@@ -35,8 +37,8 @@
       "binding": "nominee receipt runtime_metadata_sha256 (canonical JSON hash)"
     },
     "README.md": {
-      "sha256": "e7dc7ef045a4daa57f85e2213da0fe7273322737417bd16147afe8adfc73b01b",
-      "size": 26973,
+      "sha256": "48cc3f3e7659463c7bd6031d923a96f3822e7ebd62a2d5f660d779d95612f76c",
+      "size": 29930,
       "binding": "packaging record"
     },
     "LICENSE": {
@@ -48,6 +50,11 @@
       "sha256": "71c772c9388cdbdf68ca17f9e2f7f85f1ed7c210a326885a5bd64ed06bb44b3c",
       "size": 822,
       "binding": "packaging record"
+    },
+    "MODIFICATIONS.md": {
+      "sha256": "fca339612182863205137d75b60cb23dc921090c0e91a4541a382c3889cd7e95",
+      "size": 1399,
+      "binding": "packaging record"
     }
   }
 }