The method is Trajectory Utility Calibration (TUC). We train only from official LRAT trajectory-derived retrieval rows. TUC performs conservative row-internal auditing of each official query-positive-negative row: weak positives are downweighted or removed, high-confidence false-negative negatives are removed, and sample weights remain bounded.

Data Boundary

Training data source:

data/processed/lrat_broad_g16_v2_llm_full80k_conservative_weights.jsonl

Base checkpoint:

models/qwen3-embedding-0.6b-lrat-broad-g16-v2

Compliance boundary:

No generated queries.
No generated answers.
No generated passages or external evidence.
No A/B leaderboard qrels or hidden failures are used as training labels.
Strong models are used only as auditors over official LRAT query-positive-negative rows.

Training Configuration

Learning rate: 3e-7
Epochs: 0.5
Group size: 16
Batch per device: 24
Thresholds: drop_low_pos_quality<=1.0; false_negative relevance>=3.5 and confidence>=0.75; min_neg=4

Gate Evidence

Full80k gate assessment:

Verdict: pass
Report: reports/full80k_gate_assessment.json

Dense gate:

Report: reports/dense_gate_full80k_conservative_from_broad.json
Verdict: pass
Vs broad-g16-v2: recall@20 +0.0079996, recall@50 +0.0011203, hit@50 +0.0156627
Vs v5: recall@20 +0.0014922, recall@50 +0.0002979, hit@50 +0.0120482

Intended Use

This checkpoint is intended as a dense retriever in the XIR/BrowseComp-Plus-style fixed-corpus agent-search setting, where the submitted artifact is only the retriever checkpoint and the agent/judge/runtime are fixed externally.

Limitations

The checkpoint is optimized for agentic retrieval from LRAT-style trajectory supervision.
It should not be interpreted as a general-purpose factual QA model.
Leaderboard results should be used only as checkpoint gates, not as sources for new training labels.