Release dual PRE/POST pyrrho-v2-nano-g1

README.md

@@ -23,8 +23,9 @@ metrics:
 
 # pyrrho-v2-nano-g1
 
-`pyrrho-v2-nano-g1` is a small local RAG governance co-processor. It reads a user
-question plus retrieved source passages, then returns whether the evidence is
+`pyrrho-v2-nano-g1` is a small local RAG planning and governance co-processor. It can
+run before retrieval on a user query and after retrieval on the query plus
+source passages. The post-retrieval pass returns whether the evidence is
 `SUFFICIENT`, `DISPUTED`, or `INSUFFICIENT` before an answer is generated.
 
 It is not an answer generator, not a retriever, and not an open-world fact
@@ -34,8 +35,8 @@ show a dispute, retry retrieval, or ask for missing evidence.
 
 Compared with the older Pyrrho v1 line, v2 exposes a smaller native head shape:
 one evidence verdict, one failure reason, and two compact multi-label metadata
-heads. The goal is a cleaner governance contract for `fitz-sage` and other RAG
-systems.
+heads. The same model owns the `fitz-sage` pre-retrieval query-planning pass and
+the post-retrieval evidence-governance pass.
 
 ## Native V2 Heads
 
@@ -43,7 +44,7 @@ systems.
 |---|---|---|
 | `evidence_verdict` | `SUFFICIENT`, `DISPUTED`, `INSUFFICIENT` | Post-retrieval evidence sufficiency and conflict decision. |
 | `failure_mode` | `none`, `unresolved_conflict`, `missing_or_incomplete_evidence`, `wrong_scope_or_version`, `ambiguous_request` | Actionable reason when evidence is disputed or insufficient. |
-| `retrieval_intents` | `needs_lookup`, `needs_temporal_resolution`, `needs_comparison_or_set`, `needs_broad_coverage` | Query/evidence task metadata for retry and retrieval policy. |
+| `retrieval_intents` | `needs_lookup`, `needs_temporal_resolution`, `needs_comparison_or_set`, `needs_broad_coverage` | Pre-retrieval planning and post-retrieval retry metadata. |
 | `evidence_kinds` | `needs_text`, `needs_table_or_record`, `needs_code_or_symbol`, `needs_config_or_setting`, `needs_log_or_run_result`, `needs_document_layout` | Evidence-surface metadata for routing, audit, and missing-source hints. |
 
 ## Output Contract
@@ -58,8 +59,9 @@ softmax. Decode it by head:
 | `8:12` | `retrieval_intents` | sigmoid multi-label scores |
 | `12:18` | `evidence_kinds` | sigmoid multi-label scores |
 
-Most integrations should expose a structured decision object derived from those
-logits:
+Most integrations should expose structured objects derived from those logits.
+For `[PYRRHO_PRE]`, use only `retrieval_intents` and `evidence_kinds`. For
+`[PYRRHO_POST]`, use all four heads.
 
 | Field | Meaning |
 |---|---|
@@ -105,8 +107,8 @@ Example normalized output:
 ```
 
 The model does not generate answers, citations, source spans, retrieval results,
-or natural-language explanations. It classifies and scores the `(query,
-retrieved_contexts)` evidence state.
+or natural-language explanations. It classifies and scores query intent before
+retrieval and the `(query, retrieved_contexts)` evidence state after retrieval.
 
 ## Intended Use
 
@@ -124,7 +126,17 @@ a retriever, write answers, or replace human review in high-stakes settings.
 
 ## Input Format
 
+Pre-retrieval query planning:
+
+```text
+[PYRRHO_PRE]
+Question: <user query>
+```
+
+Post-retrieval evidence governance:
+
 ```text
+[PYRRHO_POST]
 Question: <user query>
 
 Sources:
@@ -168,7 +180,7 @@ contexts = [
     "The company posted net income of $4 million in Q2.",
     "The company recorded a quarterly loss of $12 million in Q3.",
 ]
-text = "Question: " + query + "\n\nSources:\n" + "\n".join(
+text = "[PYRRHO_POST]\nQuestion: " + query + "\n\nSources:\n" + "\n".join(
     f"[{i}] {context}" for i, context in enumerate(contexts, start=1)
 )
 
@@ -208,31 +220,45 @@ The repository includes both FP32 and INT8 ONNX exports:
 - `model.onnx`
 - `model_quantized.onnx`
 
-For CPU inference, load `model_quantized.onnx` through `onnxruntime` or an
-Optimum ONNX runtime wrapper. Decode the resulting 18 logits using the same
-slices shown above.
+`fitz-sage` loads `model.onnx` by default for governance accuracy. The quantized
+graph is included for integrations that explicitly trade some accuracy margin
+for smaller artifacts. Decode the resulting 18 logits using the same slices
+shown above.
 
 ## Evaluation
 
-Held-out training eval from `outputs/modernbert_base_v2_alpha_41358_active_20260704_seed42`:
+Held-out post-retrieval eval from
+`outputs/modernbert_base_v2_dual_from_g1_41358_active_20260704_seed42`:
 
 | Metric | Value |
 |---|---:|
-| overall score | 0.9497 |
-| verdict accuracy | 0.9727 |
-| false sufficient rate | 0.0455 |
-| failure accuracy | 0.9601 |
-| retrieval exact match | 0.8335 |
-| retrieval macro F1 | 0.9300 |
+| overall score | 0.9471 |
+| verdict accuracy | 0.9703 |
+| false sufficient rate | 0.0484 |
+| failure accuracy | 0.9567 |
+| retrieval exact match | 0.8308 |
+| retrieval macro F1 | 0.9277 |
 | evidence-kind exact match | 0.9809 |
 | evidence-kind macro F1 | 0.9950 |
 
+Held-out pre-retrieval query eval:
+
+| Metric | Value |
+|---|---:|
+| retrieval exact match | 0.8248 |
+| retrieval macro F1 | 0.9266 |
+| evidence-kind exact match | 0.9637 |
+| evidence-kind macro F1 | 0.9873 |
+
 Fitz-sage release-candidate checks:
 
 | Benchmark | Result |
 |---|---:|
 | balanced fixed-evidence governance sanity suite | 120/120 |
-| live fitz-sage benchmark | 86/120 |
+| live fitz-sage benchmark | 97/120 |
+| core | 19/20 |
+| holdout | 43/50 |
+| holdout2 | 35/50 |
 
 The live benchmark result is the practical integration target. The fixed-evidence
 suite is a minimal sanity check for the governance head.

manifest.json

@@ -1,18 +1,34 @@
 {
-  "schema_version": "pyrrho-v2-release-manifest-1.0",
+  "schema_version": "pyrrho-v2-release-manifest-1.1",
   "release": {
     "model_id": "yafitzdev/pyrrho-v2-nano-g1",
     "release_name": "pyrrho-v2-nano-g1",
-    "created_date": "2026-07-06",
+    "created_date": "2026-07-08",
     "base_model": "answerdotai/ModernBERT-base",
     "architecture": "ModernBertForSequenceClassification",
     "num_logits": 18,
     "max_length": 2048,
+    "input_mode": "dual",
     "trustworthy_threshold": null
   },
+  "input_contract": {
+    "pre_tag": "[PYRRHO_PRE]",
+    "post_tag": "[PYRRHO_POST]",
+    "pre_trains_heads": [
+      "retrieval_intents",
+      "evidence_kinds"
+    ],
+    "post_trains_heads": [
+      "evidence_verdict",
+      "failure_mode",
+      "retrieval_intents",
+      "evidence_kinds"
+    ]
+  },
   "source_checkpoint": {
-    "path": "outputs/modernbert_base_v2_alpha_41358_active_20260704_seed42/best_model",
-    "run_id": "modernbert_base_v2_alpha_41358_active_20260704_seed42",
+    "path": "outputs/modernbert_base_v2_dual_from_g1_41358_active_20260704_seed42/best_model",
+    "run_id": "modernbert_base_v2_dual_from_g1_41358_active_20260704_seed42",
+    "continued_from": "models/pyrrho-v2-nano-g1",
     "seed": 42,
     "training_rows": 41358,
     "dataset_pointer": "fitz_gov_v2_41358_20260703"
@@ -51,21 +67,37 @@
     ]
   },
   "eval": {
-    "eval_overall_score": 0.949743230482894,
-    "eval_verdict_accuracy": 0.9726855209088712,
-    "eval_false_sufficient_rate": 0.04551920341394026,
-    "eval_failure_accuracy": 0.9601160261058739,
-    "eval_retrieval_exact_match": 0.8334541938602852,
-    "eval_retrieval_macro_f1": 0.9299832758138473,
-    "eval_evidence_kind_exact_match": 0.9809040367416002,
-    "eval_evidence_kind_macro_f1": 0.9949836611151947
+    "post": {
+      "eval_overall_score": 0.9470664187704583,
+      "eval_verdict_accuracy": 0.9702683103698332,
+      "eval_verdict_macro_f1": 0.9591237257007766,
+      "eval_false_sufficient_rate": 0.04836415362731152,
+      "eval_failure_accuracy": 0.9567319313512207,
+      "eval_failure_macro_f1": 0.9064147136375201,
+      "eval_retrieval_exact_match": 0.8307952622673435,
+      "eval_retrieval_macro_f1": 0.9277362299701075,
+      "eval_evidence_kind_exact_match": 0.9809040367416002,
+      "eval_evidence_kind_macro_f1": 0.9949910057734287
+    },
+    "pre": {
+      "pre_eval_retrieval_exact_match": 0.8247522359197486,
+      "pre_eval_retrieval_macro_f1": 0.9266146369602828,
+      "pre_eval_evidence_kind_exact_match": 0.9637418419144308,
+      "pre_eval_evidence_kind_macro_f1": 0.9873183319207018
+    }
   },
   "fitz_sage_benchmark": {
-    "balanced_fixed_evidence_toy_cases": "120/120",
-    "live_total": "86/120",
-    "live_core": "15/20",
-    "live_holdout": "38/50",
-    "live_holdout2": "33/50"
+    "balanced_fixed_evidence_cases": "120/120",
+    "live_total": "97/120",
+    "live_core": "19/20",
+    "live_holdout": "43/50",
+    "live_holdout2": "35/50",
+    "results": {
+      "balanced": "benchmarks/results/dual_candidate_governance_balanced_v2.json",
+      "core": "benchmarks/results/dual_candidate_core_v2.json",
+      "holdout": "benchmarks/results/dual_candidate_holdout_v2.json",
+      "holdout2": "benchmarks/results/dual_candidate_holdout2_v2.json"
+    }
   },
   "artifacts": {
     "transformers": [

model.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f376b9ccead222fc056ae664fe98279bd783f0902f5965747e52c4f2603b7417
-size 599049587
+oid sha256:63982dc593fd03792177df26d7f15c9db5c41062da2f88325e63935638ab1a20
+size 599049588

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a214c3d960dd2e8438e9df9ed7073cbcd2cf04c89d3e1b0ddcbba1c0f920d812
+oid sha256:fd9b63199e997f94f2e8d334030b8412b9456434dda4836c9858a2cea6e0c3a6
 size 598489008

model_quantized.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9c2db516e0babcefed860082ca3c2d94b2467d5554356dd24d9dbd00cb155ce6
+oid sha256:4041907531982651cc1b613101427ca82c849a58e67252343237e870cb46becc
 size 150664051