pyrrho-v1-nano-g3.2 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-v1-nano-g3.1, this package adds V8.2 retrieval-control heads for retrieval action, gap type, answerability shape, retrieval modality, and an additional evidence_failure_severity scalar.
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-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.
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:
python scripts/package_multitask_encoder.py verify --package-dir models/pyrrho-v1-nano-g3.2 --device cpu
Release Selection
Seed: 1337
TRUSTWORTHY threshold: 0.34
Selection reason: Seed 1337 was selected because it has the best held-out governance result among the completed g3.2 seeds: 97.56% accuracy / 0.89% false-trustworthy at tau 0.34, while also having the strongest retrieval-modality held-out macro F1.
Held-Out Test Metrics
Metric
Result
Governance accuracy
0.9756
False-TRUSTWORTHY rate
0.0089
Query-contract accuracy
0.9492
Query-contract macro F1
0.9387
Route accuracy
0.9248
Route macro F1
0.9221
Taxonomy accuracy
0.8874
Taxonomy macro F1
0.8879
Scalar MAE
0.0634
Retrieval-action macro F1
0.8680
Gap-type macro F1
0.7537
Answerability-shape macro F1
0.7626
Retrieval-modality macro F1
0.5810
Three-seed headline from the local release summary:
Metric
Mean +/- std
Governance accuracy
97.29 +/- 0.22%
False-TRUSTWORTHY rate
1.11 +/- 0.17%
Query-contract macro F1
93.92 +/- 0.22%
Route accuracy
92.59 +/- 0.54%
Taxonomy accuracy
88.80 +/- 0.19%
Scalar MAE
0.0640 +/- 0.0005
Retrieval-action macro F1
86.69 +/- 0.23%
Gap-type macro F1
76.75 +/- 1.56%
Answerability-shape macro F1
75.53 +/- 0.54%
Retrieval-modality macro F1
55.75 +/- 1.75%
Training Data
Trained on fitz-gov V8.2.0 rows with mandatory routing.query_contract and routing.retrieval_control fields. 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.
Retrieval modality is the weakest new head; sparse subclasses such as pdf_layout should be treated as hints, not authoritative decisions.
The V8.2 retrieval-control labels are mechanically validated, but did not receive a separate independent blind-label QA pass before this local candidate package.
The license is CC BY-NC 4.0. Commercial use requires a separate license.
