pyrrho-nano-g4

pyrrho-nano-g4 is a small multitask RAG governance co-processor for anti-hallucination and retrieval-quality pipelines. It reads a user question plus retrieved source passages, then returns a calibrated evidence-state decision and auxiliary signals that fitz-sage can use before answer generation.

It is not an answer generator and not an open-world fact checker. It sits between retrieval and generation, or beside a retrieval package as a fast evidence quality layer. Compared with pyrrho-nano-g3.2, this package keeps the retrieval-control head surface but collapses answerability into the four V9 planning labels: direct_answer, synthesis_answer, set_answer, and structured_reasoning.

Governance Labels

Label	Meaning
ABSTAIN	The retrieved sources do not contain enough evidence to answer the question.
DISPUTED	The retrieved sources conflict on the answer.
TRUSTWORTHY	The retrieved sources consistently support answering the question.

Multitask Heads

Head	Labels / values	Intended use
governance	ABSTAIN, DISPUTED, TRUSTWORTHY	Post-retrieval evidence sufficiency and conflict decision.
query_contract	evidence_sufficiency, structured_lookup, temporal_grounding, exhaustive_coverage, comparison_coverage, representative_overview	Pre-retrieval routing signal for what kind of evidence the query needs.
route	science_medicine, law_policy, history_geography, technology_computing, economics_finance, culture_society, general_commonsense	Semantic route/domain signal for retrieval policy and logging.
taxonomy	23 fitz-gov taxonomy patterns	Failure/support pattern signal for audit and diagnostics.
scalars	evidence_sufficiency, query_evidence_alignment, answer_coverage, conflict_density, retrieval_retry_value, false_trustworthy_risk, evidence_failure_severity	Continuous governance signals for retry, ranking, and monitoring.
retrieval_action	answer_now, retrieve_more, broaden_search, resolve_conflict, ask_clarifying_question, structured_lookup	Retrieval policy hint for the next pipeline action.
gap_type	12 evidence-gap labels	More specific reason why retrieval is insufficient or conflicting.
answerability_shape	direct_answer, synthesis_answer, set_answer, structured_reasoning	Query-only collapsed answer shape for retrieval planning.
retrieval_modality	unstructured_text, structured_table, code, configuration, log_trace, pdf_layout, mixed	Query-only hint for the preferred retrieval substrate.

Outputs

This is a custom multitask package, not a standard single-head AutoModelForSequenceClassification artifact. The recommended runtime is pyrrho.multitask_inference.PyrrhoMultiTaskPredictor from the pyrrho repository.

The predictor returns a structured object:

Field	Meaning
governance.final_label	Final calibrated label after the TRUSTWORTHY threshold rule.
governance.raw_label	Highest-probability governance label before threshold calibration.
governance.probabilities	Probability distribution over ABSTAIN, DISPUTED, TRUSTWORTHY.
governance.threshold	TRUSTWORTHY probability threshold used by the package.
query_contract.final_label	Query-only contract prediction.
route.final_label	Query-only semantic route/domain prediction.
taxonomy.final_label	Query+evidence taxonomy-pattern prediction.
scalars	7 bounded scalar governance signals.
retrieval_action.final_label	Retrieval policy hint.
gap_type.final_label	Evidence-gap type prediction.
answerability_shape.final_label	Query-only answer-shape prediction.
retrieval_modality.final_label	Query-only retrieval-modality prediction.
timing_ms	Local inference timing for the call.

Example normalized output shape:

{
  "schema_version": "pyrrho_multitask_prediction_v1",
  "governance": {
    "raw_label": "TRUSTWORTHY",
    "final_label": "TRUSTWORTHY",
    "used_threshold_fallback": false,
    "threshold": 0.48,
    "confidence": 0.84,
    "probabilities": {
      "ABSTAIN": 0.08,
      "DISPUTED": 0.08,
      "TRUSTWORTHY": 0.84
    }
  },
  "query_contract": {
    "final_label": "structured_lookup"
  },
  "route": {
    "final_label": "economics_finance"
  },
  "taxonomy": {
    "final_label": "direct_answer"
  },
  "retrieval_action": {
    "final_label": "answer_now"
  },
  "scalars": {
    "evidence_sufficiency": 0.91,
    "query_evidence_alignment": 0.88,
    "answer_coverage": 0.86,
    "conflict_density": 0.08,
    "retrieval_retry_value": 0.12,
    "false_trustworthy_risk": 0.09,
    "evidence_failure_severity": 0.07
  }
}

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.

Intended Use

Use this model when a RAG or retrieval package needs fast local signals about:

• whether retrieved evidence is enough to answer,
• whether retrieved evidence conflicts,
• what kind of evidence the query needs before retrieval,
• which semantic/domain route the query belongs to,
• which fitz-gov support/failure pattern is active,
• what retrieval action and gap type the evidence state suggests,
• whether retrieval should retry, broaden, or escalate.

This model is not intended to write answers, verify facts outside the provided sources, replace a retriever, or replace human review in high-stakes settings.

Quick Start

Install the pyrrho package from the repository that contains this runtime, then load the package with the multitask predictor:

from huggingface_hub import snapshot_download

from pyrrho.multitask_inference import PyrrhoMultiTaskPredictor

MODEL_ID = "yafitzdev/pyrrho-nano-g4"
PACKAGE_DIR = snapshot_download(MODEL_ID)

query = "Which quarterly report is relevant?"
contexts = [
    "The Q2 report lists revenue, churn, and roadmap changes.",
]

predictor = PyrrhoMultiTaskPredictor.from_pretrained(PACKAGE_DIR, device="cpu")
result = predictor.predict(query, contexts)

print(result["governance"]["final_label"])
print(result["query_contract"]["final_label"])
print(result["route"]["final_label"])
print(result["taxonomy"]["final_label"])
print(result["retrieval_action"]["final_label"])
print(result["gap_type"]["final_label"])
print(result["scalars"])

For local package testing:

python scripts/package_multitask_encoder.py verify --package-dir models/pyrrho-nano-g4 --device cpu

Release Selection

• Seed: 1337
• TRUSTWORTHY threshold: 0.48
• Selection reason: Seed 1337 was selected from the completed official fitz-gov V9.0.0 3-seed run because it has the strongest validation composite score and the best balanced auxiliary-head tradeoff.

Held-Out Test Metrics

Metric	Result
Governance accuracy	0.9742
False-TRUSTWORTHY rate	0.0119
Query-contract accuracy	0.8756
Query-contract macro F1	0.8591
Route accuracy	0.9344
Route macro F1	0.9330
Taxonomy accuracy	0.7804
Taxonomy macro F1	0.7888
Scalar MAE	0.0687
Retrieval-action macro F1	0.8618
Gap-type macro F1	0.7055
Answerability-shape macro F1	0.9511
Retrieval-modality macro F1	0.5139

Three-seed headline from the local release summary:

Metric	Mean +/- std
Governance accuracy	97.46 +/- 0.09%
False-TRUSTWORTHY rate	1.21 +/- 0.06%
Query-contract macro F1	86.00 +/- 0.23%
Route accuracy	93.53 +/- 0.09%
Taxonomy accuracy	77.95 +/- 0.25%
Scalar MAE	0.0690 +/- 0.0003
Retrieval-action macro F1	85.92 +/- 0.43%
Gap-type macro F1	69.85 +/- 1.00%
Answerability-shape macro F1	94.92 +/- 0.19%
Retrieval-modality macro F1	51.94 +/- 0.59%

Training Data

Trained on the published fitz-gov V9.0.0 Hugging Face release with official query-grouped splits. Total prepared rows: 40,755 = 2,980 V6 rows + 7,520 V7 rows + 14,092 V8 rows + 16,163 V9 rows. Splits are train=32,625 / validation=4,104 / test=4,026. Split assignments come from v9/split_assignments.jsonl at dataset commit 874fd18d4952eec0e72b6df2264f8281615fd350. The release package records the local training config in training_config.yaml and detailed metrics in reports/summary.json.

Limitations

• This is a governance and routing co-processor, not a generator.
• The auxiliary heads are useful signals, not ground-truth explanations.
• Query-contract and route predictions are query-only and can be wrong when the user query is underspecified.
• Taxonomy and scalar outputs are trained on fitz-gov labels/signals and should be treated as decision-support metadata, not universal factual judgments.
• The four-way answerability head is intentionally collapsed for fitz-sage integration; it does not expose the old eleven detailed answerability labels.
• Retrieval modality remains the weakest auxiliary head; sparse subclasses such as pdf_layout, code, and configuration should be treated as hints, not hard guarantees.
• The license is CC BY-NC 4.0. Commercial use requires a separate license.